Latent Space: The AI Engineer Podcast

Happy holidays! We’ll be sharing snippets from Latent Space LIVE! through the break bringing you the best of 2024! We want to express our deepest appreciation to event sponsors AWS, Daylight Computer, Thoth.ai, StrongCompute, Notable Capital, and most of all all our LS supporters who helped fund the gorgeous venue and A/V production!For NeurIPS last year we did our standard conference podcast coverage interviewing selected papers (that we have now also done for ICLR and ICML), however we felt that we could be doing more to help AI Engineers 1) get more industry-relevant content, and 2) recap 2024 year in review from experts. As a result, we organized the first Latent Space LIVE!, our first in person miniconference, at NeurIPS 2024 in Vancouver. Today, we’re proud to share Loubna’s highly anticipated talk (slides here)!Synthetic DataWe called out the Synthetic Data debate at last year’s NeurIPS, and no surprise that 2024 was dominated by the rise of synthetic data everywhere:* Apple’s Rephrasing the Web, Microsoft’s Phi 2-4 and Orca/AgentInstruct, Tencent’s Billion Persona dataset, DCLM, and HuggingFace’s FineWeb-Edu, and Loubna’s own Cosmopedia extended the ideas of synthetic textbook and agent generation to improve raw web scrape dataset quality* This year we also talked to the IDEFICS/OBELICS team at HuggingFace who released WebSight this year, the first work on code-vs-images synthetic data.* We called Llama 3.1 the Synthetic Data Model for its extensive use (and documentation!) of synthetic data in its pipeline, as well as its permissive license. * Nemotron CC and Nemotron-4-340B also made a big splash this year for how they used 20k items of human data to synthesize over 98% of the data used for SFT/PFT.* Cohere introduced Multilingual Arbitrage: Optimizing Data Pools to Accelerate Multilingual Progress observing gains of up to 56.5% improvement in win rates comparing multiple teachers vs the single best teacher model* In post training, AI2’s Tülu3 (discussed by Luca in our Open Models talk) and Loubna’s Smol Talk were also notable open releases this year.This comes in the face of a lot of scrutiny and criticism, with Scale AI as one of the leading voices publishing AI models collapse when trained on recursively generated data in Nature magazine bringing mainstream concerns to the potential downsides of poor quality syndata:Part of the concerns we highlighted last year on low-background tokens are coming to bear: ChatGPT contaminated data is spiking in every possible metric:But perhaps, if Sakana’s AI Scientist pans out this year, we will have mostly-AI AI researchers publishing AI research anyway so do we really care as long as the ideas can be verified to be correct?Smol ModelsMeta surprised many folks this year by not just aggressively updating Llama 3 and adding multimodality, but also adding a new series of “small” 1B and 3B “on device” models this year, even working on quantized numerics collaborations with Qualcomm, Mediatek, and Arm. It is near unbelievable that a 1B model today can qualitatively match a 13B model of last year:and the minimum size to hit a given MMLU bar has come down roughly 10x in the last year. We have been tracking this proxied by Lmsys Elo and inference price:The key reads this year are:* MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases* Apple Intelligence Foundation Language Models* Hymba: A Hybrid-head Architecture for Small Language Models* Loubna’s SmolLM and SmolLM2: a family of state-of-the-art small models with 135M, 360M, and 1.7B parameters on the pareto efficiency frontier.* and Moondream, which we already covered in the 2024 in Vision talkFull Talk on YouTubeplease like and subscribe!Timestamps* [00:00:05] Loubna Intro* [00:00:33] The Rise of Synthetic Data Everywhere* [00:02:57] Model Collapse* [00:05:14] Phi, FineWeb, Cosmopedia - Synthetic Textbooks* [00:12:36] DCLM, Nemotron-CC* [00:13:28] Post Training - AI2 Tulu, Smol Talk, Cohere Multilingual Arbitrage* [00:16:17] Smol Models* [00:18:24] On Device Models* [00:22:45] Smol Vision Models* [00:25:14] What's NextTranscript2024 in Synthetic Data and Smol Models[00:00:00] ​[00:00:05] Loubna Intro[00:00:05] Speaker: ​I'm very happy to be here. Thank you for the invitation. So I'm going to be talking about synthetic data in 2024. And then I'm going to be talking about small on device models. So I think the most interesting thing about synthetic data this year is that like now we have it everywhere in the large language models pipeline.[00:00:33] The Rise of Synthetic Data Everywhere[00:00:33] Speaker: I think initially, synthetic data was mainly used just for post training, because naturally that's the part where we needed human annotators. And then after that, we realized that we don't really have good benchmarks to [00:01:00] measure if models follow instructions well, if they are creative enough, or if they are chatty enough, so we also started using LLMs as judges.[00:01:08] Speaker: Thank you. And I think this year and towards the end of last year, we also went to the pre training parts and we started generating synthetic data for pre training to kind of replace some parts of the web. And the motivation behind that is that you have a lot of control over synthetic data. You can control your prompt and basically also the kind of data that you generate.[00:01:28] Speaker: So instead of just trying to filter the web, you could try to get the LLM to generate what you think the best web pages could look like and then train your models on that. So this is how we went from not having synthetic data at all in the LLM pipeline to having it everywhere. And so the cool thing is like today you can train an LLM with like an entirely synthetic pipeline.[00:01:49] Speaker: For example, you can use our Cosmopedia datasets and you can train a 1B model on like 150 billion tokens that are 100 percent synthetic. And those are also of good quality. And then you can [00:02:00] instruction tune the model on a synthetic SFT dataset. You can also do DPO on a synthetic dataset. And then to evaluate if the model is good, you can use.[00:02:07] Speaker: A benchmark that uses LLMs as a judge, for example, MTBench or AlpacaEvil. So I think this is like a really mind blowing because like just a few years ago, we wouldn't think this is possible. And I think there's a lot of concerns about model collapse, and I'm going to talk about that later. But we'll see that like, if we use synthetic data properly and we curate it carefully, that shouldn't happen.[00:02:29] Speaker: And the reason synthetic data is very popular right now is that we have really strong models, both open and closed. It is really cheap and fast to use compared to human annotations, which cost a lot and take a lot of time. And also for open models right now, we have some really good inference frameworks.[00:02:47] Speaker: So if you have enough GPUs, it's really easy to spawn these GPUs and generate like a lot of synthetic data. Some examples are VLM, TGI, and TensorRT.[00:02:57] Model Collapse[00:02:57] Speaker: Now let's talk about the elephant in the room, model [00:03:00] collapse. Is this the end? If you look at the media and all of like, for example, some papers in nature, it's really scary because there's a lot of synthetic data out there in the web.[00:03:09] Speaker: And naturally we train on the web. So we're going to be training a lot of synthetic data. And if model collapse is going to happen, we should really try to take that seriously. And the other issue is that, as I said, we think, a lot of people think the web is polluted because there's a lot of synthetic data.[00:03:24] Speaker: And for example, when we're building fine web datasets here at Guillerm and Hinek, we're interested in like, how much synthetic data is there in the web? So there isn't really a method to properly measure the amount of synthetic data or to save a webpage synthetic or not. But one thing we can do is to try to look for like proxy words, for example, expressions like as a large language model or words like delve that we know are actually generated by chat GPT.[00:03:49] Speaker: We could try to measure the amount of these words in our data system and compare them to the previous years. For example, here, we measured like a, these words ratio in different dumps of common crawl. [00:04:00] And we can see that like the ratio really increased after chat GPT's release. So if we were to say that synthetic data amount didn't change, you would expect this ratio to stay constant, which is not the case.[00:04:11] Speaker: So there's a lot of synthetic data probably on the web, but does this really make models worse? So what we did is we trained different models on these different dumps. And we then computed their performance on popular, like, NLP benchmarks, and then we computed the aggregated score. And surprisingly, you can see that the latest DOMs are actually even better than the DOMs that are before.[00:04:31] Speaker: So if there's some synthetic data there, at least it did not make the model's worse. Yeah, which is really encouraging. So personally, I wouldn't say the web is positive with Synthetic Data. Maybe it's even making it more rich. And the issue with like model collapse is that, for example, those studies, they were done at like a small scale, and you would ask the model to complete, for example, a Wikipedia paragraph, and then you would train it on these new generations, and you would do that every day.[00:04:56] Speaker: iteratively. I think if you do that approach, it's normal to [00:05:00] observe this kind of behavior because the quality is going to be worse because the model is already small. And then if you train it just on its generations, you shouldn't expect it to become better. But what we're really doing here is that we take a model that is very large and we try to distill its knowledge into a model that is smaller.[00:05:14] Phi, FineWeb, Cosmopedia - Synthetic Textbooks[00:05:14] Speaker: And in this way, you can expect to get like a better performance for your small model. And using synthetic data for pre-training has become really popular. After the textbooks are all you need papers where Microsoft basically trained a series of small models on textbooks that were using a large LLM.[00:05:32] Speaker: And then they found that these models were actually better than models that are much larger. So this was really interesting. It was like first of its time, but it was also met with a lot of skepticism, which is a good thing in research. It pushes you to question things because the dataset that they trained on was not public, so people were not really sure if these models are really good or maybe there's just some data contamination.[00:05:55] Speaker: So it was really hard to check if you just have the weights of the models. [00:06:00] And as Hugging Face, because we like open source, we tried to reproduce what they did. So this is our Cosmopedia dataset. We basically tried to follow a similar approach to what they documented in the paper. And we created a synthetic dataset of textbooks and blog posts and stories that had almost 30 billion tokens.[00:06:16] Speaker: And we tried to train some models on that. And we found that like the key ingredient to getting a good data set that is synthetic is trying as much as possible to keep it diverse. Because if you just throw the same prompts as your model, like generate like a textbook about linear algebra, and even if you change the temperature, the textbooks are going to look alike.[00:06:35] Speaker: So there's no way you could scale to like millions of samples. And the way you do that is by creating prompts that have some seeds that make them diverse. In our case, the prompt, we would ask the model to generate a textbook, but make it related to an extract from a webpage. And also we try to frame it within, to stay within topic.[00:06:55] Speaker: For example, here, we put like an extract about cardiovascular bioimaging, [00:07:00] and then we ask the model to generate a textbook related to medicine that is also related to this webpage. And this is a really nice approach because there's so many webpages out there. So you can. Be sure that your generation is not going to be diverse when you change the seed example.[00:07:16] Speaker: One thing that's challenging with this is that you want the seed samples to be related to your topics. So we use like a search tool to try to go all of fine web datasets. And then we also do a lot of experiments with the type of generations we want the model to generate. For example, we ask it for textbooks for middle school students or textbook for college.[00:07:40] Speaker: And we found that like some generation styles help on some specific benchmarks, while others help on other benchmarks. For example, college textbooks are really good for MMLU, while middle school textbooks are good for benchmarks like OpenBookQA and Pico. This is like a sample from like our search tool.[00:07:56] Speaker: For example, you have a top category, which is a topic, and then you have some [00:08:00] subtopics, and then you have the topic hits, which are basically the web pages in fine web does belong to these topics. And here you can see the comparison between Cosmopedia. We had two versions V1 and V2 in blue and red, and you can see the comparison to fine web, and as you can see throughout the training training on Cosmopedia was consistently better.[00:08:20] Speaker: So we managed to get a data set that was actually good to train these models on. It's of course so much smaller than FineWeb, it's only 30 billion tokens, but that's the scale that Microsoft data sets was, so we kind of managed to reproduce a bit what they did. And the data set is public, so everyone can go there, check if everything is all right.[00:08:38] Speaker: And now this is a recent paper from NVIDIA, Neumatron CC. They took things a bit further, and they generated not a few billion tokens, but 1. 9 trillion tokens, which is huge. And we can see later how they did that. It's more of, like, rephrasing the web. So we can see today that there's, like, some really huge synthetic datasets out there, and they're public, so, [00:09:00] like, you can try to filter them even further if you want to get, like, more high quality corpses.[00:09:04] Speaker: So for this, rephrasing the web this approach was suggested in this paper by Pratyush, where basically in this paper, they take some samples from C4 datasets, and then they use an LLM to rewrite these samples into a better format. For example, they ask an LLM to rewrite the sample into a Wikipedia passage or into a Q& A page.[00:09:25] Speaker: And the interesting thing in this approach is that you can use a model that is Small because it doesn't, rewriting doesn't require knowledge. It's just rewriting a page into a different style. So the model doesn't need to have like knowledge that is like extensive of what is rewriting compared to just asking a model to generate a new textbook and not giving it like ground truth.[00:09:45] Speaker: So here they rewrite some samples from C4 into Q& A, into Wikipedia, and they find that doing this works better than training just on C4. And so what they did in Nemo Trans CC is a similar approach. [00:10:00] They rewrite some pages from Common Crawl for two reasons. One is to, like improve Pages that are low quality, so they rewrite them into, for example, Wikipedia page, so they look better.[00:10:11] Speaker: And another reason is to create more diverse datasets. So they have a dataset that they already heavily filtered, and then they take these pages that are already high quality, and they ask the model to rewrite them in Question and Answer format. into like open ended questions or like multi choice questions.[00:10:27] Speaker: So this way they can reuse the same page multiple times without fearing like having multiple duplicates, because it's the same information, but it's going to be written differently. So I think that's also a really interesting approach for like generating synthetic data just by rephrasing the pages that you already have.[00:10:44] Speaker: There's also this approach called Prox where they try to start from a web page and then they generate a program which finds how to write that page to make it better and less noisy. For example, here you can see that there's some leftover metadata in the web page and you don't necessarily want to keep that for training [00:11:00] your model.[00:11:00] Speaker: So So they train a model that can generate programs that can like normalize and remove lines that are extra. So I think this approach is also interesting, but it's maybe less scalable than the approaches that I presented before. So that was it for like rephrasing and generating new textbooks.[00:11:17] Speaker: Another approach that I think is really good and becoming really popular for using synthetic data for pre training is basically building a better classifiers. For filtering the web for example, here we release the data sets called fine web edu. And the way we built it is by taking Llama3 and asking it to rate the educational content of web pages from zero to five.[00:11:39] Speaker: So for example, if a page is like a really good textbook that could be useful in a school setting, it would get a really high score. And if a page is just like an advertisement or promotional material, it would get a lower score. And then after that, we take these synthetic annotations and we train a classifier on them.[00:11:57] Speaker: It's a classifier like a BERT model. [00:12:00] And then we run this classifier on all of FineWeb, which is a 15 trillion tokens dataset. And then we only keep the pages that have like a score that's higher than 3. So for example, in our case, we went from 15 trillion tokens to 3. to just 1. 5 trillion tokens. Those are really highly educational.[00:12:16] Speaker: And as you can see here, a fine web EDU outperforms all the other public web datasets by a larger margin on a couple of benchmarks here, I show the aggregated score and you can see that this approach is really effective for filtering web datasets to get like better corpuses for training your LLMs.[00:12:36] DCLM, Nemotron-CC[00:12:36] Speaker: Others also try to do this approach. There's, for example, the DCLM datasets where they also train the classifier, but not to detect educational content. Instead, they trained it on OpenHermes dataset, which is a dataset for instruction tuning. And also they explain like IAM5 subreddits, and then they also get really high quality dataset which is like very information dense and can help [00:13:00] you train some really good LLMs.[00:13:01] Speaker: And then Nemotron Common Crawl, they also did this approach, but instead of using one classifier, they used an ensemble of classifiers. So they used, for example, the DCLM classifier, and also classifiers like the ones we used in FineWebEducational, and then they combined these two. Scores into a, with an ensemble method to only retain the best high quality pages, and they get a data set that works even better than the ones we develop.[00:13:25] Speaker: So that was it for like synthetic data for pre-training.[00:13:28] Post Training - AI2 Tulu, Smol Talk, Cohere Multilingual Arbitrage[00:13:28] Speaker: Now we can go back to post training. I think there's a lot of interesting post training data sets out there. One that was released recently, the agent instructs by Microsoft where they basically try to target some specific skills. And improve the performance of models on them.[00:13:43] Speaker: For example, here, you can see code, brain teasers, open domain QA, and they managed to get a dataset that outperforms that's when fine tuning Mistral 7b on it, it outperforms the original instruct model that was released by Mistral. And as I said, to get good synthetic data, you really [00:14:00] have to have a framework to make sure that your data is diverse.[00:14:03] Speaker: So for example, for them, they always. And then they see the generations on either source code or raw text documents, and then they rewrite them to make sure they're easier to generate instructions from, and then they use that for their like instruction data generation. There's also the Tool3SFT mixture, which was released recently by Allen AI.[00:14:23] Speaker: It's also really good quality and it covers a wide range of tasks. And the way they make sure that this dataset is diverse is by using personas from the persona hub datasets. Which is basically a data set of like I think over a million personas. And for example, in the tool mixture to generate like a new code snippet, they would give like the model persona, for example, a machine learning researcher interested in neural networks, and then ask it to generate like a coding problem.[00:14:49] Speaker: This way you make sure that your data set is really diverse, and then you can further filter the data sets, for example, using the reward models. We also released a dataset called Smalltalk, [00:15:00] and we also tried to cover the wide range of tasks, and as you can see here, for example, when fine tuning Mistral 7b on the dataset, we also outperformed the original Mistral instructs on a number of benchmarks, notably on mathematics and instruction following with ifevil.[00:15:18] Speaker: Another paper that's really interesting I wanted to mention is this one called Multilingual Data Arbitrage by Cohere. And basically they want to generate a data set for post training that is multilingual. And they have a really interesting problem. It's the fact that there isn't like one model that's really good at all the languages they wanted.[00:15:36] Speaker: So what they do is that like they use not just one teacher model, but multiple teachers. And then they have a router which basically sends the prompts they have to all these models. And then they get the completions and they have a reward model that traces all these generations and only keeps the best one.[00:15:52] Speaker: And this is like arbitrage and finance. So well, I think what's interesting in this, it shows that like synthetic data, it doesn't have to come from a single model. [00:16:00] And because we have so many good models now, you could like pull these models together and get like a dataset that's really high quality and that's diverse and that's covers all your needs.[00:16:12] Speaker: I was supposed to put a meme there, but. Yeah, so that was it for like a synthetic data.[00:16:17] Smol Models[00:16:17] Speaker: Now we can go to see what's happening in the small models field in 2024. I don't know if you know, but like now we have some really good small models. For example, Lama 3. 2 1B is. It matches Lama 2. 13b from, that was released last year on the LMSYS arena, which is basically the default go to leaderboard for evaluating models using human evaluation.[00:16:39] Speaker: And as you can see here, the scores of the models are really close. So I think we've made like hugely forward in terms of small models. Of course, that's one, just one data point, but there's more. For example, if you look at this chart from the Quint 2. 5 blog post, it shows that today we have some really good models that are only like 3 billion parameters [00:17:00] and 4 billion that score really high on MMLU.[00:17:03] Speaker: Which is a really popular benchmark for evaluating models. And you can see here that the red, the blue dots have more than 65 on MMLU. And the grey ones have less. And for example, Llama33b had less. So now we have a 3b model that outperforms a 33b model that was released earlier. So I think now people are starting to realize that like, we shouldn't just scale and scale models, but we should try to make them more efficient.[00:17:33] Speaker: I don't know if you knew, but you can also chat with a 3B plus model on your iPhone. For example, here, this is an app called PocketPal, where you can go and select a model from Hugging Face. It has a large choice. For example, here we loaded the 5. 3. 5, which is 3. 8 billion parameters on this iPhone. And we can chat with this and you can see that even the latency is also acceptable.[00:17:57] Speaker: For example, here, I asked it to give me a joke about [00:18:00] NeurIPS. So let's see what it has to say.[00:18:06] Speaker: Okay, why did the neural network attend NeurIPS? Because it heard there would be a lot of layers and fun and it wanted to train its sense of humor. So not very funny, but at least it can run on device. Yeah, so I think now we have good small models, but we also have like good frameworks and tools to use these small models.[00:18:24] On Device Models[00:18:24] Speaker: So I think we're really close to having like really on edge and on device models that are really good. And I think for a while we've had this narrative. But just training larger models is better. Of course, this is supported by science scaling laws. As you can see here, for example, when we scale the model size, the loss is lower and obviously you get a better model.[00:18:46] Speaker: But and we can see this, for example, in the GPT family of models, how we went from just a hundred million parameters to more than a trillion. parameters. And of course, we all observed the performance improvement when using the latest model. But [00:19:00] one thing that we shouldn't forget is that when we scale the model, we also scale the inference costs and time.[00:19:05] Speaker: And so the largest models were are going to cost so much more. So I think now instead of just building larger models, we should be focusing on building more efficient models. It's no longer a race for the largest models since these models are really expensive to run and they require like a really good infrastructure to do that and they cannot run on, for example, consumer hardware.[00:19:27] Speaker: And when you try to build more efficient models that match larger models, that's when you can really unlock some really interesting on device use cases. And I think a trend that we're noticing now is the trend of training smaller models longer. For example, if you compare how much, how long LLAMA was trained compared to LLAMA3, there is a huge increase in the pre training length.[00:19:50] Speaker: LLAMA was trained on 1 trillion tokens, but LLAMA3 8b was trained on 15 trillion tokens. So Meta managed to get a model that's the same size, but But it performs so much [00:20:00] better by choosing to like spend the sacrifice during training, because as we know, training is a one time cost, but inference is something that's ongoing.[00:20:08] Speaker: If we want to see what are like the small models reads in 2024, I think this mobile LLM paper by Meta is interesting. They try to study different models that are like have the less than 1 billion parameters and find which architecture makes most sense for these models. For example, they find that depth is more important than width.[00:20:29] Speaker: So it's more important to have models that have like more layers than just one. making them more wide. They also find that GQA helps, that tying the embedding helps. So I think it's a nice study overall for models that are just a few hundred million parameters. There's also the Apple intelligence tech report, which is interesting.[00:20:48] Speaker: So for Apple intelligence, they had two models, one that was like on server and another model that was on device. It had 3 billion parameters. And I think the interesting part is that they trained this model using [00:21:00] pruning. And then distillation. And for example, they have this table where they show that, like, using pruning and distillation works much better than training from scratch.[00:21:08] Speaker: And they also have some interesting insights about, like, how they specialize their models on specific tasks, like, for example, summarization and rewriting. There's also this paper by NVIDIA that was released recently. I think you've already had a talk about, like, hybrid models that was all interesting.[00:21:23] Speaker: And this model, they used, like, a hybrid architecture between state space models and transformers. And they managed to train a 1B model that's really performant without needing to train it on a lot of tokens. And regarding our work, we just recently released SmallM2, so it's a series of three models, which are the best in class in each model size.[00:21:46] Speaker: For example, our 1. 7b model outperforms Lama 1b and also Qt 2. 5. And how we managed to train this model is the following. That's where you spent a lot of time trying to curate the pre training datasets. We did a lot of [00:22:00] ablations, trying to find which datasets are good and also how to mix them. We also created some new math and code datasets that we're releasing soon.[00:22:08] Speaker: But you basically really spent a lot of time trying to find what's the best mixture that you can train these models on. And then we spent some time trying to like we also trained these models for very long. For example, small M1 was trained only on 1 trillion tokens, but this model is trained on 11 trillion tokens.[00:22:24] Speaker: And we saw that the performance kept improving. The models didn't really plateau mid training, which I think is really interesting. It shows that you can train such small models for very long and keep getting performance gains. What's interesting about SmallLM2 is that it's fully open. We also released, like the pre training code base, the fine tuning code, the datasets, and also evaluation in this repository.[00:22:45] Smol Vision Models[00:22:45] Speaker: Also there's, like, really interesting small models for text, but also for vision. For example, here you can see SmallVLM, which is a 2B model that's really efficient. It doesn't consume a lot of RAM, and it also has a good performance. There's also Moondream 0. [00:23:00] 5b, which was released recently. It's like the smallest visual language model.[00:23:04] Speaker: And as you can see, there isn't like a big trade off compared to Moondream 2b. So now I showed you that we have some really good small models. We also have the tools to use them, but why should you consider using small models and when? I think, like, small models are really interesting because of the on device feature.[00:23:23] Speaker: Because these models are small and they can run fast, you can basically run them on your laptop, but also on your mobile phone. And this means that your dataset stays locally. You don't have to send your queries to third parties. And this really enhances privacy. That was, for example, one of the big selling points for Apple Intelligence.[00:23:42] Speaker: Also, right now, we really have a lot of work to do. So many frameworks to do on device inference. For example, there's MLX, MLC, Llama, CPP, Transformers, JS. So we have a lot of options and each of them have like great features. So you have so many options for doing that. Small models are also really powerful if you choose to specialize them.[00:24:00][00:24:00] Speaker: For example, here there's a startup called Numind, which took small LM and then they fine tuned it on text extraction datasets. And they managed to get a model that's not very far from models that are much larger. So I think text extraction is like one use case where small models can be really performant and it makes sense to use them instead of just using larger models.[00:24:19] Speaker: You can also chat with these models in browser. For example, here, you can go there, you can load the model, you can even turn off your internet and just start chatting with the model locally. Speaking of text extraction, if you don't want to fine tune the models, there's a really good method of structure generation.[00:24:36] Speaker: We can basically force the models to follow a JSON schema that you defined. For example, here, we try to force the model to follow a schema for extracting key information from GitHub issues. So you can input free text, which is a complaint about a GitHub repository, something not working. And then you can run it there and the model can extract anything that is relevant for your GitHub issue creation.[00:24:58] Speaker: For example, the [00:25:00] priority, for example, here, priority is high, the type of the issue bug, and then a title and the estimation of how long this will take to fix. And you can just like do this in the browser, you can transform your text into a GitHub issue that's properly formatted.[00:25:14] What's Next[00:25:14] Speaker: So what's next for synthetic data and small models?[00:25:18] Speaker: I think that domain specific synthetic data is going to be, it's already important, it's going to be even more important. For example, generating synthetic data for math. I think this really would help improve the reasoning of a lot of models. And a lot of people are doing it, for example, Quint 2. 12 math, everyone's trying to reproduce a one.[00:25:37] Speaker: And so I think for synthetic data, trying to specialize it on some domains is going to be really important. And then for small models, I think specializing them through fine tuning, it's also going to be really important because I think a lot of companies are just trying to use these large models because they are better.[00:25:53] Speaker: But on some tasks, I think you can already get decent performance with small models. So you don't need to Pay like a [00:26:00] cost that's much larger just to make your model better at your task by a few percent. And this is not just for text. And I think it also applies for other modalities like vision and audio.[00:26:11] Speaker: And I think you should also watch out for on device frameworks and applications. For example, like the app I showed, or lama, all these frameworks are becoming really popular and I'm pretty sure that we're gonna get like more of them in 2025. And users really like that. Maybe for other, I should also say hot take.[00:26:28] Speaker: I think that like in AI, we just started like with fine tuning, for example, trying to make BERT work on some specific use cases, and really struggling to do that. And then we had some models that are much larger. So we just switched to like prompt engineering to get the models And I think we're going back to fine tuning where we realize these models are really costly.[00:26:47] Speaker: It's better to use just a small model or try to specialize it. So I think it's a little bit of a cycle and we're going to start to see like more fine tuning and less of just like a prompt engineering the models. So that was my talk. Thank you for following. And if you have [00:27:00] any questions, we can take them now. Get full access to Latent Space at www.latent.space/subscribe

Happy holidays! We’ll be sharing snippets from Latent Space LIVE! through the break bringing you the best of 2024! We want to express our deepest appreciation to event sponsors AWS, Daylight Computer, Thoth.ai, StrongCompute, Notable Capital, and most of all all our LS supporters who helped fund the gorgeous venue and A/V production!For NeurIPS last year we did our standard conference podcast coverage interviewing selected papers (that we have now also done for ICLR and ICML), however we felt that we could be doing more to help AI Engineers 1) get more industry-relevant content, and 2) recap 2024 year in review from experts. As a result, we organized the first Latent Space LIVE!, our first in person miniconference, at NeurIPS 2024 in Vancouver.Of perennial interest, particularly at academic conferences, is scaled-up architecture research as people hunt for the next Attention Is All You Need. We have many names for them: “efficient models”, “retentive networks”, “subquadratic attention” or “linear attention” but some of them don’t even have any lineage with attention - one of the best papers of this NeurIPS was Sepp Hochreiter’s xLSTM, which has a particularly poetic significance as one of the creators of the LSTM returning to update and challenge the OG language model architecture:So, for lack of a better term, we decided to call this segment “the State of Post-Transformers” and fortunately everyone rolled with it.We are fortunate to have two powerful friends of the pod to give us an update here:* Together AI: with CEO Vipul Ved Prakash and CTO Ce Zhang joining us to talk about how they are building Together together as a quote unquote full stack AI startup, from the lowest level kernel and systems programming to the highest level mathematical abstractions driving new model architectures and inference algorithms, with notable industry contributions from RedPajama v2, Flash Attention 3, Mamba 2, Mixture of Agents, BASED, Sequoia, Evo, Dragonfly, Dan Fu's ThunderKittens and many more research projects this year* Recursal AI: with CEO Eugene Cheah who has helped lead the independent RWKV project while also running Featherless AI. This year, the team has shipped RWKV v5, codenamed Eagle, to 1.5 billion Windows 10 and Windows 11 machines worldwide, to support Microsoft's on-device, energy-usage-sensitive Windows Copilot usecases, and has launched the first updates on RWKV v6, codenamed Finch and GoldFinch. On the morning of Latent Space Live, they also announced QRWKV6, a Qwen 32B model modified with RWKV linear attention layers. We were looking to host a debate between our speakers, but given that both of them were working on post-transformers alternativesFull Talk on YoutubePlease like and subscribe!LinksAll the models and papers they picked:* Earlier Cited Work* Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention* Hungry hungry hippos: Towards language modeling with state space models* Hyena hierarchy: Towards larger convolutional language models* Mamba: Linear-Time Sequence Modeling with Selective State Spaces* S4: Efficiently Modeling Long Sequences with Structured State Spaces* Just Read Twice (Arora et al)* Recurrent large language models that compete with Transformers in language modeling perplexity are emerging at a rapid rate (e.g., Mamba, RWKV). Excitingly, these architectures use a constant amount of memory during inference. However, due to the limited memory, recurrent LMs cannot recall and use all the information in long contexts leading to brittle in-context learning (ICL) quality. A key challenge for efficient LMs is selecting what information to store versus discard. In this work, we observe the order in which information is shown to the LM impacts the selection difficulty. * To formalize this, we show that the hardness of information recall reduces to the hardness of a problem called set disjointness (SD), a quintessential problem in communication complexity that requires a streaming algorithm (e.g., recurrent model) to decide whether inputted sets are disjoint. We empirically and theoretically show that the recurrent memory required to solve SD changes with set order, i.e., whether the smaller set appears first in-context. * Our analysis suggests, to mitigate the reliance on data order, we can put information in the right order in-context or process prompts non-causally. Towards that end, we propose: (1) JRT-Prompt, where context gets repeated multiple times in the prompt, effectively showing the model all data orders. This gives 11.0±1.3 points of improvement, averaged across 16 recurrent LMs and the 6 ICL tasks, with 11.9× higher throughput than FlashAttention-2 for generation prefill (length 32k, batch size 16, NVidia H100). We then propose (2) JRT-RNN, which uses non-causal prefix-linear-attention to process prompts and provides 99% of Transformer quality at 360M params., 30B tokens and 96% at 1.3B params., 50B tokens on average across the tasks, with 19.2× higher throughput for prefill than FA2.* Jamba: A 52B Hybrid Transformer-Mamba Language Model* We present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. * Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. * This flexible architecture allows resource- and objective-specific configurations. In the particular configuration we have implemented, we end up with a powerful model that fits in a single 80GB GPU.* Built at large scale, Jamba provides high throughput and small memory footprint compared to vanilla Transformers, and at the same time state-of-the-art performance on standard language model benchmarks and long-context evaluations. Remarkably, the model presents strong results for up to 256K tokens context length. * We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. We also describe several interesting properties of these architectures which the training and evaluation of Jamba have revealed, and plan to release checkpoints from various ablation runs, to encourage further exploration of this novel architecture. We make the weights of our implementation of Jamba publicly available under a permissive license.* SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers* We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: * (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. * (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. * (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. * (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. * As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024×1024 resolution image. Sana enables content creation at low cost. * RWKV: Reinventing RNNs for the Transformer Era* Transformers have revolutionized almost all natural language processing (NLP) tasks but suffer from memory and computational complexity that scales quadratically with sequence length. In contrast, recurrent neural networks (RNNs) exhibit linear scaling in memory and computational requirements but struggle to match the same performance as Transformers due to limitations in parallelization and scalability. * We propose a novel model architecture, Receptance Weighted Key Value (RWKV), that combines the efficient parallelizable training of transformers with the efficient inference of RNNs.* Our approach leverages a linear attention mechanism and allows us to formulate the model as either a Transformer or an RNN, thus parallelizing computations during training and maintains constant computational and memory complexity during inference. * We scale our models as large as 14 billion parameters, by far the largest dense RNN ever trained, and find RWKV performs on par with similarly sized Transformers, suggesting future work can leverage this architecture to create more efficient models. This work presents a significant step towards reconciling trade-offs between computational efficiency and model performance in sequence processing tasks.* LoLCATs: On Low-Rank Linearizing of Large Language Models* Recent works show we can linearize large language models (LLMs) -- swapping the quadratic attentions of popular Transformer-based LLMs with subquadratic analogs, such as linear attention -- avoiding the expensive pretraining costs. However, linearizing LLMs often significantly degrades model quality, still requires training over billions of tokens, and remains limited to smaller 1.3B to 7B LLMs. * We thus propose Low-rank Linear Conversion via Attention Transfer (LoLCATs), a simple two-step method that improves LLM linearizing quality with orders of magnitudes less memory and compute. * We base these steps on two findings. * First, we can replace an LLM's softmax attentions with closely-approximating linear attentions, simply by training the linear attentions to match their softmax counterparts with an output MSE loss ("attention transfer").* Then, this enables adjusting for approximation errors and recovering LLM quality simply with low-rank adaptation (LoRA). * LoLCATs significantly improves linearizing quality, training efficiency, and scalability. We significantly reduce the linearizing quality gap and produce state-of-the-art subquadratic LLMs from Llama 3 8B and Mistral 7B v0.1, leading to 20+ points of improvement on 5-shot MMLU. * Furthermore, LoLCATs does so with only 0.2% of past methods' model parameters and 0.4% of their training tokens. * Finally, we apply LoLCATs to create the first linearized 70B and 405B LLMs (50x larger than prior work). * When compared with prior approaches under the same compute budgets, LoLCATs significantly improves linearizing quality, closing the gap between linearized and original Llama 3.1 70B and 405B LLMs by 77.8% and 78.1% on 5-shot MMLU.Timestamps* [00:02:27] Intros* [00:03:16] Why Scale Context Lengths? or work on Efficient Models* [00:06:07] The Story of SSMs* [00:09:33] Idea 1: Approximation -> Principled Modeling* [00:12:14] Idea 3: Selection* [00:15:07] Just Read Twice* [00:16:51] Idea 4: Test Time Compute* [00:17:32] Idea 2: Hardware & Kernel Support* [00:19:49] RWKV vs SSMs* [00:24:24] RWKV Arch* [00:26:15] QWRKWv6 launch* [00:30:00] What's next* [00:33:21] Hot Takes - does anyone really need long context?Transcript[00:00:00] AI Charlie: We're back at Latent Space Live, our first mini conference held at NeurIPS 2024 in Vancouver. This is Charlie, your AI co host. As a special treat this week, we're recapping the best of 2024 going domain by domain. We sent out a survey to the over 900 of you who told us what you wanted, and then invited the best speakers in the Latent Space Network to cover each field.[00:00:24] AI Charlie: 200 of you joined us in person throughout the day, with over 2200 watching live online. Thanks Our next keynote covers the State of Transformers alternative architectures, with a special joint presentation with Dan Fu of Together AI and Eugene Chia of Recursal AI and Featherless AI. We've featured both Together and Recursal on the pod before, with CEO Veepal Vedprakash introducing them.[00:00:49] AI Charlie: And CTO CE Zhang joining us to talk about how they are building together together as a quote unquote full stack AI startup from the lowest level kernel and systems [00:01:00] programming to the highest level mathematical abstractions driving new model architectures and inference algorithms with notable industry contributions from Red Pajama V2, Flash Attention 3, Mamba 2, Mixture of Agents.[00:01:15] AI Charlie: Based, Sequoia, Evo, Dragonfly, Danfoo's Thunder Kittens, and many more research projects this year. As for Recursal and Featherless, we were the first podcast to feature RWKV last year, and this year the team has shipped RWKV v5, codenamed Eagle, to 1. 5 billion Windows 10 and Windows 11 machines worldwide to support Microsoft's on device, end Energy Usage Sensitive Windows Copilot Use Cases and has launched the first updates on RWKV v6, codenamed Finch and Goldfinch.[00:01:53] AI Charlie: On the morning of Latent Space Live, they also announced QRdata UKv6, a QEN32B model [00:02:00] modified with RDWKV linear attention layers. Eugene has also written the most single most popular guest post on the Latent Space blog this year. Yes, we do take guest posts on what he has discovered about the H100 GPU inference NeoCloud market since the successful launch of Featherless AI this year.[00:02:20] AI Charlie: As always, don't forget to check the show notes for the YouTube link to their talk as well as their slides. Watch out and take care.[00:02:27] Intros[00:02:27] Dan Fu: Yeah, so thanks so much for having us. So this is going to be a little bit of a two part presentation. My name is Dan. I'm at Together AI, and I'll be joining UCSD as faculty in about a year. And Eugene, you want to introduce yourself?[00:02:46] Eugene Cheah: Eugene, I lead the art activity team, and I, I'm CEO of Featherless, and we both work on this new post transformer architecture space.[00:02:55] Dan Fu: Yeah, so yeah, so today we're really excited to talk to you a little bit [00:03:00] about that. So first I'm going to give a broad overview of kind of the last few years of progress in non post transformer architectures. And then afterwards Eugene will tell us a little bit about the latest and the greatest and the latest frontier models in this space.[00:03:16] Why Scale Context Lengths? or work on Efficient Models[00:03:16] Dan Fu: So, the story starts with Scaling. So this is probably a figure or something like this that you've seen very recently. Over the last five to six years, we've seen models really scale up in parameter size, and that's brought with it a bunch of new capabilities, like the ability to talk to you and tell you sometimes how to use your Colab screens.[00:03:35] Dan Fu: But another place where we've seen scaling especially recently is scaling in context length. So this can mean Having more text inputs for your models, but it can also mean things like taking a lot of visual token inputs image inputs to your models or generating lots of outputs. And one thing that's been really exciting over the last few months or so is that we're, we're seeing scaling, not only during training time, but also [00:04:00] during test time.[00:04:00] Dan Fu: So this is one of the, the, this is the iconic image from the OpenAI 01 release. Not only are we starting to scale train time compute, but we're also starting to scale test time compute. Now if you're familiar with our attention and our transformer architectures today, this graph on the right might look a little bit scary.[00:04:19] Dan Fu: And one of the reasons is that the implications are a little bit Interesting. So what does it mean if we want to continue having smarter and smarter models? Do we just need to start building bigger, bigger data centers, spending more flops? Is this this little Dolly 3, we need more flops, guys? Is this going to be the future of all of AI?[00:04:39] Dan Fu: Or is there a better way, another path forward? Maybe we can get the same capabilities that we've gotten used to, But for a lot less compute, a lot less flops. And one of the things that we're going to talk about today is specifically looking at that core attention operator in some of these models.[00:04:57] Dan Fu: And the reason is that so this is just some, some [00:05:00] basic you know, scaling curves, but attention has compute that scales quadratically in the context length. So that means that if you're doing something like test time compute and you want to spend a bunch of tokens thinking about what comes next, the longer that that goes the, the, the more tokens you spend on that, that compute grows quadratically in that.[00:05:19] Dan Fu: One of the questions that we're interested in is, can we take that basic sequence model, that basic sequence primitive at the bottom, and get it to scale better? Can we scale in, let's say, n to the 3 halves or n log n? So in, in the first part of the talk, so we just went over the introduction. What I'm gonna do over the next few slides is just talk about some of the key advances and ideas that have shown over the past few years since maybe early 2020 to, to now that shown promise that this might actually be possible.[00:05:48] Dan Fu: That you can actually get potentially the same quality that we want while scale, while scaling better. So to do that, we're and, and basically the, the story that we're gonna look is we're gonna start to see [00:06:00] how. So this is a basic graph of just the past couple years of progress of perplexity where that blue line, that dotted blue line, is attention.[00:06:07] The Story of SSMs[00:06:07] Dan Fu: It's your basic transformer, full dense attention. And then the dots coming down are some of the methods that you'll see in this presentation today. We're going to turn the clock back all the way to 2020. So this, this, this question of can we make attention subquadratic? Basically, as soon as we said attention is all you need, People started asking this question.[00:06:28] Dan Fu: So we have this quadratic attention operator. Can we do better? I'll briefly talk about why attention is quadratic. And the basic thing that happens, if you're not familiar, is that you have these inputs, these keys and queries. And what you do in this attention matrix, this S matrix over here, is that you're using, you're comparing every token in your input to every other token.[00:06:49] Dan Fu: So when I try to do something like upload a whole book to Gemini, what happens beyond the Maybe not Gemini, because we don't necessarily know what architecture is. But let's say we upload it to LLAMA, what happens beyond [00:07:00] the scenes, behind the scenes, is that it's going to take every single word in that book and compare it to every other word.[00:07:05] Dan Fu: And this has been a really, it's, it's led to some pretty impressive things. But it's kind of a brute forcing of the way that you would try to interpret a interpret something. And what attention does in particular is the, and then what attention, sorry, don't want to. Okay, no, no laser pointer. What, what attention does afterwards is that instead of always operating in this quadratic thing, it takes a row wise softmax over this matrix, and then multiplies it by this values matrix.[00:07:32] Dan Fu: So, one of the key points to notice is that the output size is always going to be the same as the inputs, at least in standard self attention. So one of the first things that folks tried to do around 2020 is this thing called linear attention, which is just, just noticing that if we take out this softmax from here, if we take out this non linearity in the middle of the attention operation, and then if you compute the keys and the values operation first, you actually never hit this quadratic bottleneck.[00:07:57] Dan Fu: So that, that's potentially a way [00:08:00] to get a lot more computationally efficient. And there are various ways to do this by basically using feature maps or try to approximate this overall attention computation. But some of this work sort of started to hit a wall in 2020. And the basic challenges were, were two.[00:08:16] Dan Fu: So one was quality. It was back then, it was kind of hard to, to get good quality with these linear attention operators. The other one was actually hardware efficiency. So these, this feature map that was just shown by a simplify simplify here. Actually ends up being quite computationally expensive if you just implement it naively.[00:08:34] Dan Fu: So you started having these operators that not only were you sure, you're not really sure if they have the same quality, but also they're actually just wall clock slower. So you kind of end up getting the worst of both worlds. So this was the the stage. So that kind of sets the stage for four years ago.[00:08:49] Dan Fu: Keep this in mind because linear attention is actually going to come back in a few years once we have a better understanding. But one of the works that started kicking off this, this [00:09:00] mini revolution in post transformer architectures was this idea called states based model. So here the seminal work is, is one about our work queue in 2022.[00:09:09] Dan Fu: And this, this piece of work really brought together a few ideas from, from some long running research research lines of work. The first one was, and this is really one of the keys to, to closing the gap in quality was just using things that, that if you talk to a, a, an electrical engineer off the street, they might know off, off the, like the back of their hand.[00:09:33] Idea 1: Approximation -> Principled Modeling[00:09:33] Dan Fu: But taking some of those properties with how we model dynamical systems in signal processing and then using those ideas to model the inputs, the, the text tokens in, for example a transformer like Next Token Prediction Architecture. So some of those early states-based model papers were looking at this relatively, relatively simple recurrent update model that comes from maybe chapter one of a signal processing class.[00:09:59] Dan Fu: But then using [00:10:00] some principle theory about how you should do that recurrent update in order to really get the most that you can out of your hidden state, out of your out of your sequence. So that, that was one key idea for quality and. When this was eventually realized, you started to see a bunch of benchmarks that were pretty sticky for a few years.[00:10:20] Dan Fu: Things like long range arena, some long sequence evaluation benchmarks, There was stuff in time series, time series analysis. They started to, you started to see the quality tick up in meaningful ways. But the other key thing that What's so influential about these states based models is that they also had a key idea about how you can compute these things efficiently.[00:10:45] Dan Fu: So if you go back to your machine learning 101 class where you learned about RNNs, one thing that you may have learned is that they don't paralyze as well as detention, because if you just run them naively, you have to do this kind of sequential update to process new tokens, [00:11:00] whereas in attention, you can process all the tokens in parallel at one time.[00:11:04] Dan Fu: One of the key insights behind the S4 paper was that these recurrent models, you could take them and you could also formulate them as a convolution. And in particular, with a convolution, you could, instead of using a PyTorch conv1d operation, you can compute that with the FFT. And that would give you n log n compute in the in the sequence length n with an operator that was relatively well optimized for modern hardware.[00:11:28] Dan Fu: So those are really, I'd say, the two key ideas in 2022 that started allowing these breakthroughs to happen in these non transformer architectures. So, these ideas about how to principally model sorry, how to model the recurrent updates of a mo of, of a sequence in a principled way, and also these key ideas in how you can compute it efficiently by turning it into a convolution and then scaling it up with the FFT.[00:11:53] Dan Fu: Along those same lines, so afterwards we started putting out some work on specialized kernels, so just [00:12:00] like we have flash attention for transformers, we also have works like flash fft conf, and if you look at these lines of work oftentimes when, whenever you see a new architecture, you see a new primitive one of the, one of the table stakes now is, do you have an efficient kernel so that you can actually get wall clock speed up?[00:12:14] Idea 3: Selection[00:12:14] Dan Fu: So by 2022, We are starting to have these models that had promising quality primitives, but and, and also promising wall clocks. So you could actually see regimes where they were better than transformers in meaningful ways. That being said, there were, there's still sometimes a quality gap, particularly for language modeling.[00:12:33] Dan Fu: And because languages, It's so core to what we do in sequence modeling these days the, the next, the next key idea that I'm going to talk about is this idea of selection mechanisms. And this is basically an idea of, so you have this recurrent state that you're keeping around that just summarizes everything that, that came before.[00:12:50] Dan Fu: And to get a good sequence model, one of the things that you really need to be able to do is have the model learn what's the best way to pick out pieces from that recurrent [00:13:00] state. So one of the, one of the major ideas here in a line of work called H3, Hungry Hungry Hippos, and also these hyena models were One way you can do this is by just adding some simple element wise gates.[00:13:13] Dan Fu: So versions of these ideas have been around for decades. If you squint at the LSTM paper you, you can probably find, find this gating mechanism. But turns out you can take those old ideas, add them into these new. state space models, and then you can see quality start to pick up. If you've heard of the Mamba model, this also takes the selection to the next level by actually making some changes in that fundamental recurrent state space.[00:13:40] Dan Fu: So, it's not only just this gating that happens around the SSM layer, but also you can actually make The ABCD matrices of your state space model, you can make them data dependent, which will allow you to even better select out different pieces from your hidden state depending on what you're seeing. I'll also point out if you look at the [00:14:00] bottom right of this figure, there's this little triangle with a GPU SRAM, GPU HBM, and this, this is just continuing that trend of when you have a new architecture you, you, you also release it with a kernel to, to, to show that it is hardware efficient, that it, that it can be hardware efficient on modern hardware.[00:14:17] Dan Fu: The, the, one of the next cool things that happened is once we had this understanding of these are the basic pieces, these are the basic principles behind some of the sequence models linear attention actually started to come back. So in earlier this year, there was a model called BASED the, from Simran Arora and, and some other folks, that combined a more principled version of linear attention that basically the, the, the, the two second summary is that it used a Taylor approximation of the softmax attention, combined that with a simple sliding window attention and was starting to able, starting to be able to expand the Pareto frontier of how much data can you recall from your sequence, versus how small is your recurrent state size.[00:14:58] Dan Fu: So those orange dots [00:15:00] are, at the top there, are just showing smaller sequences that can recall more memory.[00:15:07] Just Read Twice[00:15:07] Dan Fu: And the last major idea I think that has been influential in this line of work and is very relatively late breaking just a few months ago, is just the basic idea that when you have these models that are fundamentally more efficient in the sequence length, you maybe don't want to prompt them or use them in exactly the same way.[00:15:26] Dan Fu: So this was a really cool paper called Just Read Twice, also from Simran. That basically said, hey, all these efficient models can process tokens so much more efficiently than transformers that they can sometimes have unfair advantages compared to a simple transformer token. So, or sorry, a simple transformer model.[00:15:44] Dan Fu: So take, for example the standard, the standard use case of you have some long document, you're going to pass it in as input, and then you're going to ask some question about it. One problem you might imagine for a recurrent model where you have a fixed state size is, let's say that [00:16:00] you're. Article is very long, and you're trying to ask about some really niche thing.[00:16:04] Dan Fu: You can imagine it might be hard for the model to know ahead of time what information to put into the hidden state. But these, these, these models are so much more efficient that you can do something really stupid, like, you can just put the document write down the document, write down the question, write down the document again, and then write down the question again, and then this time, the second time that you go over that document, you know exactly what to look for.[00:16:25] Dan Fu: And the cool thing about this is, so this is, And this this results in better quality, especially on these recall intensive tasks. But the other interesting thing is it really takes advantage of the more efficient architectures that, that we're having here. So one of the other, I think, influential ideas in this line of work is if you change the fundamental compute capabilities of your model and the way that it scales, you can actually start to query it at test time differently.[00:16:51] Idea 4: Test Time Compute[00:16:51] Dan Fu: And this actually, of course, goes back to those slides on test time compute. So while everybody's looking at, say, test time compute for big transformer models, [00:17:00] I think potentially a really interesting research question is, how can you take those and how does it change with this new next generation of models?[00:17:09] Dan Fu: So the, I'll just briefly summarize what some of those key ideas were and then talk and then show you briefly kind of what the state of the art is today. So, so the four key ideas are instead of just doing a simple linear attention approximation, instead take ideas that we know from other fields like signal processing, do a more principled approach to your modeling of the sequence.[00:17:32] Idea 2: Hardware & Kernel Support[00:17:32] Dan Fu: Another key idea throughout all these lines of work is you really want. Hardware and kernel support from day one. So, so even if your model is theoretically more efficient if somebody goes and runs it and it's two times slower one of the things that, that we've learned is that if, if you're in that situation, it's, it's just gonna be dead on arrival.[00:17:49] Dan Fu: So you want to be designing your architectures one of the key, key machine learning ideas that has been important for the quality is just making sure that you encode different ways that you can [00:18:00] select from your hidden state and, and really focus on that as a key decider of quality. And finally, I think one of the, the, the emerging new, new things for, for this line of work and something that's quite interesting is, What are the right test time paradigms for these models?[00:18:15] Dan Fu: How do they change relative to relative to what you might do for a standard transformer? I'll briefly end this section. So I've labeled this slide where we are yesterday because Eugene is going to talk about some new models that he released literally this morning. But as of yesterday, some of the really cool results out of the, these efficient alternative models were so AI2 trained this hybrid MOE called Jamba.[00:18:40] Dan Fu: That, that, that seems, that is currently the state of the art for these non transformer architectures. There's this NVIDIA and MIT put out this new diffusion model called SANA recently that one of their key key observations is that you can take a standard diffusion transformer diffusion model, replace the layers with linear [00:19:00] attention, and then that lets you scale to much larger much larger images, much, much Much larger sequences more efficiently.[00:19:07] Dan Fu: And and one thing that I don't think anybody would have called when a few years ago is that one of those gated SSM, gated states based models ended up on the cover of Science because a great group of folks went and trained some DNA models. So that's Michael Polley, Eric Yuen from from Stanford and the Arc Institute.[00:19:26] Dan Fu: So it's, we're really at an exciting time in 2024 where these non transformer, post transformer architectures are showing promise across a wide range. Across a wide range of, of modalities, of applications, and, and of tasks. And with that, I'll pass it on to Eugene, who can tell you a little bit about the latest and greatest with RWKV.[00:19:49] RWKV vs SSMs[00:19:49] Eugene Cheah: So, that's useful? Yeah. You're talking to here. Oh, I'm talking to here. Okay. So, yeah, two streams. Yeah. So, I think one common questions that we tend to get asked, right, is what's the difference between [00:20:00] RWKV and state space? So I think one of the key things to really understand, right the difference between the two groups, right, is that we are actually more like an open source, random internet meets academia kind of situation.[00:20:11] Eugene Cheah: Like, most of us never wrote any paper, but we, we basically look at RNNs and linear intention when intention is all you need came out, and then we decided to like, hey there is a quadratic scaling problem. Why don't we try fixing that instead? So, so, so we end up developing our own branch, but we end up sharing ideas back and forth.[00:20:30] Eugene Cheah: So, and, and we do all this actively in Discord, GitHub, etc. This was so bad for a few years, right, that basically, the average group's H index was so close to zero, right, Illuter. ai actually came in and helped us write our first paper. Great, now our H index is now three, apparently. So, so, so, but, but the thing is, like, a lot of these experiments led to results, and, and, essentially, essentially, we we took the same ideas from linear attention, [00:21:00] and we built on it.[00:21:01] Eugene Cheah: So, to take a step back into, like, how does RWKB handle its own attention mechanic and achieve the same goals of, like, O and compute, respectively, and in focus of our overall goal to make AI accessible to everyone, regardless of language, nation, or compute, that's our goal. We actually train our models primarily on over a hundred languages, which is another topic altogether.[00:21:23] Eugene Cheah: And our goal is to train to even 200 languages to cover all languages in the world. But at the same time, we work on this architecture, To lower the compute cost so that people can run it on Raspberry Pis and on anything. So, how did RWKB break the dependency of LSTM token flow? Because I think to understand architecture, right, it's probably easier to understand it from the RNN lens.[00:21:46] Eugene Cheah: Because that's where we built on. We all, we all state space kind of like try to, try to start anew and took lessons from that and say, So there's a little bit of divergence there. And AKA, this our version of linear attention. So to take step back [00:22:00] all foundation models, be it transformers or non transformers at a very high level, right?[00:22:05] Eugene Cheah: Pumps in the token. I mean, text that things into embeddings and go through a lot of layers. Generate a lot of states where the QKV cache or be iron in states or RW KB states. And outputs and embedding, they are not the same thing. And we just take more layers and more embeddings. And somehow that magically works.[00:22:23] Eugene Cheah: So, if you, if you remember your ancient RNN lessons which we, which we, which we we call best learning these days the general idea is that you have the embedding information flowing all the way up, and when, and you take that information and you flow it back down, and then you process it as part of your LSTM layers.[00:22:41] Eugene Cheah: So, this is how it generally works. Kapati is quoted saying that RNNs are actually unreasonably effective. The problem is this is not scalable. To start doing work on the second token, you need to wait for the first token. And then you need to, and likewise for the third token and fourth token, yada yada.[00:22:55] Eugene Cheah: That is CPU land, not GPU land. So, so, so, you [00:23:00] can have a H100 and you can't even use 1 percent of it. So, so that's kind of why RNNs didn't really take off in the direction that we wanted, like, billions of parameters when it comes to training. So, what did RDAP KV version 0 do? Boom. We just did the dumbest, lamest thing.[00:23:13] Eugene Cheah: Sorry, this is the bottleneck for RNN. We did the dumb thing of removing that line. And it kind of worked. It trained. It sucked, but it kind of worked. Then we were like, hey, then no one cared because the loss was crap, but how do we improve that? And that's essentially where we move forward, because if you see this kind of flow, right, you can actually get your GPU saturated quickly, where it essentially cascades respectively.[00:23:41] Eugene Cheah: So I'm just waiting for this to loop again. So it's like, once you get your first layer, your token to be computed finish. You start to cascade your compute all the way until you are, Hey, I'm using 100 percent of the GPU. So we, we worked on it, and we started going along the principle of that as long as we keep this general architecture [00:24:00] where, where we can cascade and, and be highly efficient with our architecture, nothing is sacred in our architecture.[00:24:06] Eugene Cheah: And we have done some crazy ideas. In fact, you ask us, if you ask me to explain some things in the paper, right, officially in the paper, I'll say we had this idea and we wrote it this way. The reality is someone came with a code, we tested it, it worked, and then we rationalized later. So, so the general[00:24:24] RWKV Arch[00:24:24] Eugene Cheah: The idea behind rwkbr is that we generally have two major blocks that we do.[00:24:30] Eugene Cheah: We call time mix and channel mix. And time mix generally handles handles long term memory states, where essentially, where essentially where we apply the matrix multiplication and Cilu activation functions into processing an input embedding and an output embedding. I'm oversimplifying it because this, This calculation changed every version and we have, like, version 7 right now.[00:24:50] Eugene Cheah: ChannelMix is similar to Base in the sense that it does shorter term attention, where it just looks at the sister token, or the token before it, because [00:25:00] there's a shift in the token shift matrix. I don't really want to go too much into the papers itself, because, like, we do have three papers on this.[00:25:09] Eugene Cheah: Basically, RWKB, RNN for the transformer, ERA, Ego and Pinch, RWKB, Matrix Value State. This is the updated version 5, version 6. And Goldfinch is our, is, is, is, is our hybrid model respectively. We are writing the paper already for V seven and which is, which is for R wk V seven. Called, named Goose, or architectures are named by Bird.[00:25:30] Eugene Cheah: And, I'm going to cover as well, qrwkb, and mama100k, and rwkb, and Where did that lead to? Great! Because we are all GPU poor and to be clear, like, most of this research is done, like, only on a handful H100s, which I had one Google researcher told me that was, like, his experiment budget for a single researcher.[00:25:48] Eugene Cheah: So, our entire organization has less compute than a single researcher in Google. So We, we, one of the things that we explored into was to how do we convert transformer models instead? Because [00:26:00] someone already paid that billion dollars, a million dollars onto training, so why don't we take advantage of those weights?[00:26:05] Eugene Cheah: And, and to, I believe, together AI worked on the lockets for, for the Lambda side of things, and, and we took some ideas from there as well, and we essentially did that for RWKB.[00:26:15] QWRKWv6 launch[00:26:15] Eugene Cheah: And that led to, Q RWKB6, which we just dropped today, a 32 bit instruct preview model, where we took the Quen 32 bit instruct model, freeze the feedforward layer, remove the QKB attention layer, and replace it with RWKB linear layers.[00:26:32] Eugene Cheah: So to be clear, this means we do not have the rwkv channel mix layer, we only have the time mix layer. But but once we do that, we train the rwkv layer. Important is that the feedforward layer needs to be frozen, so the new attention can be learned. And then we unfreeze the feedforward layer, and train all the layers together with a custom learning rate schedule, so that they can learn how to work together.[00:26:54] Eugene Cheah: The end result, surprisingly, And, to be honest, to the frustration of the R. W. [00:27:00] KV MOE team, which ended up releasing the model on the same day, was that, with just a few hours of training on two nodes, we managed to get it to be on par, kind of, with the original QUAN32B model. So, in fact, when the first run, right, that completely confused us, it was like, and I was telling Daniel Goldstein, Smirky, who kind of leads most of our research coordination, When you pitched me this idea, you told me at best you'll get the same level of performance.[00:27:26] Eugene Cheah: You didn't tell me the challenge and score and Winograd score will shoot up. I don't know what's happening there. But it did. MMLU score dropping, that was expected. Because if you think about it, when we were training all the layers, right, we were essentially Like, Frankenstein this thing, and we did brain damage to the feedforward network layer 2 with the new RWKB layers.[00:27:47] Eugene Cheah: But, 76%, hey, somehow it's retained, and we can probably further train this. We didn't even spend more than 3 days training this, so there's a lot more that can be done, hence the preview. This brings up [00:28:00] a big question, because We are already now in the process of converting to 7TB. We are now, this is actually extremely compute efficient to test our attention mechanic.[00:28:10] Eugene Cheah: It's like, it becomes a shortcut. We can, we are already planning to do our version 7 and our hybrid architecture for it. Because we don't need to train from scratch. And we get a really good model out of it. And the other thing that is uncomfortable to say is that because we are doing right now on the 70b is that if this scales correctly to 128k context length, I'm not even talking about a million 128, majority of enterprise workload today is just on 70b at under 32k context length.[00:28:41] Eugene Cheah: That means if this works and the benchmark matches it, It means we can replace the vast majority of current AI workload, unless you want super long context. And then sorry, can someone give us more GPUs? Because we do need the VRAM for super long context, sadly. So yeah, that's what we are working on, and essentially, [00:29:00] we are excited about this to just push it further.[00:29:02] Eugene Cheah: And this conversion process, to be clear, I don't think it's going to be exclusive to RWKB. It probably will work for Mamba as well, I don't see why not. And we will probably see more ideas, or more experiments, or more hybrids, or Yeah, like, one of the weirdest things that I wanted to say outright, and I confirmed this with the Black Mamba team and the Jamba team, which because we did the GoFinch hybrid model, is that none of us understand why a hard hybrid with a state based model to be R.[00:29:28] Eugene Cheah: QA state space and transformer performs better when, than the baseline of both. It's like, it's like when you train one, you expect, and then you replace, you expect the same results. That's our pitch. That's our claim. But somehow when we jam both together, it outperforms both. And that's like one area of emulation that, like, we only have four experiments, plus four teams, that a lot more needs to be done.[00:29:51] Eugene Cheah: But, but these are things that excite me, essentially, because that is what it's potentially we can move ahead for. Which brings us to what comes next.[00:30:00] What's next[00:30:00] [00:30:00][00:30:00] Dan Fu: So, this part is kind of just some, where we'll talk a little bit about stuff that, that we're excited about. Maybe have some wild speculation on, on what, what's, what's coming next.[00:30:12] Dan Fu: And, of course this is also the part that will be more open to questions. So, a couple things that, that I'm excited about is continued hardware model co design for, for these models. So one of the things that we've put out recently is this library called ThunderKittens. It's a CUDA library.[00:30:29] Dan Fu: And one of the things that, that we found frustrating is every time that we built one of these new architectures, and I'm sure you had the exact same experience, we'd have to go and spend two months in CUDA land, like writing these, these new efficient things. And. If we decided to change one thing in PyTorch, like one line of PyTorch code is like a week of CUDA code at least.[00:30:47] Dan Fu: So one of our goals with, with a library like Thunderkitten, so we, we just broke down what are the key principles, what are the key hardware things what are the key, Compute pieces that you get from the hardware. So for example on [00:31:00] H100 everything is really revolves around a warp group matrix multiply operation.[00:31:06] Dan Fu: So you really want your operation to be able to split into relatively small matrix, matrix multiply operations. So like multiplying two 64 by 64 matrices, for example. And so if you know that ahead of time when you're designing your model, that probably gives you you know, some information about how you set the state sizes, how you set the update, how you set the update function.[00:31:27] Dan Fu: So with Thunderkittens we basically built a whole library just around this basic idea that all your basic compute primitives should not be a float, but it should be a matrix, and everything should just be matrix compute. And we've been using that to, to try to both re implement some existing architectures, and also start to design code.[00:31:44] Dan Fu: Some new ones that are really designed with this core with a tensor core primitive in mind. Another thing that that we're, that at least I'm excited about is we, over the last four or five years, we've really been looking at language models as the next thing. But if you've been paying [00:32:00] attention to Twitter there's been a bunch of new next generation models that are coming out.[00:32:04] Dan Fu: So there, there are. So, video generation models that can run real time, that are supported by your mouse and your keyboard, that I'm told if you play with them that, you know, that they only have a few seconds of memory. Can we take that model, can we give it a very long context length so that you could actually maybe generate an entire game state at a time?[00:32:25] Dan Fu: What does that look like for the model? You're certainly not going to do a giant quadratic attention computation to try to run that. Maybe, maybe use some of these new models, or some of these new video generation models that came out. So Sora came out I don't know, two days ago now. But with super long queue times and super long generation times.[00:32:43] Dan Fu: So that's probably a quadratic attention operation at the, at the bottom of it. What if we could remove that and get the same quality, but a lot faster generation time? Or some of the demos that we saw from Paige earlier today. You know, if I have a super long conversation with my [00:33:00] Gemini bot, what if I wanted to remember everything that it's seen in the last week?[00:33:06] Dan Fu: I mean, maybe you don't for personal reasons, but what if I did, you know? What does that mean for the architecture? And I think, you know, that's certainly something I'm pretty excited about. I'm sure you're excited about it too. So, I think we were supposed to have some hot takes, but I honestly don't remember what our hot takes were.[00:33:21] Hot Takes - does anyone really need long context?[00:33:21] Eugene Cheah: Yeah, including the next slide. Hot takes, yes, these are our[00:33:25] Dan Fu: hot takes.[00:33:25] Eugene Cheah: I think the big one on Twitter that we saw, that we shared, was the question is like, is RAG relevant? In the case of, like, the future of, like, state based models?[00:33:38] Dan Fu: Let's see, I haven't played too much with RAG. But when I have. I'll say I found it was a little bit challenging to do research on it because we had this experience over and over again, where you could have any, an embedding model of any quality, so you could have a really, really bad embedding model, or you could have a really, really [00:34:00] good one, By any measure of good.[00:34:03] Dan Fu: And for the final RAG application, it kind of didn't matter. That's what I'll say about RAG while I'm being recorded. I know it doesn't actually answer the question, but[00:34:13] Eugene Cheah: Yeah, so I think a lot of folks are like, extremely excited of the idea of RWKB or State Space potentially having infinite context.[00:34:21] Eugene Cheah: But I think the reality is that when we say infinite context, we just mean a different kind of infinite context, or you, or as it's previously covered, you need to test the model differently. So, think of it more along the lines of the human. Like, I don't remember what I ate for breakfast yesterday.[00:34:37] Eugene Cheah: Yeah, that's the statement that I'll say. And And we humans are not quadratic transformers. If we did, if let's say we increased our brain size for every second we live, we would have exploded by the time we are 5 years old or something like that. And, and I think, I think basically fundamentally for us, right, be it whether we, regardless of whether RWKB, statespace, XLSTM, [00:35:00] etc, our general idea is that instead of that expanding state, that increase in computational cost, what if we have a fixed state size?[00:35:08] Eugene Cheah: And Information theory detects that that fixed state size will have a limit. Just how big of a limit is a question, like, we, like, RWKB is running at 40 megabytes for, for its state. Its future version might run into 400 megabytes. That is like millions of tokens in, if you're talking about mathematically, the maximum possibility.[00:35:29] Eugene Cheah: It's just that I guess we were all more inefficient about it, so maybe we hit 100, 000. And that's kind of like the work we are doing, trying to like push it and maximize it. And that's where the models will start differing, because it will choose to forget things, it will choose to remember things. And that's why I think that there might be some element of right, but it may not be the same right.[00:35:49] Eugene Cheah: It may be the model learn things, and it's like, hmm, I can't remember that, that article. Let me do a database search, to search. Just like us humans, when we can't remember the article in the company. We do a search on Notion. [00:36:00][00:36:00] Dan Fu: I think something that would be really interesting is if you could have facts that are, so right now, the one intuition about language models is that all those parameters are around just to store random facts about the world.[00:36:14] Dan Fu: And this intuition comes from the observation that if you take a really small language model, it can do things like talk to you, or kind of has like the The style of conversation, it can learn that, but where it will usually fall over compared to a much larger one is it'll just be a lot less factual about things that it knows or that it can do.[00:36:32] Dan Fu: But that points to all those weights that we're spending, all that SGD that we're spending to train these models are just being used to store facts. And we have things like databases that are pretty good at storing facts. So I think one thing that would be really interesting is if we could actually have some sort of outside data store that a language model can can look at that that maybe is you know, has has some sort of gradient descent in it, but but would be quite interesting.[00:36:58] Dan Fu: And then maybe you could edit it, delete [00:37:00] facts, you know, change who's president so that it doesn't, it doesn't get lost.[00:37:04] Vibhu: Can we open up Q& A and hot takes for the audience? I have a hot take Q& A. Do these scale? When, when 405B state space model, RAG exists, no one does long context, who's throwing in 2 million token questions, hot takes?[00:37:24] Dan Fu: The, the who's throwing in 2 million token question, I think, is, is a really good question. So I actually, I was going to offer that as a hot take. I mean, my hot take was going to be that long context doesn't matter. I know I just gave a whole talk about it, but you know, what, what's the point of doing research if you can't, you know, play both sides.[00:37:40] Dan Fu: But I think one of the, so I think for both of us, the reason that we first got into this was just from the first principled questions of there's this quadratic thing. Clearly intelligence doesn't need to be quadratic. What is going on? Can we understand it better? You know, since then it's kind of turned into a race, which has [00:38:00] been exciting to watch, like, how much context you can take in.[00:38:03] Dan Fu: But I think it's right. Nobody is actually putting in a two million context prompt into these models. And, and, you know, if they are, maybe we can go, go You know, design a better model to do that particular thing. Yeah, what do you think about that? So you've also been working on this. Do you think long context matters?[00:38:19] Eugene Cheah: So I'm going to burn a bit. How many of you remember the news of Google Gemini supporting 3 million contacts, right? Raise your hand.[00:38:28] Vibhu: Yeah, 2 million.[00:38:29] Eugene Cheah: Oh, it's 2 million.[00:38:31] Eugene Cheah: Yeah, how many of you actually tried that? See?[00:38:34] Vibhu: I use it a lot. You? You work for MindsTV. I use it a lot.[00:38:41] Eugene Cheah: So, for some people that has used, and I think, I think that's the, that's might be, like, this is where my opinion starts to differ, because I think the big labs may have a bigger role in this, because Like, even for RWKB, even when we train non contacts, the reason why I say VRAM is a problem is that because when we did the, we need to backprop [00:39:00] against the states, we actually need to maintain the state in between the tokens by the token length.[00:39:05] Eugene Cheah: So that means we need to actually roll out the whole 1 million contacts if we are actually training 1 million. Which is the same for transformers, actually, but it just means we don't magically reuse the VRAM consumption in the training time space. So that is one of the VRAM bottlenecks, and I'm neither OpenAI nor Google, so donate GPUs if you have too much of them.[00:39:27] Eugene Cheah: But then, putting it back to another paradigm, right, is that I think O1 style reasoning might be actually pushing that direction downwards. In my opinion, this is my partial hot take is that if, let's say you have a super big model, And let's say you have a 70B model that may take double the tokens, but gets the same result.[00:39:51] Eugene Cheah: Strictly speaking, a 70B, and this is even for transformer or non transformer, right? We we'll take less less resources than that 400 B [00:40:00] model, even if it did double the amount thinking. And if that's the case, and we are still all trying to figure this out, maybe the direction for us is really getting the sub 200 B to be as fast as efficient as possible.[00:40:11] Eugene Cheah: We a very efficient architecture that some folks happen to be working on to, to just reason it out over larger and larger context thing.[00:40:20] Question: Yeah. One thing I'm super interested in is. Models that can watch forever? Obviously you cannot train something on infinite context length. How are y'all thinking about that, where you run on a much longer context length than is possible to train on?[00:40:38] Dan Fu: Yeah, it's a, it's a great question. So I think when I think you guys probably had tweets along these lines, too. When we first started doing these things, because these are all recurrent models in theory you could just run it forever. You could just run it forever. And at the very least it won't, it won't like error out on your crash.[00:40:57] Dan Fu: There's another question of whether it can actually [00:41:00] use what it's seen in that infinite context. And I think there, so one place where probably the research and architectures ran faster Then another research is actually the benchmarks for long context. So you turn it on forever. You want to do everything or watch everything.[00:41:16] Dan Fu: What is it that you actually wanted to do? Can we actually build some benchmarks for that? Then measure what's happening. And then ask the question, can the models do it? Is there something else that they need? Yeah, I think that if I were to turn back the clock to 2022, that's probably one of the things I would have done differently, which would have been actually get some long context benchmarks out at the same time as we started pushing context length on all these models.[00:41:41] Eugene Cheah: I will also say the use case. So like, I think we both agree that there's no Infinite memory and the model needs to be able to learn and decide. I think what we have observed for, I think this also fits the state space model, is that one of the key advantages of this alternate attention mechanic that is not based on token position is that the model don't suddenly become crazy when you go past the [00:42:00] 8k training context tank, or a million context tank.[00:42:03] Eugene Cheah: It's actually still stable. It's still able to run, it's still able to rationalize. It just starts forgetting things. But some of these things are still there in latent memory. Some of these things are still somewhat there. That's the whole point of why reading twice works. Things like that. And one of the biggest pushes in this direction is that I think both Statespace and RWKB have Separate papers by other researchers where they use this architecture for time series data.[00:42:26] Eugene Cheah: Weather modeling. So, you are not asking what was the weather five days ago. You're asking what's the weather tomorrow based on the infinite length that we, as long as this Earth and the computer will keep running. So, so, and they found that it is like, better than existing, like, transformer or existing architecture in modeling this weather data.[00:42:47] Eugene Cheah: Control for the param size and stuff. I'm quite sure there are people with larger models. So, so there are things that, that in this case, right, there is future applications if your question is just what's next and not what's 10 years ago.[00:42:59] Dan Fu: Thanks so [00:43:00] much for having us. Get full access to Latent Space at www.latent.space/subscribe

Happy holidays! We’ll be sharing snippets from Latent Space LIVE! through the break bringing you the best of 2024! We want to express our deepest appreciation to event sponsors AWS, Daylight Computer, Thoth.ai, StrongCompute, Notable Capital, and most of all our LS supporters who helped fund the venue and A/V production!For NeurIPS last year we did our standard conference podcast coverage interviewing selected papers (that we have now also done for ICLR and ICML), however we felt that we could be doing more to help AI Engineers 1) get more industry-relevant content, and 2) recap 2024 year in review from experts. As a result, we organized the first Latent Space LIVE!, our first in person miniconference, at NeurIPS 2024 in Vancouver.Since Nathan Lambert ( Interconnects ) joined us for the hit RLHF 201 episode at the start of this year, it is hard to overstate how much Open Models have exploded this past year. In 2023 only five names were playing in the top LLM ranks, Mistral, Mosaic's MPT, TII UAE's Falcon, Yi from Kai-Fu Lee's 01.ai, and of course Meta's Llama 1 and 2. This year a whole cast of new open models have burst on the scene, from Google's Gemma and Cohere's Command R, to Alibaba's Qwen and Deepseek models, to LLM 360 and DCLM and of course to the Allen Institute's OLMo, OL MOE, Pixmo, Molmo, and Olmo 2 models. We were honored to host Luca Soldaini, one of the research leads on the Olmo series of models at AI2.Pursuing Open Model research comes with a lot of challenges beyond just funding and access to GPUs and datasets, particularly the regulatory debates this year across Europe, California and the White House. We also were honored to hear from and Sophia Yang, head of devrel at Mistral, who also presented a great session at the AI Engineer World's Fair Open Models track!Full Talk on YouTubePlease like and subscribe!Timestamps* 00:00 Welcome to Latent Space Live * 00:12 Recap of 2024: Best Moments and Keynotes * 01:22 Explosive Growth of Open Models in 2024 * 02:04 Challenges in Open Model Research * 02:38 Keynote by Luca Soldani: State of Open Models * 07:23 Significance of Open Source AI Licenses * 11:31 Research Constraints and Compute Challenges * 13:46 Fully Open Models: A New Trend * 27:46 Mistral's Journey and Innovations * 32:57 Interactive Demo: Lachat Capabilities * 36:50 Closing Remarks and NetworkingTranscriptSession3Audio[00:00:00] AI Charlie: Welcome to Latent Space Live, our first mini conference held at NeurIPS 2024 in Vancouver. This is Charlie, your AI co host. As a special treat this week, we're recapping the best of 2024 going domain by domain. We sent out a survey to the over 900 of you who told us what you wanted, and then invited the best speakers in the latent space network to cover each field.[00:00:28] AI Charlie: 200 of you joined us in person throughout the day, with over 2, 200 watching live online. Our next keynote covers the state of open models in 2024, with Luca Soldani and Nathan Lambert of the Allen Institute for AI, with a special appearance from Dr. Sophia Yang of Mistral. Our first hit episode of 2024 was with Nathan Lambert on RLHF 201 back in January.[00:00:57] AI Charlie: Where he discussed both reinforcement learning for language [00:01:00] models and the growing post training and mid training stack with hot takes on everything from constitutional AI to DPO to rejection sampling and also previewed the sea change coming to the Allen Institute. And to Interconnects, his incredible substack on the technical aspects of state of the art AI training.[00:01:18] AI Charlie: We highly recommend subscribing to get access to his Discord as well. It is hard to overstate how much open models have exploded this past year. In 2023, only five names were playing in the top LLM ranks. Mistral, Mosaics MPT, and Gatsby. TII UAE's Falcon, Yi, from Kaifu Lee's 01. ai, And of course, Meta's Lama 1 and 2.[00:01:43] AI Charlie: This year, a whole cast of new open models have burst on the scene. From Google's Jemma and Cohere's Command R, To Alibaba's Quen and DeepSeq models, to LLM360 and DCLM, and of course, to the Allen Institute's OLMO, [00:02:00] OLMOE, PIXMO, MOLMO, and OLMO2 models. Pursuing open model research comes with a lot of challenges beyond just funding and access to GPUs and datasets, particularly the regulatory debates this year across Europe.[00:02:14] AI Charlie: California and the White House. We also were honored to hear from Mistral, who also presented a great session at the AI Engineer World's Fair Open Models track. As always, don't forget to check the show notes for the YouTube link to their talk, as well as their slides. Watch out and take care.[00:02:35] Luca Intro[00:02:35] Luca Soldaini: Cool. Yeah, thanks for having me over. I'm Luca. I'm a research scientist at the Allen Institute for AI. I threw together a few slides on sort of like a recap of like interesting themes in open models for, for 2024. Have about maybe 20, 25 minutes of slides, and then we can chat if there are any questions.[00:02:57] Luca Soldaini: If I can advance to the next slide. [00:03:00] Okay, cool. So I did the quick check of like, to sort of get a sense of like, how much 2024 was different from 2023. So I went on Hugging Face and sort of get, tried to get a picture of what kind of models were released in 2023 and like, what do we get in 2024?[00:03:16] Luca Soldaini: 2023 we get, we got things like both LLAMA 1 and 2, we got Mistral, we got MPT, Falcon models, I think the YI model came in at the end. Tail end of the year. It was a pretty good year. But then I did the same for 2024. And it's actually quite stark difference. You have models that are, you know, reveling frontier level.[00:03:38] Luca Soldaini: Performance of what you can get from closed models from like Quen, from DeepSeq. We got Llama3. We got all sorts of different models. I added our own Olmo at the bottom. There's this growing group of like, Fully open models that I'm going to touch on a little bit later. But you know, just looking at the slides, it feels like 2024 [00:04:00] was just smooth sailing, happy knees, much better than previous year.[00:04:04] Luca Soldaini: And you know, you can plot you can pick your favorite benchmark Or least favorite, I don't know, depending on what point you're trying to make. And plot, you know, your closed model, your open model and sort of spin it in ways that show that, oh, you know open models are much closer to where closed models are today versus to Versus last year where the gap was fairly significant.[00:04:29] Luca Soldaini: So one thing that I think I don't know if I have to convince people in this room, but usually when I give this talks about like open models, there is always like this background question in, in, in people's mind of like, why should we use open models? APIs argument, you know, it's, it's. Just an HTTP request to get output from a, from one of the best model out there.[00:04:53] Luca Soldaini: Why do I have to set up infra and use local models? And there are really like two answer. There is the more [00:05:00] researchy answer for this, which is where it might be. Background lays, which is just research. If you want to do research on language models, research thrives on, on open models, there is like large swath of research on modeling, on how these models behave on evaluation and inference on mechanistic interpretability that could not happen at all if you didn't have open models they're also for AI builders, they're also like.[00:05:30] Luca Soldaini: Good use cases for using local models. You know, you have some, this is like a very not comprehensive slides, but you have things like there are some application where local models just blow closed models out of the water. So like retrieval, it's a very clear example. We might have like constraints like Edge AI applications where it makes sense.[00:05:51] Luca Soldaini: But even just like in terms of like stability, being able to say this model is not changing under the hood. It's, there's plenty of good cases for, [00:06:00] for open models. And the community is just not models. Is I stole this slide from one of the Quent2 announcement blog posts. But it's super cool to see like how much tech exists around open models and serving them on making them efficient and hosting them.[00:06:18] Luca Soldaini: It's pretty cool. And so. It's if you think about like where the term opens come from, comes from like the open source really open models meet the core tenants of, of open, of open source specifically when it comes around collaboration, there is truly a spirit, like through these open models, you can build on top of other people.[00:06:41] Luca Soldaini: innovation. We see a lot of these even in our own work of like, you know, as we iterate in the various versions of Alma it's not just like every time we collect from scratch all the data. No, the first step is like, okay, what are the cool data sources and datasets people have put [00:07:00] together for language model for training?[00:07:01] Luca Soldaini: Or when it comes to like our post training pipeline We one of the steps is you want to do some DPO and you use a lot of outputs of other models to improve your, your preference model. So it's really having like an open sort of ecosystem benefits and accelerates the development of open models.[00:07:23] The Definition of Open Models[00:07:23] Luca Soldaini: One thing that we got in 2024, which is not a specific model, but I thought it was really significant, is we first got we got our first open source AI definition. So this is from the open source initiative they've been generally the steward of a lot of the open source licenses when it comes to software and so they embarked on this journey in trying to figure out, okay, How does a license, an open source license for a model look like?[00:07:52] Luca Soldaini: Majority of the work is very dry because licenses are dry. So I'm not going to walk through the license step by [00:08:00] step, but I'm just going to pick out one aspect that is very good and then one aspect that personally feels like it needs improvement on the good side. This this open source AI license actually.[00:08:13] Luca Soldaini: This is very intuitive. If you ever build open source software and you have some expectation around like what open source looks like for software for, for AI, sort of matches your intuition. So, the weights need to be fairly available the code must be released with an open source license and there shouldn't be like license clauses that block specific use cases.[00:08:39] Luca Soldaini: So. Under this definition, for example, LLAMA or some of the QUEN models are not open source because the license says you can't use this model for this or it says if you use this model you have to name the output this way or derivative needs to be named that way. Those clauses don't meet open source [00:09:00] definition and so they will not be covered.[00:09:02] Luca Soldaini: The LLAMA license will not be covered under the open source definition. It's not perfect. One of the thing that, um, internally, you know, in discussion with with OSI, we were sort of disappointed is around the language. For data. So you might imagine that an open source AI model means a model where the data is freely available.[00:09:26] Luca Soldaini: There were discussion around that, but at the end of the day, they decided to go with a softened stance where they say a model is open source if you provide sufficient detail information. On how to sort of replicate the data pipeline. So you have an equivalent system, sufficient, sufficiently detailed.[00:09:46] Luca Soldaini: It's very, it's very fuzzy. Don't like that. An equivalent system is also very fuzzy. And this doesn't take into account the accessibility of the process, right? It might be that you provide enough [00:10:00] information, but this process costs, I don't know, 10 million to do. Now the open source definition. Like, any open source license has never been about accessibility, so that's never a factor in open source software, how accessible software is.[00:10:14] Luca Soldaini: I can make a piece of open source, put it on my hard drive, and never access it. That software is still open source, the fact that it's not widely distributed doesn't change the license, but practically there are expectations of like, what we want good open sources to be. So, it's, It's kind of sad to see that the data component in this license is not as, as, Open as some of us would like would like it to be.[00:10:40] Challenges for Open Models[00:10:40] Luca Soldaini: and I linked a blog post that Nathan wrote on the topic that it's less rambly and easier to follow through. One thing that in general, I think it's fair to say about the state of open models in 2024 is that we know a lot more than what we knew in, [00:11:00] in 2023. Like both on the training data, like And the pre training data you curate on like how to do like all the post training, especially like on the RL side.[00:11:10] Luca Soldaini: You know, 2023 was a lot of like throwing random darts at the board. I think 2024, we have clear recipes that, okay, don't get the same results as a closed lab because there is a cost in, in actually matching what they do. But at least we have a good sense of like, okay, this is, this is the path to get state of the art language model.[00:11:31] Luca Soldaini: I think that one thing that it's a downside of 2024 is that I think we are more research constrained in 2023. It feels that, you know, the barrier for compute that you need to, to move innovation along as just being right rising and rising. So like, if you go back to this slide, there is now this, this cluster of models that are sort of released by the.[00:11:57] Luca Soldaini: Compute rich club. Membership is [00:12:00] hotly debated. You know, some people don't want to be. Called the rich because it comes to expectations. Some people want to be called rich, but I don't know, there's debate, but like, these are players that have, you know, 10, 000, 50, 000 GPUs at minimum. And so they can do a lot of work and a lot of exploration and improving models that it's not very accessible.[00:12:21] Luca Soldaini: To give you a sense of like how I personally think about. Research budget for each part of the, of the language model pipeline is like on the pre training side, you can maybe do something with a thousand GPUs, really you want 10, 000. And like, if you want real estate of the art, you know, your deep seek minimum is like 50, 000 and you can scale to infinity.[00:12:44] Luca Soldaini: The more you have, the better it gets. Everyone on that side still complains that they don't have enough GPUs. Post training is a super wide sort of spectrum. You can do as little with like eight GPUs as long as you're able to [00:13:00] run, you know, a good version of, say, a LLAMA model, you can do a lot of work there.[00:13:05] Luca Soldaini: You can scale a lot of the methodology, just like scales with compute, right? If you're interested in you know, your open replication of what OpenAI's O1 is you're going to be on the 10K spectrum of our GPUs. Inference, you can do a lot with very few resources. Evaluation, you can do a lot with, well, I should say at least one GPUs if you want to evaluate GPUs.[00:13:30] Luca Soldaini: Open models but in general, like if you are, if you care a lot about intervention to do on this model, which it's my prefer area of, of research, then, you know, the resources that you need are quite, quite significant. Yeah. One other trends that has emerged in 2024 is this cluster of fully open models.[00:13:54] Luca Soldaini: So Omo the model that we built at ai, two being one of them and you know, it's nice [00:14:00] that it's not just us. There's like a cluster of other mostly research efforts who are working on this. And so it's good to to give you a primer of what like fully open means. So fully open, the easy way to think about it is instead of just releasing a model checkpoint that you run, you release a full recipe so that other people working on it.[00:14:24] Luca Soldaini: Working on that space can pick and choose whatever they want from your recipe and create their own model or improve on top of your model. You're giving out the full pipeline and all the details there instead of just like the end output. So I pull up the screenshot from our recent MOE model.[00:14:43] Luca Soldaini: And like for this model, for example, we released the model itself. Data that was trained on, the code, both for training and inference all the logs that we got through the training run, as well as every intermediate checkpoint and like the fact that you release different part of the pipeline [00:15:00] allows others to do really cool things.[00:15:02] Luca Soldaini: So for example, this tweet from early this year from folks in news research they use our pre training data to do a replication of the BitNet paper in the open. So they took just a Really like the initial part of a pipeline and then the, the thing on top of it. It goes both ways.[00:15:21] Luca Soldaini: So for example, for the Olmo2 model a lot of our pre trained data for the first stage of pre training was from this DCLM initiative that was led by folks Ooh, a variety of ins a variety of institutions. It was a really nice group effort. But you know, for When it was nice to be able to say, okay, you know, the state of the art in terms of like what is done in the open has improved.[00:15:46] AI2 Models - Olmo, Molmo, Pixmo etc[00:15:46] Luca Soldaini: We don't have to like do all this work from scratch to catch up the state of the art. We can just take it directly and integrate it and do our own improvements on top of that. I'm going to spend a few minutes doing like a [00:16:00] shameless plug for some of our fully open recipes. So indulge me in this.[00:16:05] Luca Soldaini: So a few things that we released this year was, as I was mentioning, there's OMOE model which is, I think still is state of the art MOE model in its size class. And it's also. Fully open, so every component of this model is available. We released a multi modal model called Molmo. Molmo is not just a model, but it's a full recipe of how you go from a text only model to a multi modal model, and we apply this recipe on top of Quent checkpoints, on top of Olmo checkpoints, as well as on top of OlmoE.[00:16:37] Luca Soldaini: And I think there'd be a replication doing that on top of Mistral as well. The post training side we recently released 2. 0. 3. Same story. This is a recipe on how you go from a base model to A state of the art post training model. We use the Tulu recipe on top of Olmo, on top of Llama, and then there's been open replication effort [00:17:00] to do that on top of Quen as well.[00:17:02] Luca Soldaini: It's really nice to see like, you know, when your recipe sort of, it's kind of turnkey, you can apply it to different models and it kind of just works. And finally, the last thing we released this year was Olmo 2, which so far is the best state of the art. Fully open language model a Sera combines aspect from all three of these previous models.[00:17:22] Luca Soldaini: What we learn on the data side from MomoE and what we learn on like making models that are easy to adapt from the Momo project and the Tulu project. I will close with a little bit of reflection of like ways this, this ecosystem of open models like it's not all roses. It's not all happy. It feels like day to day, it's always in peril.[00:17:44] Luca Soldaini: And, you know, I talked a little bit about like the compute issues that come with it. But it's really not just compute. One thing that is on top of my mind is due to like the environment and how you know, growing feelings about like how AI is treated. [00:18:00] It's actually harder to get access to a lot of the data that was used to train a lot of the models up to last year.[00:18:06] Luca Soldaini: So this is a screenshot from really fabulous work from Shane Longpre who's, I think is in Europe about Just access of like diminishing access to data for language model pre training. So what they did is they went through every snapshot of common crawl. Common crawl is this publicly available scrape of the, of a subset of the internet.[00:18:29] Luca Soldaini: And they looked at how For any given website whether a website that was accessible in say 2017, what, whether it was accessible or not in 2024. And what they found is as a reaction to like the close like of the existence of closed models like OpenAI or Cloud GPT or Cloud a lot of content owners have blanket Blocked any type of crawling to your website.[00:18:57] Luca Soldaini: And this is something that we see also internally at [00:19:00] AI2. Like one project that we started this year is we wanted to, we wanted to understand, like, if you're a good citizen of the internet and you crawl following sort of norms and policy that have been established in the last 25 years, what can you crawl?[00:19:17] Luca Soldaini: And we found that there's a lot of website where. The norms of how you express preference of whether to crawl your data or not are broken. A lot of people would block a lot of crawling, but do not advertise that in RobustDXT. You can only tell that they're crawling, that they're blocking you in crawling when you try doing it.[00:19:37] Luca Soldaini: Sometimes you can't even crawl the robots. txt to, to check whether you're allowed or not. And then a lot of websites there's, there's like all these technologies that historically have been, have existed to make websites serving easier such as Cloudflare or DNS. They're now being repurposed for blocking AI or any type of crawling [00:20:00] in a way that is Very opaque to the content owners themselves.[00:20:04] Luca Soldaini: So, you know, you go to these websites, you try to access them and they're not available and you get a feeling it's like, Oh, someone changed, something changed on the, on the DNS side that it's blocking this and likely the content owner has no idea. They're just using a Cloudflare for better, you know, load balancing.[00:20:25] Luca Soldaini: And this is something that was sort of sprung on them with very little notice. And I think the problem is this, this blocking or ideas really, it impacts people in different ways. It disproportionately helps companies that have a headstart, which are usually the closed labs and it hurts incoming newcomer players where either have now to do things in a sketchy way or you're never going to get that content that the closed lab might have.[00:20:54] Luca Soldaini: So there's a lot, it was a lot of coverage. I'm going to plug Nathan's blog post again. That is, [00:21:00] that I think the title of this one is very succinct which is like, we're actually not, You know, before thinking about running out of training data, we're actually running out of open training data. And so if we want better open models they should be on top of our mind.[00:21:13] Regulation and Lobbying[00:21:13] Luca Soldaini: The other thing that has emerged is that there is strong lobbying efforts on trying to define any kind of, AI as like a new extremely risky and I want to be precise here. Like the problem is now, um, like the problem is not not considering the risk of this technology. Every technology has risks that, that should always be considered.[00:21:37] Luca Soldaini: The thing that it's like to me is sorry, is ingenious is like just putting this AI on a pedestal and calling it like, An unknown alien technology that has like new and undiscovered potentials to destroy humanity. When in reality, all the dangers I think are rooted in [00:22:00] dangers that we know from existing software industry or existing issues that come with when using software on on a lot of sensitive domains, like medical areas.[00:22:13] Luca Soldaini: And I also noticed a lot of efforts that have actually been going on and trying to make this open model safe. I pasted one here from AI2, but there's actually like a lot of work that has been going on on like, okay, how do you make, if you're distributing this model, Openly, how do you make it safe?[00:22:31] Luca Soldaini: How, what's the right balance between accessibility on open models and safety? And then also there's annoying brushing of sort of concerns that are then proved to be unfounded under the rug. You know, if you remember the beginning of this year, it was all about bio risk of these open models.[00:22:48] Luca Soldaini: The whole thing fizzled because as being Finally, there's been like rigorous research, not just this paper from Cohere folks, but it's been rigorous research showing [00:23:00] that this is really not a concern that we should be worried about. Again, there is a lot of dangerous use of AI applications, but this one was just like, A lobbying ploy to just make things sound scarier than they actually are.[00:23:15] Luca Soldaini: So I got to preface this part. It says, this is my personal opinion. It's not my employer, but I look at things like the SP 1047 from, from California. And I think we kind of dodged a bullet on, on this legislation. We, you know, the open source community, a lot of the community came together at the last, sort of the last minute and did a very good effort trying to explain all the negative impact of this bill.[00:23:43] Luca Soldaini: But There's like, I feel like there's a lot of excitement on building these open models or like researching on these open models. And lobbying is not sexy it's kind of boring but it's sort of necessary to make sure that this ecosystem can, can really [00:24:00] thrive. This end of presentation, I have Some links, emails, sort of standard thing in case anyone wants to reach out and if folks have questions or anything they wanted to discuss.[00:24:13] Luca Soldaini: Is there an open floor? I think we have Sophia[00:24:16] swyx: who wants to who one, one very important open model that we haven't covered is Mistral. Ask her on this slide. Yeah, yeah. Well, well, it's nice to have the Mistral person talk recap the year in Mistral. But while Sophia gets set up, does anyone have like, just thoughts or questions about the progress in this space?[00:24:32] Questions - Incentive Alignment[00:24:32] swyx: Do you always have questions?[00:24:34] Quesiton: I'm very curious how we should build incentives to build open models, things like Francois Chollet's ArcPrize, and other initiatives like that. What is your opinion on how we should better align incentives in the community so that open models stay open?[00:24:49] Luca Soldaini: The incentive bit is, like, really hard.[00:24:51] Luca Soldaini: Like, even It's something that I actually, even we think a lot about it internally because like building open models is risky. [00:25:00] It's very expensive. And so people don't want to take risky bets. I think the, definitely like the challenges like our challenge, I think those are like very valid approaches for it.[00:25:13] Luca Soldaini: And then I think in general, promoting, building, so, any kind of effort to participate in this challenge, in those challenges, if we can promote doing that on top of open models and sort of really lean into like this multiplier effect, I think that is a good way to go. If there were more money for that.[00:25:35] Luca Soldaini: For efforts like research efforts around open models. There's a lot of, I think there's a lot of investments in companies that at the moment are releasing their model in the open, which is really cool. But it's usually more because of commercial interest and not wanting to support this, this like open models in the longterm, it's a really hard problem because I think everyone is operating sort of [00:26:00] in what.[00:26:01] Luca Soldaini: Everyone is at their local maximum, right? In ways that really optimize their position on the market. Global maximum is harder to achieve.[00:26:11] Question2: Can I ask one question? No.[00:26:12] Luca Soldaini: Yeah.[00:26:13] Question2: So I think one of the gap between the closed and open source models is the mutability. So the closed source models like chat GPT works pretty good on the low resource languages, which is not the same on the open, open source models, right?[00:26:27] Question2: So is it in your plan to improve on that?[00:26:32] Luca Soldaini: I think in general,[00:26:32] Luca Soldaini: yes, is I think it's. I think we'll see a lot of improvements there in, like, 2025. Like, there's groups like, Procurement English on the smaller side that are already working on, like, better crawl support, multilingual support. I think what I'm trying to say here is you really want to be experts.[00:26:54] Luca Soldaini: who are actually in those countries that teach those languages to [00:27:00] participate in the international community. To give you, like, a very easy example I'm originally from Italy. I think I'm terribly equipped to build a model that works well in Italian. Because one of the things you need to be able to do is having that knowledge of, like, okay, how do I access, you know, how Libraries, or content that is from this region that covers this language.[00:27:23] Luca Soldaini: I've been in the US long enough that I no longer know. So, I think that's the efforts that folks in Central Europe, for example, are doing. Around like, okay, let's tap into regional communities. To get access you know, to bring in collaborators from those areas. I think it's going to be, like, very crucial for getting products there.[00:27:46] Mistral intro[00:27:46] Sophia Yang: Hi everyone. Yeah, I'm super excited to be here to talk to you guys about Mistral. A really short and quick recap of what we have done, what kind of models and products we have released in the [00:28:00] past year and a half. So most of you We have already known that we are a small startup funded about a year and a half ago in Paris in May, 2003, it was funded by three of our co founders, and in September, 2003, we released our first open source model, Mistral 7b yeah, how, how many of you have used or heard about Mistral 7b?[00:28:24] Sophia Yang: Hey, pretty much everyone. Thank you. Yeah, it's our Pretty popular and community. Our committee really loved this model, and in December 23, we, we released another popular model with the MLE architecture Mr. A X seven B and oh. Going into this year, you can see we have released a lot of things this year.[00:28:46] Sophia Yang: First of all, in February 2004, we released MrSmall, MrLarge, LeChat, which is our chat interface, I will show you in a little bit. We released an embedding model for, you [00:29:00] know, converting your text into embedding vectors, and all of our models are available. The, the big cloud resources. So you can use our model on Google cloud, AWS, Azure Snowflake, IBM.[00:29:16] Sophia Yang: So very useful for enterprise who wants to use our model through cloud. And in April and May this year, we released another powerful open source MOE model, AX22B. And we also released our first code. Code Model Coastal, which is amazing at 80 plus languages. And then we provided another fine tuning service for customization.[00:29:41] Sophia Yang: So because we know the community love to fine tune our models, so we provide you a very nice and easy option for you to fine tune our model on our platform. And also we released our fine tuning code base called Menstrual finetune. It's open source, so feel free to take it. Take a look and.[00:29:58] Sophia Yang: More models. [00:30:00] On July 2, November this year, we released many, many other models. First of all is the two new small, best small models. We have Minestra 3B great for Deploying on edge devices we have Minstrel 8B if you used to use Minstrel 7B, Minstrel 8B is a great replacement with much stronger performance than Minstrel 7B.[00:30:25] Sophia Yang: We also collaborated with NVIDIA and open sourced another model, Nemo 12B another great model. And Just a few weeks ago, we updated Mistral Large with the version 2 with the updated, updated state of the art features and really great function calling capabilities. It's supporting function calling in LatentNate.[00:30:45] Sophia Yang: And we released two multimodal models Pixtral 12b. It's this open source and Pixtral Large just amazing model for, models for not understanding images, but also great at text understanding. So. Yeah, a [00:31:00] lot of the image models are not so good at textual understanding, but pixel large and pixel 12b are good at both image understanding and textual understanding.[00:31:09] Sophia Yang: And of course, we have models for research. Coastal Mamba is built on Mamba architecture and MathRoll, great with working with math problems. So yeah, that's another model.[00:31:29] Sophia Yang: Here's another view of our model reference. We have several premier models, which means these models are mostly available through our API. I mean, all of the models are available throughout our API, except for Ministry 3B. But for the premier model, they have a special license. Minstrel research license, you can use it for free for exploration, but if you want to use it for enterprise for production use, you will need to purchase a license [00:32:00] from us.[00:32:00] Sophia Yang: So on the top row here, we have Minstrel 3b and 8b as our premier model. Minstrel small for best, best low latency use cases, MrLarge is great for your most sophisticated use cases. PixelLarge is the frontier class multimodal model. And, and we have Coastral for great for coding and then again, MrEmbedding model.[00:32:22] Sophia Yang: And The bottom, the bottom of the slides here, we have several Apache 2. 0 licensed open way models. Free for the community to use, and also if you want to fine tune it, use it for customization, production, feel free to do so. The latest, we have Pixtros 3 12b. We also have Mr. Nemo mum, Coastal Mamba and Mastro, as I mentioned, and we have three legacy models that we don't update anymore.[00:32:49] Sophia Yang: So we recommend you to move to our newer models if you are still using them. And then, just a few weeks ago, [00:33:00] we did a lot of, uh, improvements to our code interface, Lachette. How many of you have used Lachette? Oh, no. Only a few. Okay. I highly recommend Lachette. It's chat. mistral. ai. It's free to use.[00:33:16] Sophia Yang: It has all the amazing capabilities I'm going to show you right now. But before that, Lachette in French means cat. So this is actually a cat logo. If you You can tell this is the cat eyes. Yeah. So first of all, I want to show you something Maybe let's, let's take a look at image understanding.[00:33:36] Sophia Yang: So here I have a receipts and I want to ask, just going to get the prompts. Cool. So basically I have a receipt and I said I ordered I don't know. Coffee and the sausage. How much do I owe? Add a 18 percent tip. So hopefully it was able to get the cost of the coffee and the [00:34:00] sausage and ignore the other things.[00:34:03] Sophia Yang: And yeah, I don't really understand this, but I think this is coffee. It's yeah. Nine, eight. And then cost of the sausage, we have 22 here. And then it was able to add the cost, calculate the tip, and all that. Great. So, it's great at image understanding, it's great at OCR tasks. So, if you have OCR tasks, please use it.[00:34:28] Sophia Yang: It's free on the chat. It's also available through our API. And also I want to show you a Canvas example. A lot of you may have used Canvas with other tools before. But, With Lachat, it's completely free again. Here, I'm asking it to create a canvas that's used PyScript to execute Python in my browser.[00:34:51] Sophia Yang: Let's see if it works. Import this. Okay, so, yeah, so basically it's executing [00:35:00] Python here. Exactly what we wanted. And the other day, I was trying to ask Lachat to create a game for me. Let's see if we can make it work. Yeah, the Tetris game. Yep. Let's just get one row. Maybe. Oh no. Okay. All right. You get the idea. I failed my mission. Okay. Here we go. Yay! Cool. Yeah. So as you can see, Lachet can write, like, a code about a simple game pretty easily. And you can ask Lachet to explain the code. Make updates however you like. Another example. There is a bar here I want to move.[00:35:48] Sophia Yang: Okay, great, okay. And let's go back to another one. Yeah, we also have web search capabilities. Like, you can [00:36:00] ask what's the latest AI news. Image generation is pretty cool. Generate an image about researchers. Okay. In Vancouver? Yeah, it's Black Forest Labs flux Pro. Again, this is free, so Oh, cool.[00:36:19] Sophia Yang: I guess researchers here are mostly from University of British Columbia. That's smart. Yeah. So this is Laia ira. Please feel free to use it. And let me know if you have any feedback. We're always looking for improvement and we're gonna release a lot more powerful features in the coming years.[00:36:37] Sophia Yang: Thank you. Get full access to Latent Space at www.latent.space/subscribe

Happy holidays! We’ll be sharing snippets from Latent Space LIVE! through the break bringing you the best of 2024! We want to express our deepest appreciation to event sponsors AWS, Daylight Computer, Thoth.ai, StrongCompute, Notable Capital, and most of all all our LS supporters who helped fund the gorgeous venue and A/V production!For NeurIPS last year we did our standard conference podcast coverage interviewing selected papers (that we have now also done for ICLR and ICML), however we felt that we could be doing more to help AI Engineers 1) get more industry-relevant content, and 2) recap 2024 year in review from experts. As a result, we organized the first Latent Space LIVE!, our first in person miniconference, at NeurIPS 2024 in Vancouver.The single most requested domain was computer vision, and we could think of no one better to help us recap 2024 than our friends at Roboflow, who was one of our earliest guests in 2023 and had one of this year’s top episodes in 2024 again. Roboflow has since raised a $40m Series B!LinksAll the trends and papers they picked:* Isaac Robinson* Sora (see our Video Diffusion pod) - extending diffusion from images to video* SAM 2: Segment Anything in Images and Videos (see our SAM2 pod) - extending prompted masks to full video object segmentation* DETR Dominancy: DETRs show Pareto improvement over YOLOs* RT-DETR: DETRs Beat YOLOs on Real-time Object Detection* LW-DETR: A Transformer Replacement to YOLO for Real-Time Detection* D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement* Peter Robicheaux* MMVP (Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs)* * Florence 2 (Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks) * PalíGemma / PaliGemma 2* PaliGemma: A versatile 3B VLM for transfer* PaliGemma 2: A Family of Versatile VLMs for Transfer* AlMv2 (Multimodal Autoregressive Pre-training of Large Vision Encoders) * Vik Korrapati - MoondreamFull Talk on YouTubeWant more content like this? Like and subscribe to stay updated on our latest talks, interviews, and podcasts.Transcript/Timestamps[00:00:00] Intro[00:00:05] AI Charlie: welcome to Latent Space Live, our first mini conference held at NeurIPS 2024 in Vancouver. This is Charlie, your AI co host. When we were thinking of ways to add value to our academic conference coverage, we realized that there was a lack of good talks, just recapping the best of 2024, going domain by domain.[00:00:36] AI Charlie: We sent out a survey to the over 900 of you. who told us what you wanted, and then invited the best speakers in the Latent Space Network to cover each field. 200 of you joined us in person throughout the day, with over 2, 200 watching live online. Our second featured keynote is The Best of Vision 2024, with Peter Robichaud and Isaac [00:01:00] Robinson of Roboflow, with a special appearance from Vic Corrapati of Moondream.[00:01:05] AI Charlie: When we did a poll of our attendees, the highest interest domain of the year was vision. And so our first port of call was our friends at Roboflow. Joseph Nelson helped us kickstart our vision coverage in episode 7 last year, and this year came back as a guest host with Nikki Ravey of Meta to cover segment Anything 2.[00:01:25] AI Charlie: Roboflow have consistently been the leaders in open source vision models and tooling. With their SuperVision library recently eclipsing PyTorch's Vision library. And Roboflow Universe hosting hundreds of thousands of open source vision datasets and models. They have since announced a 40 million Series B led by Google Ventures.[00:01:46] AI Charlie: Woohoo.[00:01:48] Isaac's picks[00:01:48] Isaac Robinson: Hi, we're Isaac and Peter from Roboflow, and we're going to talk about the best papers of 2024 in computer vision. So, for us, we defined best as what made [00:02:00] the biggest shifts in the space. And to determine that, we looked at what are some major trends that happened and what papers most contributed to those trends.[00:02:09] Isaac Robinson: So I'm going to talk about a couple trends, Peter's going to talk about a trend, And then we're going to hand it off to Moondream. So, the trends that I'm interested in talking about are These are a major transition from models that run on per image basis to models that run using the same basic ideas on video.[00:02:28] Isaac Robinson: And then also how debtors are starting to take over the real time object detection scene from the YOLOs, which have been dominant for years.[00:02:37] Sora, OpenSora and Video Vision vs Generation[00:02:37] Isaac Robinson: So as a highlight we're going to talk about Sora, which from my perspective is the biggest paper of 2024, even though it came out in February. Is the what?[00:02:48] Isaac Robinson: Yeah. Yeah. So just it's a, SORA is just a a post. So I'm going to fill it in with details from replication efforts, including open SORA and related work, such as a stable [00:03:00] diffusion video. And then we're also going to talk about SAM2, which applies the SAM strategy to video. And then how debtors, These are the improvements in 2024 to debtors that are making them a Pareto improvement to YOLO based models.[00:03:15] Isaac Robinson: So to start this off, we're going to talk about the state of the art of video generation at the end of 2023, MagVIT MagVIT is a discrete token, video tokenizer akin to VQ, GAN, but applied to video sequences. And it actually outperforms state of the art handcrafted video compression frameworks.[00:03:38] Isaac Robinson: In terms of the bit rate versus human preference for quality and videos generated by autoregressing on these discrete tokens generate some pretty nice stuff, but up to like five seconds length and, you know, not super detailed. And then suddenly a few months later we have this, which when I saw it, it was totally mind blowing to me.[00:03:59] Isaac Robinson: 1080p, [00:04:00] a whole minute long. We've got light reflecting in puddles. That's reflective. Reminds me of those RTX demonstrations for next generation video games, such as Cyberpunk, but with better graphics. You can see some issues in the background if you look closely, but they're kind of, as with a lot of these models, the issues tend to be things that people aren't going to pay attention to unless they're looking for.[00:04:24] Isaac Robinson: In the same way that like six fingers on a hand. You're not going to notice is a giveaway unless you're looking for it. So yeah, as we said, SORA does not have a paper. So we're going to be filling it in with context from the rest of the computer vision scene attempting to replicate these efforts. So the first step, you have an LLM caption, a huge amount of videos.[00:04:48] Isaac Robinson: This, this is a trick that they introduced in Dolly 3, where they train a image captioning model to just generate very high quality captions for a huge corpus and then train a diffusion model [00:05:00] on that. Their Sora and their application efforts also show a bunch of other steps that are necessary for good video generation.[00:05:09] Isaac Robinson: Including filtering by aesthetic score and filtering by making sure the videos have enough motion. So they're not just like kind of the generators not learning to just generate static frames. So. Then we encode our video into a series of space time latents. Once again, SORA, very sparse in details.[00:05:29] Isaac Robinson: So the replication related works, OpenSORA actually uses a MAG VIT V2 itself to do this, but swapping out the discretization step with a classic VAE autoencoder framework. They show that there's a lot of benefit from getting the temporal compression, which makes a lot of sense as the Each sequential frames and videos have mostly redundant information.[00:05:53] Isaac Robinson: So by compressing against, compressing in the temporal space, you allow the latent to hold [00:06:00] a lot more semantic information while avoiding that duplicate. So, we've got our spacetime latents. Possibly via, there's some 3D VAE, presumably a MAG VATV2 and then you throw it into a diffusion transformer.[00:06:19] Isaac Robinson: So I think it's personally interesting to note that OpenSORA is using a MAG VATV2, which originally used an autoregressive transformer decoder to model the latent space, but is now using a diffusion diffusion transformer. So it's still a transformer happening. Just the question is like, is it?[00:06:37] Isaac Robinson: Parameterizing the stochastic differential equation is, or parameterizing a conditional distribution via autoregression. It's also it's also worth noting that most diffusion models today, the, the very high performance ones are switching away from the classic, like DDPM denoising diffusion probability modeling framework to rectified flows.[00:06:57] Isaac Robinson: Rectified flows have a very interesting property that as [00:07:00] they converge, they actually get closer to being able to be sampled with a single step. Which means that in practice, you can actually generate high quality samples much faster. Major problem of DDPM and related models for the past four years is just that they require many, many steps to generate high quality samples.[00:07:22] Isaac Robinson: So, and naturally, the third step is throwing lots of compute at the problem. So I didn't, I never figured out how to manage to get this video to loop, but we see very little compute, medium compute, lots of compute. This is so interesting because the the original diffusion transformer paper from Facebook actually showed that, in fact, the specific hyperparameters of the transformer didn't really matter that much.[00:07:48] Isaac Robinson: What mattered was that you were just increasing the amount of compute that the model had. So, I love how in the, once again, little blog posts, they don't even talk about [00:08:00] like the specific hyperparameters. They say, we're using a diffusion transformer, and we're just throwing more compute at it, and this is what happens.[00:08:08] Isaac Robinson: OpenSora shows similar results. The primary issue I think here is that no one else has 32x compute budget. So we end up with these we end up in the middle of the domain and most of the related work, which is still super, super cool. It's just a little disappointing considering the context. So I think this is a beautiful extension of the framework that was introduced in 22 and 23 for these very high quality per image generation and then extending that to videos.[00:08:39] Isaac Robinson: It's awesome. And it's GA as of Monday, except no one can seem to get access to it because they keep shutting down the login.[00:08:46] SAM and SAM2[00:08:46] Isaac Robinson: The next, so next paper I wanted to talk about is SAM. So we at Roboflow allow users to label data and train models on that data. Sam, for us, has saved our users 75 years of [00:09:00] labeling time.[00:09:00] Isaac Robinson: We are the, to the best of my knowledge, the largest SAM API that exists. We also, SAM also allows us to have our users train just pure bounding box regression models and use those to generate high quality masks which has the great side effect of requiring less training data to have a meaningful convergence.[00:09:20] Isaac Robinson: So most people are data limited in the real world. So anything that requires less data to get to a useful thing is that super useful. Most of our users actually run their object per frame object detectors on every frame in a video, or maybe not most, but many, many. And so Sam follows into this category of taking, Sam 2 falls into this category of taking something that really really works and applying it to a video which has the wonderful benefit of being plug and play with most of our Many of our users use cases.[00:09:53] Isaac Robinson: We're, we're still building out a sufficiently mature pipeline to take advantage of that, but it's, it's in the works. [00:10:00] So here we've got a great example. We can click on cells and then follow them. You even notice the cell goes away and comes back and we can still keep track of it which is very challenging for existing object trackers.[00:10:14] Isaac Robinson: High level overview of how SAM2 works. We there's a simple pipeline here where we can give, provide some type of prompt and it fills out the rest of the likely masks for that object throughout the rest of the video. So here we're giving a bounding box in the first frame, a set of positive negative points, or even just a simple mask.[00:10:36] Isaac Robinson: I'm going to assume people are somewhat familiar with SAM. So I'm going to just give a high level overview of how SAM works. You have an image encoder that runs on every frame. SAM two can be used on a single image, in which case the only difference between SAM two and SAM is that image encoder, which Sam used a standard VIT [00:11:00] Sam two replaced that with a hara hierarchical encoder, which gets approximately the same results, but leads to a six times faster inference, which is.[00:11:11] Isaac Robinson: Excellent, especially considering how in a trend of 23 was replacing the VAT with more efficient backbones. In the case where you're doing video segmentation, the difference is that you actually create a memory bank and you cross attend the features from the image encoder based on the memory bank.[00:11:31] Isaac Robinson: So the feature set that is created is essentially well, I'll go more into it in a couple of slides, but we take the features from the past couple frames, plus a set of object pointers and the set of prompts and use that to generate our new masks. Then we then fuse the new masks for this frame with the.[00:11:57] Isaac Robinson: Image features and add that to the memory bank. [00:12:00] It's, well, I'll say more in a minute. The just like SAM, the SAM2 actually uses a data engine to create its data set in that people are, they assembled a huge amount of reference data, used people to label some of it and train the model used the model to label more of it and asked people to refine the predictions of the model.[00:12:20] Isaac Robinson: And then ultimately the data set is just created from the engine Final output of the model on the reference data. It's very interesting. This paradigm is so interesting to me because it unifies a model in a dataset in a way that is very unique. It seems unlikely that another model could come in and have such a tight.[00:12:37] Isaac Robinson: So brief overview of how the memory bank works, the paper did not have a great visual, so I'm just, I'm going to fill in a bit more. So we take the last couple of frames from our video. And we take the last couple of frames from our video attend that, along with the set of prompts that we provided, they could come from the future, [00:13:00] they could come from anywhere in the video, as well as reference object pointers, saying, by the way, here's what we've found so far attending to the last few frames has the interesting benefit of allowing it to model complex object motion without actually[00:13:18] Isaac Robinson: By limiting the amount of frames that you attend to, you manage to keep the model running in real time. This is such an interesting topic for me because one would assume that attending to all of the frames is super essential, or having some type of summarization of all the frames is super essential for high performance.[00:13:35] Isaac Robinson: But we see in their later ablation that that actually is not the case. So here, just to make sure that there is some benchmarking happening, we just compared to some of the stuff that's came out prior, and indeed the SAM2 strategy does improve on the state of the art. This ablation deep in their dependencies was super interesting to me.[00:13:59] Isaac Robinson: [00:14:00] We see in section C, the number of memories. One would assume that increasing the count of memories would meaningfully increase performance. And we see that it has some impact, but not the type that you'd expect. And that it meaningfully decreases speed, which justifies, in my mind, just having this FIFO queue of memories.[00:14:20] Isaac Robinson: Although in the future, I'm super interested to see A more dedicated summarization of all of the last video, not just a stacking of the last frames. So that another extension of beautiful per frame work into the video domain.[00:14:42] Realtime detection: DETRs > YOLO[00:14:42] Isaac Robinson: The next trend I'm interested in talking about is this interesting at RoboFlow, we're super interested in training real time object detectors.[00:14:50] Isaac Robinson: Those are bread and butter. And so we're doing a lot to keep track of what is actually happening in that space. We are finally starting to see something change. So, [00:15:00] for years, YOLOs have been the dominant way of doing real time object detection, and we can see here that they've essentially stagnated.[00:15:08] Isaac Robinson: The performance between 10 and 11 is not meaningfully different, at least, you know, in this type of high level chart. And even from the last couple series, there's not. A major change so YOLOs have hit a plateau, debtors have not. So we can look here and see the YOLO series has this plateau. And then these RT debtor, LW debtor, and Define have meaningfully changed that plateau so that in fact, the best Define models are plus 4.[00:15:43] Isaac Robinson: 6 AP on Cocoa at the same latency. So three major steps to accomplish this. The first RT deditor, which is technically a 2023 paper preprint, but published officially in 24, so I'm going to include that. I hope that's okay. [00:16:00] That is showed that RT deditor showed that we could actually match or out speed YOLOs.[00:16:04] Isaac Robinson: And then LWdebtor showed that pre training is hugely effective on debtors and much less so on YOLOs. And then DeFine added the types of bells and whistles that we expect from these types, this, this arena. So the major improvements that RTdebtor shows was Taking the multi scale features that debtors typically pass into their encoder and decoupling them into a much more efficient transformer encoder.[00:16:30] Isaac Robinson: The transformer is of course, quadratic complexity. So decreasing the amount of stuff that you pass in at once is super helpful for increasing your runtime or increasing your throughput. So that change basically brought us up to yellow speed and then they do a hardcore analysis on. Benchmarking YOLOs, including the NMS step.[00:16:54] Isaac Robinson: Once you once you include the NMS in the latency calculation, you see that in fact, these debtors [00:17:00] are outperforming, at least this time, the the, the YOLOs that existed. Then LW debtor goes in and suggests that in fact, the frame, the huge boost here is from pre training. So, this is the define line, and this is the define line without pre training.[00:17:19] Isaac Robinson: It's within range, it's still an improvement over the YOLOs, but Really huge boost comes from the benefit of pre training. When YOLOx came out in 2021, they showed that they got much better results by having a much, much longer training time, but they found that when they did that, they actually did not benefit from pre training.[00:17:40] Isaac Robinson: So, you see in this graph from LWdebtor, in fact, YOLOs do have a real benefit from pre training, but it goes away as we increase the training time. Then, the debtors converge much faster. LWdebtor trains for only 50 epochs, RTdebtor is 60 epochs. So, one could assume that, in fact, [00:18:00] the entire extra gain from pre training is that you're not destroying your original weights.[00:18:06] Isaac Robinson: By relying on this long training cycle. And then LWdebtor also shows superior performance to our favorite data set, Roboflow 100 which means that they do better on the real world, not just on Cocoa. Then Define throws all the bells and whistles at it. Yellow models tend to have a lot of very specific complicated loss functions.[00:18:26] Isaac Robinson: This Define brings that into the debtor world and shows consistent improvement on a variety of debtor based frameworks. So bring these all together and we see that suddenly we have almost 60 AP on Cocoa while running in like 10 milliseconds. Huge, huge stuff. So we're spending a lot of time trying to build models that work better with less data and debtors are clearly becoming a promising step in that direction.[00:18:56] Isaac Robinson: The, what we're interested in seeing [00:19:00] from the debtors in this, this trend to next is. Codetter and the models that are currently sitting on the top of the leaderboard for large scale inference scale really well as you switch out the backbone. We're very interested in seeing and having people publish a paper, potentially us, on what happens if you take these real time ones and then throw a Swingy at it.[00:19:23] Isaac Robinson: Like, do we have a Pareto curve that extends from the real time domain all the way up to the super, super slow but high performance domain? We also want to see people benchmarking in RF100 more, because that type of data is what's relevant for most users. And we want to see more pre training, because pre training works now.[00:19:43] Isaac Robinson: It's super cool.[00:19:48] Peter's Picks[00:19:48] Peter Robicheaux: Alright, so, yeah, so in that theme one of the big things that we're focusing on is how do we get more out of our pre trained models. And one of the lenses to look at this is through sort of [00:20:00] this, this new requirement for like, how Fine grained visual details and your representations that are extracted from your foundation model.[00:20:08] Peter Robicheaux: So it's sort of a hook for this Oh, yeah, this is just a list of all the the papers that I'm going to mention I just want to make sure I set an actual paper so you can find it later[00:20:18] MMVP (Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs)[00:20:18] Peter Robicheaux: Yeah, so sort of the big hook here is that I make the claim that LLMs can't see if you go to if you go to Claude or ChatGPT you ask it to see this Watch and tell me what time it is, it fails, right?[00:20:34] Peter Robicheaux: And so you could say, like, maybe, maybe the Like, this is, like, a very classic test of an LLM, but you could say, Okay, maybe this, this image is, like, too zoomed out, And it just, like, it'll do better if we increase the resolution, And it has easier time finding these fine grained features, Like, where the watch hands are pointing.[00:20:53] Peter Robicheaux: Nodice. And you can say, okay, well, maybe the model just doesn't know how to tell time from knowing the position of the hands. But if you actually prompt [00:21:00] it textually, it's very easy for it to tell the time. So this to me is proof that these LLMs literally cannot see the position of the watch hands and it can't see those details.[00:21:08] Peter Robicheaux: So the question is sort of why? And for you anthropic heads out there, cloud fails too. So the, the, my first pick for best paper of 2024 Envision is this MMVP paper, which tries to investigate the Why do LLMs not have the ability to see fine grained details? And so, for instance, it comes up with a lot of images like this, where you ask it a question that seems very visually apparent to us, like, which way is the school bus facing?[00:21:32] Peter Robicheaux: And it gets it wrong, and then, of course, it makes up details to support its wrong claim. And so, the process by which it finds these images is sort of contained in its hypothesis for why it can't. See these details. So it hypothesizes that models that have been initialized with, with Clip as their vision encoder, they don't have fine grained details and the, the features extracted using Clip because Clip sort of doesn't need to find these fine grained [00:22:00] details to do its job correctly, which is just to match captions and images, right?[00:22:04] Peter Robicheaux: And sort of at a high level, even if ChatGPT wasn't initialized with Clip and wasn't trained contrastively at all. The vision encoder wasn't trained contrastively at all. Still, in order to do its job of capturing the image it could do a pretty good job without actually finding the exact position of all the objects and visual features in the image, right?[00:22:21] Peter Robicheaux: So This paper finds a set of difficult images for these types of models. And the way it does it is it looks for embeddings that are similar in clip space, but far in DynaV2 space. So DynaV2 is a foundation model that was trained self supervised purely on image data. And it kind of uses like some complex student teacher framework, but essentially, and like, it patches out like certain areas of the image or like crops with certain areas of the image and tries to make sure that those have consistent representations, which is a way for it to learn very fine grained visual features.[00:22:54] Peter Robicheaux: And so if you take things that are very close in clip space and very far in DynaV2 space, you get a set of images [00:23:00] that Basically, pairs of images that are hard for a chat GPT and other big language models to distinguish. So, if you then ask it questions about this image, well, as you can see from this chart, it's going to answer the same way for both images, right?[00:23:14] Peter Robicheaux: Because to, to, from the perspective of the vision encoder, they're the same image. And so if you ask a question like, how many eyes does this animal have? It answers the same for both. And like all these other models, including Lava do the same thing, right? And so this is the benchmark that they create, which is like finding clip, like clip line pairs, which is pairs of images that are similar in clip space and creating a data set of multiple choice questions based off of those.[00:23:39] Peter Robicheaux: And so how do these models do? Well, really bad. Lava, I think, So, so, chat2BT and Jim and I do a little bit better than random guessing, but, like, half of the performance of humans who find these problems to be very easy. Lava is, interestingly, extremely negatively correlated with this dataset. It does much, much, much, much worse [00:24:00] than random guessing, which means that this process has done a very good job of identifying hard images for, for Lava, specifically.[00:24:07] Peter Robicheaux: And that's because Lava is basically not trained for very long and is initialized from Clip, and so You would expect it to do poorly on this dataset. So, one of the proposed solutions that this paper attempts is by basically saying, Okay, well if clip features aren't enough, What if we train the visual encoder of the language model also on dyno features?[00:24:27] Peter Robicheaux: And so it, it proposes two different ways of doing this. One, additively which is basically interpolating between the two features, and then one is interleaving, which is just kind of like training one on the combination of both features. So there's this really interesting trend when you do the additive mixture of features.[00:24:45] Peter Robicheaux: So zero is all clip features and one is all DynaV2 features. So. It, as you, so I think it's helpful to look at the right most chart first, which is as you increase the number of DynaV2 features, your model does worse and worse and [00:25:00] worse on the actual language modeling task. And that's because DynaV2 features were trained completely from a self supervised manner and completely in image space.[00:25:08] Peter Robicheaux: It knows nothing about text. These features aren't really compatible with these text models. And so you can train an adapter all you want, but it seems that it's in such an alien language that it's like a very hard optimization for this. These models to solve. And so that kind of supports what's happening on the left, which is that, yeah, it gets better at answering these questions if as you include more dyna V two features up to a point, but then you, when you oversaturate, it completely loses its ability to like.[00:25:36] Peter Robicheaux: Answer language and do language tasks. So you can also see with the interleaving, like they essentially double the number of tokens that are going into these models and just train on both, and it still doesn't really solve the MMVP task. It gets Lava 1. 5 above random guessing by a little bit, but it's still not close to ChachiPT or, you know, Any like human performance, obviously.[00:25:59] Peter Robicheaux: [00:26:00] So clearly this proposed solution of just using DynaV2 features directly, isn't going to work. And basically what that means is that as a as a vision foundation model, DynaV2 is going to be insufficient for language tasks, right?[00:26:14] Florence 2 (Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks)[00:26:14] Peter Robicheaux: So my next pick for best paper of 2024 would be Florence 2, which tries to solve this problem by incorporating not only This dimension of spatial hierarchy, which is to say pixel level understanding, but also in making sure to include what they call semantic granularity, which ends up, the goal is basically to have features that are sufficient for finding objects in the image, so they're, they're, they have enough pixel information, but also can be talked about and can be reasoned about.[00:26:44] Peter Robicheaux: And that's on the semantic granularity axis. So here's an example of basically three different paradigms of labeling that they do. So they, they create a big dataset. One is text, which is just captioning. And you would expect a model that's trained [00:27:00] only on captioning to have similar performance like chat2BT and like not have spatial hierarchy, not have features that are meaningful at the pixel level.[00:27:08] Peter Robicheaux: And so they add another type, which is region text pairs, which is essentially either classifying a region or You're doing object detection or doing instance segmentation on that region or captioning that region. And then they have text phrased region annotations, which is essentially a triple. And basically, not only do you have a region that you've described, you also find it's like, It's placed in a descriptive paragraph about the image, which is basically trying to introduce even more like semantic understanding of these regions.[00:27:39] Peter Robicheaux: And so like, for instance, if you're saying a woman riding on the road, right, you have to know what a woman is and what the road is and that she's on top of it. And that's, that's basically composing a bunch of objects in this visual space, but also thinking about it semantically, right? And so the way that they do this is they take basically they just dump Features from a vision encoder [00:28:00] straight into a encoder decoder transformer.[00:28:03] Peter Robicheaux: And then they train a bunch of different tasks like object detection and so on as a language task. And I think that's one of the big things that we saw in 2024 is these, these vision language models operating in, on pixel space linguistically. So they introduced a bunch of new tokens to point to locations and[00:28:22] Peter Robicheaux: So how does it work? How does it actually do? We can see if you look at the graph on the right, which is using the, the Dino, the the Dino framework your, your pre trained Florence 2 models transfer very, very well. They get 60%, 60 percent map on Cocoa, which is like approaching state of the art and they train[00:28:42] Vik Korrapati: with, and they[00:28:43] Peter Robicheaux: train with a much more more efficiently.[00:28:47] Peter Robicheaux: So they, they converge a lot faster, which both of these things are pointing to the fact that they're actually leveraging their pre trained weights effectively. So where is it falling short? So these models, I forgot to mention, Florence is a 0. 2 [00:29:00] billion and a 0. 7 billion parameter count. So they're very, very small in terms of being a language model.[00:29:05] Peter Robicheaux: And I think that. This framework, you can see saturation. So, what this graph is showing is that if you train a Florence 2 model purely on the image level and region level annotations and not including the pixel level annotations, like this, segmentation, it actually performs better as an object detector.[00:29:25] Peter Robicheaux: And what that means is that it's not able to actually learn all the visual tasks that it's trying to learn because it doesn't have enough capacity.[00:29:32] PalíGemma / PaliGemma 2[00:29:32] Peter Robicheaux: So I'd like to see this paper explore larger model sizes, which brings us to our next big paper of 2024 or two papers. So PolyGemma came out earlier this year.[00:29:42] Peter Robicheaux: PolyGemma 2 was released, I think like a week or two ago. Oh, I forgot to mention, you can actually train You can, like, label text datasets on RoboFlow and you can train a Florence 2 model and you can actually train a PolyGemma 2 model on RoboFlow, which we got into the platform within, like, 14 hours of release, which I was really excited about.[00:29:59] Peter Robicheaux: So, anyway, so [00:30:00] PolyGemma 2, so PolyGemma is essentially doing the same thing, but instead of doing an encoder decoder, it just dumps everything into a decoder only transformer model. But it also introduced the concept of location tokens to point to objects in pixel space. PolyGemma 2, so PolyGemma uses Gemma as the language encoder, and it uses Gemma2B.[00:30:17] Peter Robicheaux: PolyGemma 2 introduces using multiple different sizes of language encoders. So, the way that they sort of get around having to do encoder decoder is they use the concept of prefix loss. Which basically means that when it's generating, tokens autoregressively, it's all those tokens in the prefix, which is like the image that it's looking at and like a description of the task that it's trying to do.[00:30:41] Peter Robicheaux: They're attending to each other fully, full attention. Which means that, you know, it can sort of. Find high level it's easier for the, the prefix to color, to color the output of the suffix and also to just find like features easily. So this is sort of [00:31:00] an example of like one of the tasks that was trained on, which is like, you describe the task in English and then you give it all these, like, You're asking for it to segment these two classes of objects, and then it finds, like, their locations using these tokens, and it finds their masks using some encoding of the masks into tokens.[00:31:24] Peter Robicheaux: And, yeah, so, one of my critiques, I guess, of PolyGemma 1, at least, is that You find that performance saturates as a pre trained model after only 300 million examples seen. So, what this graph is representing is each blue dot is a performance on some downstream task. And you can see that after seeing 300 million examples, It sort of does equally well on all of the downtrend tasks that they tried it on, which was a lot as 1 billion examples, which to me also kind of suggests a lack of capacity for this model.[00:31:58] Peter Robicheaux: PolyGemma2, [00:32:00] you can see the results on object detection. So these were transferred to to Coco. And you can see that this sort of also points to an increase in capacity being helpful to the model. You can see as. Both the resolution increases, and the parameter count of the language model increases, performance increases.[00:32:16] Peter Robicheaux: So resolution makes sense, obviously, it helps to find small images, or small objects in the image. But it also makes sense for another reason, which is that it kind of gives the model a thinking register, and it gives it more tokens to, like, process when making its predictions. But yeah, you could, you could say, oh, 43.[00:32:30] Peter Robicheaux: 6, that's not that great, like Florence 2 got 60. But this is not Training a dino or a debtor on top of this language or this image encoder. It's doing the raw language modeling task on Cocoa. So it doesn't have any of the bells and whistles. It doesn't have any of the fancy losses. It doesn't even have bipartite graph matching or anything like that.[00:32:52] Peter Robicheaux: Okay, the big result and one of the reasons that I was really excited about this paper is that they blow everything else away [00:33:00] on MMVP. I mean, 47. 3, sure, that's nowhere near human accuracy, which, again, is 94%, but for a, you know, a 2 billion language, 2 billion parameter language model to be chat2BT, that's quite the achievement.[00:33:12] Peter Robicheaux: And that sort of brings us to our final pick for paper of the year, which is AIMV2. So, AIMV2 sort of says, okay, Maybe this language model, like, maybe coming up with all these specific annotations to find features and with high fidelity and pixel space isn't actually necessary. And we can come up with an even simpler, more beautiful idea for combining you know, image tokens and pixel tokens in a way that's interfaceable for language tasks.[00:33:44] Peter Robicheaux: And this is nice because it can scale, you can come up with lots more data if you don't have to come up with all these annotations, right? So the way that it works. is it does something very, very similar to PolyGemo, where you have a vision encoder that dumps image tokens into a decoder only transformer.[00:33:59] Peter Robicheaux: But [00:34:00] the interesting thing is that it also autoregressively tries to learn the mean squared error of the image tokens. So instead of having to come up with fancy object detection or semantic, or segment, or segmentation labels, you can just try to reconstruct the image and have it learn fine grained features that way.[00:34:16] Peter Robicheaux: And it does this in kind of, I think, a beautiful way that's kind of compatible with the PolyGemma line of thinking, which is randomly sampling a prefix line of thinking Prefix length and using only this number of image tokens as the prefix. And so doing a similar thing with the causal. So the causal with prefix is the, the attention mask on the right.[00:34:35] Peter Robicheaux: So it's doing full block attention with some randomly sampled number of image tokens to then reconstruct the rest of the image and the downstream caption for that image. And so, This is the dataset that they train on. It's image or internet scale data, very high quality data created by the data filtering networks paper, essentially which is maybe The best clip data that exists.[00:34:59] Peter Robicheaux: [00:35:00] And we can see that this is finally a model that doesn't saturate. It's even at the highest parameter count, it's, it appears to be, oh, at the highest parameter account, it appears to be improving in performance with more and more samples seen. And so you can sort of think that. You know, if we just keep bumping the parameter count and increasing the example scene, which is the, the, the line of thinking for language models, then it'll keep getting better.[00:35:27] Peter Robicheaux: So how does it actually do at finding, oh, it also improves with resolution, which you would expect for a model that This is the ImageNet classification accuracy, but yeah, it does better if you increase the resolution, which means that it's actually leveraging and finding fine grained visual features.[00:35:44] Peter Robicheaux: And so how does that actually do compared to CLIP on Cocoa? Well, you can see that if you slap a transformer detection head on it, Entry now in Cocoa, it's just 60. 2, which is also within spitting distance of Soda, which means that it does a very good job of [00:36:00] finding visual features, but you could say, okay, well, wait a second.[00:36:03] Peter Robicheaux: Clip got to 59. 1, so. Like, how does this prove your claim at all? Because doesn't that mean like clip, which is known to be clip blind and do badly on MMVP, it's able to achieve a very high performance on fine, on this fine grained visual features task of object detection, well, they train on like, Tons of data.[00:36:24] Peter Robicheaux: They train on like objects, 365, Cocoa, Flickr and everything else. And so I think that this benchmark doesn't do a great job of selling how good of a pre trained model MV2 is. And we would like to see the performance on fewer data as examples and not trained to convergence on object detection. So seeing it in the real world on like a dataset, like RoboFlow 100, I think would be quite interesting.[00:36:48] Peter Robicheaux: And our, our, I guess our final, final pick for paper of 2024 would be Moondream. So introducing Vic to talk about that.[00:36:54] swyx: But overall, that was exactly what I was looking for. Like best of 2024, an amazing job. Yeah, you can, [00:37:00] if there's any other questions while Vic gets set up, like vision stuff,[00:37:07] swyx: yeah,[00:37:11] swyx: Vic, go ahead. Hi,[00:37:13] Vik Korrapati / Moondream[00:37:13] question: well, while we're getting set up, hi, over here, thanks for the really awesome talk. One of the things that's been weird and surprising is that the foundation model companies Even these MLMs, they're just like worse than RT Tether at detection still. Like, if you wanted to pay a bunch of money to auto label your detection dataset, If you gave it to OpenAI or Cloud, that would be like a big waste.[00:37:37] question: So I'm curious, just like, even Pali Gemma 2, like is worse. So, so I'm curious to hear your thoughts on like, how come, Nobody's cracked the code on like a generalist that really you know, beats a specialist model in computer vision like they have in in LLM land.[00:38:00][00:38:01] Isaac Robinson: Okay. It's a very, very interesting question. I think it depends on the specific domain. For image classification, it's basically there. In the, in AIMv2 showed, a simple attentional probe on the pre trained features gets like 90%, which is as well as anyone does. The, the, the, the bigger question, like, why isn't it transferring to object detection, especially like real time object detection.[00:38:25] Isaac Robinson: I think, in my mind, there are two answers. One is, object detection is really, really, really the architectures are super domain specific. You know, we see these, all these super, super complicated things, and it's not super easy to, to, to build something that just transfers naturally like that, whereas image classification, you know, clip pre training transfers super, super quickly.[00:38:48] Isaac Robinson: And the other thing is, until recently, the real time object detectors didn't even really benefit from pre training. Like, you see the YOLOs that are like, essentially saturated, showing very little [00:39:00] difference with pre training improvements, with using pre trained model at all. It's not surprising, necessarily, that People aren't looking at the effects of better and better pre training on real time detection.[00:39:12] Isaac Robinson: Maybe that'll change in the next year. Does that answer your question?[00:39:17] Peter Robicheaux: Can you guys hear me? Yeah, one thing I want to add is just like, or just to summarize, basically, is that like, Until 2024, you know, we haven't really seen a combination of transformer based object detectors and fancy losses, and PolyGemma suffers from the same problem, which is basically to say that these ResNet, or like the convolutional models, they have all these, like, extreme optimizations for doing object detection, but essentially, I think it's kind of been shown now that convolution models like just don't benefit from pre training and just don't like have the level of intelligence of transformer models.[00:39:56] swyx: Awesome. Hi,[00:39:59] Vik Korrapati: can [00:40:00] you hear me?[00:40:01] swyx: Cool. I hear you. See you. Are you sharing your screen?[00:40:04] Vik Korrapati: Hi. Might have forgotten to do that. Let me do[00:40:07] swyx: that. Sorry, should have done[00:40:08] Vik Korrapati: that.[00:40:17] swyx: Here's your screen. Oh, classic. You might have to quit zoom and restart. What? It's fine. We have a capture of your screen.[00:40:34] swyx: So let's get to it.[00:40:35] Vik Korrapati: Okay, easy enough.[00:40:49] Vik Korrapati: All right. Hi, everyone. My name is Vic. I've been working on Moondream for almost a year now. Like Shawn mentioned, I just went and looked and it turns out the first version I released December [00:41:00] 29, 2023. It's been a fascinating journey. So Moonbeam started off as a tiny vision language model. Since then, we've expanded scope a little bit to also try and build some tooling, client libraries, et cetera, to help people really deploy it.[00:41:13] Vik Korrapati: Unlike traditional large models that are focused at assistant type use cases, we're laser focused on building capabilities that developers can, sorry, it's yeah, we're basically focused on building capabilities that developers can use to build vision applications that can run anywhere. So, in a lot of cases for vision more so than for text, you really care about being able to run on the edge, run in real time, etc.[00:41:40] Vik Korrapati: So That's really important. We have we have different output modalities that we support. There's query where you can ask general English questions about an image and get back human like answers. There's captioning, which a lot of our users use for generating synthetic datasets to then train diffusion models and whatnot.[00:41:57] Vik Korrapati: We've done a lot of work to minimize those sessions there. [00:42:00] So that's. Use lot. We have open vocabulary object detection built in similar to a couple of more recent models like Palagem, et cetera, where rather than having to train a dedicated model, you can just say show me soccer balls in this image or show me if there are any deer in this image, it'll detect it.[00:42:14] Vik Korrapati: More recently, earlier this month, we released pointing capability where if all you're interested in is the center of an object you can just ask it to point out where that is. This is very useful when you're doing, you know, I automation type stuff. Let's see, LA we, we have two models out right now.[00:42:33] Vik Korrapati: There's a general purpose to be para model, which runs fair. Like it's, it's it's fine if you're running on server. It's good for our local Amma desktop friends and it can run on flagship, flagship mobile phones, but it never. so much for joining us today, and we'll see you in the [00:43:00] next one. Less memory even with our not yet fully optimized inference client.[00:43:06] Vik Korrapati: So the way we built our 0. 5b model was to start with the 2 billion parameter model and prune it while doing continual training to retain performance. We, our objective during the pruning was to preserve accuracy across a broad set of benchmarks. So the way we went about it was to estimate the importance of different components of the model, like attention heads, channels MLP rows and whatnot using basically a technique based on the gradient.[00:43:37] Vik Korrapati: I'm not sure how much people want to know details. We'll be writing a paper about this, but feel free to grab me if you have more questions. Then we iteratively prune a small chunk that will minimize loss and performance retrain the model to recover performance and bring it back. The 0. 5b we released is more of a proof of concept that this is possible.[00:43:54] Vik Korrapati: I think the thing that's really exciting about this is it makes it possible for for developers to build using the 2B param [00:44:00] model and just explore, build their application, and then once they're ready to deploy figure out what exactly they need out of the model and prune those capabilities into a smaller form factor that makes sense for their deployment target.[00:44:12] Vik Korrapati: So yeah, very excited about that. Let me talk to you folks a little bit about another problem I've been working on recently, which is similar to the clocks example we've been talking about. We had a customer reach out who was talking about, like, who had a bunch of gauges out in the field. This is very common in manufacturing and oil and gas, where you have a bunch of analog devices that you need to monitor.[00:44:34] Vik Korrapati: It's expensive to. And I was like, okay, let's have humans look at that and monitor stuff and make sure that the system gets shut down when the temperature goes over 80 or something. So I was like, yeah, this seems easy enough. Happy to, happy to help you distill that. Let's, let's get it going. Turns out our model couldn't do it at all.[00:44:51] Vik Korrapati: I went and looked at other open source models to see if I could just generate a bunch of data and learn from that. Did not work either. So I was like, let's look at what the folks with [00:45:00] hundreds of billions of dollars in market cap have to offer. And yeah, that doesn't work either. My hypothesis is that like the, the way these models are trained are using a large amount of image text data scraped from the internet.[00:45:15] Vik Korrapati: And that can be biased. In the case of gauges, most gauge images aren't gauges in the wild, they're product images. Detail images like these, where it's always set to zero. It's paired with an alt text that says something like GIVTO, pressure sensor, PSI, zero to 30 or something. And so the models are fairly good at picking up those details.[00:45:35] Vik Korrapati: It'll tell you that it's a pressure gauge. It'll tell you what the brand is, but it doesn't really learn to pay attention to the needle over there. And so, yeah, that's a gap we need to address. So naturally my mind goes to like, let's use synthetic data to, Solve this problem. That works, but it's problematic because it turned out we needed millions of synthetic gauge images to get to reasonable performance.[00:45:57] Vik Korrapati: And thinking about it, reading a gauge is like [00:46:00] not a one, like it's not a zero short process in our minds, right? Like if you had to tell me the reading in Celsius for this, Real world gauge. There's two dials on there. So first you have to figure out which one you have to be paying attention to, like the inner one or the outer one.[00:46:14] Vik Korrapati: You look at the tip of the needle, you look at what labels it's between, and you count how many and do some math to figure out what that probably is. So what happens if we just add that as a Chain of thought to give the model better understanding of the different sub, to allow the model to better learn the subtasks it needs to perform to accomplish this goal.[00:46:37] Vik Korrapati: So you can see in this example, this was actually generated by the latest version of our model. It's like, okay, Celsius is the inner scale. It's between 50 and 60. There's 10 ticks. So the second tick, it's a little debatable here, like there's a weird shadow situation going on, the dial is off, so I don't know what the ground truth is, but it works okay.[00:46:57] Vik Korrapati: There's points on there that are, the points [00:47:00] over there are actually grounded. I don't know if this is easy to see, but when I click on those, there's a little red dot that moves around on the image. The model actually has to predict where this points are, I was already trying to do this with bounding boxes, but then Malmo came out with pointing capabilities.[00:47:15] Vik Korrapati: And it's like pointing is a much better paradigm to to represent this. We see pretty good results. This one's actually for clock reading. I couldn't find our chart for gauge reading at the last minute. So the light. Blue chart is with our rounded chain of thought. This measures, we have, we built a clock reading benchmark about 500 images.[00:47:37] Vik Korrapati: This measures accuracy on that. You can see it's a lot more sample efficient when you're using the chain of thought to model. Another big benefit from this approach is like, you can kind of understand how the model is. it and how it's failing. So in this example, the actual correct reading is 54 Celsius, the model output [00:48:00] 56, not too bad but you can actually go and see where it messed up. Like it got a lot of these right, except instead of saying it was on the 7th tick, it actually predicted that it was the 8th tick and that's why it went with 56.[00:48:14] Vik Korrapati: So now that you know that this. Failing in this way, you can adjust how you're doing the chain of thought to maybe say like, actually count out each tick from 40, instead of just trying to say it's the eighth tick. Or you might say like, okay, I see that there's that middle thing, I'll count from there instead of all the way from 40.[00:48:31] Vik Korrapati: So helps a ton. The other thing I'm excited about is a few short prompting or test time training with this. Like if a customer has a specific gauge that like we're seeing minor errors on, they can give us a couple of examples where like, if it's miss detecting the. Needle, they can go in and correct that in the chain of thought.[00:48:49] Vik Korrapati: And hopefully that works the next time. Now, exciting approach, we only apply it to clocks and gauges. The real question is, is it going to generalize? Probably, like, there's some science [00:49:00] from text models that when you train on a broad number of tasks, it does generalize. And I'm seeing some science with our model as well.[00:49:05] Vik Korrapati: So, in addition to the image based chain of thought stuff, I also added some spelling based chain of thought to help it understand better understand OCR, I guess. I don't understand why everyone doesn't do this, by the way. Like, it's trivial benchmark question. It's Very, very easy to nail. But I also wanted to support it for stuff like license plate, partial matching, like, hey, does any license plate in this image start with WHA or whatever?[00:49:29] Vik Korrapati: So yeah, that sort of worked. All right, that, that ends my story about the gauges. If you think about what's going on over here it's interesting that like LLMs are showing enormous. Progress in reasoning, especially with the latest set of models that we've seen, but we're not really seeing, I have a feeling that VLMs are lagging behind, as we can see with these tasks that should be very simple for a human to do [00:50:00] that are very easy to find VLMs failing at.[00:50:04] Vik Korrapati: My hypothesis on why this is the case is because On the internet, there's a ton of data that talks about how to reason. There's books about how to solve problems. There's books critiquing the books about how to solve problems. But humans are just so good at perception that we never really talk about it.[00:50:20] Vik Korrapati: Like, maybe in art books where it's like, hey, to show that that mountain is further away, you need to desaturate it a bit or whatever. But the actual data on how to, like, look at images is, isn't really present. Also, the Data we have is kind of sketched. The best source of data we have is like image all text pairs on the internet and that's pretty low quality.[00:50:40] Vik Korrapati: So yeah, I, I think our solution here is really just we need to teach them how to operate on individual tasks and figure out how to scale that out. All right. Yep. So conclusion. At Moondream we're trying to build amazing PLMs that run everywhere. Very hard problem. Much work ahead, but we're making a ton of progress and I'm really excited [00:51:00] about If anyone wants to chat about more technical details about how we're doing this or interest in collaborating, please, please hit me up.[00:51:08] Isaac Robinson: Yeah,[00:51:09] swyx: like, I always, when people say, when people say multi modality, like, you know, I always think about vision as the first among equals in all the modalities. So, I really appreciate having the experts in the room. Get full access to Latent Space at www.latent.space/subscribe