Luca Soldaini: preference

Transcript context

…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. 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. 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. 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. 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. 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. 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. 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. ute 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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. 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.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.