#229 Inside Meta's Biggest and Best Open-Source AI Model Yet | Thomas Scialom, Co-Creator of Llama3

I think we are reaching a level of intelligence where this kind of agentic behavior will start to increase more and more. It's already the case. And for that you need a model that is able to format a structural output to reason, to write a plan to evaluate step by step, refine its own output.

Thomas Jamm, it's great to have you on the show.

Thanks for having me and nice to see you.

You are a senior staff research scientist on LLMs at meta's AI organization. You've been integral to the training of the LLAMA models, arguably one of the most widely used and best open source LLMs. Today. Maybe to set the stage and start off our chat, how has it been like being at the center of such an important moment in AI today?

Right. Look, it's weird, um, maybe when a quick anecdote I can tell you is I'm based in Paris where it's very AI hyped but less than in San Francisco. And when I went with my wife in San Francisco like a few months ago, she was like, we met people in the street or like, you know, oh, you're the guy from Lama. And she realized it's something.

Yeah.

Like here all my friends are, oh, you're working on AI. Cool. Do you do that? That's how I answer. Yeah.

And yeah, in San Francisco, people were stopping you on the street. So yeah, I think, yeah. I recently saw a video of Sam Altman with paparazzi behind him. It's a pretty interesting moment in AI today. So you just unveiled llama 405B which is an incredible milestone in open source AI. Maybe walk us through llama 405 being a bit more depth, which I think is the biggest open source LLM out there. What makes it unique to other open source LLMs?

Well, the size of course, but beyond that, and that just one scaling dimension is a scale we train it on. Again, like 15 trillion tokens. With respect to the preview released we did at the time, we have done much more post training. So the alignment is also much better and so the resulting performance follows. Let's see the results on the arena leaderboard, for instance, and other benchmarks. But I'm hopeful that it will be the best performing model out for open source. It has long context. It has, although you may not release it like multimodal and other capabilities. So it's really filling the gap with the frontier models.

And we're definitely going to talk about deep dive into the training of LLAMA 405B and also all of these models. But Maybe when we're speaking about training of these models, this is quite a technical challenge. When thinking about the scale of llama 405B, maybe before we talk about llama 405B, just general training these large language models, what are some of the biggest challenges in developing these models and how do you navigate these challenges?

The biggest challenge to me is how to balance exploration and exploitation. And let me elaborate. When you do some experiments, you cannot do that at 405B on 15 trend tokens and do like 20 ablations. That's just not possible. That's the kind of run even for smaller size, to be honest, like 70, that you put a lot of resources of this one. And so you need to be confident with the EP parameters, with the data, the training mix everything. And so how to choose that you need to do smaller scale experiments and to then infer the right parameters at a larger scale, you need to infer the right dynamics. Maybe like a simple example, what is the impact of multiple epoch? So seeing repeating the same data multiple times at a larger scale, there's more weights, the model memorize more. And so you could have some new phenomenon that you didn't observe at a smaller scale. And how to trade off this unoptimal large scale run because we just couldn't grid search everything. There's just the exploration you want to do at a smaller scale is something tricky. And so how even to balance the resources between exploration and exploitation is very changing.

Okay, and maybe walk us through some of the best practices that you guys have developed while balancing, while having that balancing act between exploitation exploration. What would be your advice for the small handful of other teams looking to train large language models?

I think in general the best advice is to follow first principle like okay, garbage in, garbage out, for instance for the data and data is probably one of the most important thing. So making sure that the data you collect is high quality, reviewing it manually with some processes, with quality analysis, some classifiers, we use Lama 2 to do that and select the tokens or pre training and making sure you validate your ideas and your intuition with a run at a smaller scale and you see the improvement. It's never guaranteed, but it's likely to lead to some gains at a larger scale. And I will say at every stage of the pipeline being very, very rigorous and meticulous to get the best effort to have the best class pipeline.

Okay, that's great. And you mentioned here the data and making sure that the data is high quality and doing qualitative analysis, I'd love to learn a bit more about that because massive amount of data is going into training these models. How do you make sure that this data is set up in a way that you're able to ingest it into these models and have confidence in the output? So I'd love to see kind of what are the processes that you adopt to make sure that the data is high quality because that's a big aspect of making sure that the model is successful in the long run.

For pre training in particular. I don't know if you want me later to dive more in post training and ihf, but you just collect basically web tokens. Let's say you have like famous data sets like Common Pro or the PI. But the truth is that the web is garbage. It's just random. If you look at a random paragraph or document from this data set, it's random noise. And clearly there's something wrong to train your model on 50% random noise. It's a waste of compute. So the first basic thing to do is can you train a classifier or using some heuristics or machine learning models to select what is the right token. Then can you do some topic classification? Maybe you can rebalance your data mixture such that it covers not like 0.5% some domains, but it's more like balance. You don't have a hole in your racket in the knowledge. Then can you add additional sources? What is your pipeline to clean the data like if you convert? For instance, we developed NUGA last year that performs a scientific ocr. Just putting a lot of effort in this model so that you end with a high quality text from a PDF on scientific text leads to improvement on reasoning for instance.

Okay, that's really great. And this here we're talking about the pre training aspect of it. But you mentioned as well there's a lot of work that happens on post training and and reinforcement learning from human feedback. So I'd love to learn as well about that aspect of training a model and what that looks like in practice.

Maybe the first interesting thing to say is overall the general stage for people is you do supervised fine tuning so that you start from your pre trained model which is just the next token prediction of finishing sentences and you align it to follow instructions in the chat model style truth of a supervised fine tuning. In general, it's like what humans will have written as the expected answer from a chat model. And then you do reinforcement learning with human feedback. An interesting thing to note here is we stated that in llama 2 generation from the model itself with truth, reinforcement, learning, feedback. Synthetic data are higher quality than human written outputs for a lot of these. So if you follow this principle on what should we initiate LLAMA 3 model with super high fine tuning data, we could ask humans or take what we did for llama 2 and take the supervised fine tuning data we used. But this is actually worse than what llama2 generated. So we use llama2 to generate the first pass, the first round of supervised fine tuning data. So from the round one and we had multiple rounds of RLHF, we use llama 2 to initiate with synthetic data. And once you have that, you have a llama 3 supervisedly fine tuned on llama 2 outputs, prompt and outputs. And then you start your France, you collect new preferences with this model. You train your Reward model with llama 3, you sample with your llama 3, you do reinforcement learning and so on.

That's fascinating. And essentially you reach a point where you have kind of a self sustaining loop of producing data because you're improving the model which creates supervised fine tuning data which improves the model. And at what point do you reach a limit in improvement in that phase?

So we have so many open research questions on lhs and this is one of them. We haven't seen plateau really yet, but we are also in a process where we try to nudge annotators to always ask harder and harder prompts at the edge of the model capability in a way that, well, write a poem about language model is something that now the model does pretty well. So we will ask annotators to ask less about that and more about something more complex, but not too complex because again, the process of annotation here is that you ask a prompt, you sample two answers from your model, and you ask the annotator which one he prefers. And so if the two answers are bad, you don't have any signal, basically. So you need to always be at the age maybe. One difference with Lama 2 I can emphasize is the T shirt mode we had, which is basically if the two answers are bad, we can sometimes not always enter in what we call the teacher mode, where the annotator can basically write a critic prompt saying oh, this answer, even if I prefer it to the other one, it's pretty bad for this and this reason, and then edit the answer to make it better.

And that creates another like that's an additional way to retrain the model with.

Human and to move away to have a shortcut when the model is just bad.

Okay, that's very fascinating. And so that's one of the open questions, is that that has yet to plateau. Right. And I think this is a great segue to discuss the broader generative AI landscape as well, because there's quite a few open questions on LLMs today and what the next generation of LLMs will look like. You know, from the outside looking in, you definitely have a lot more insight than I do. You know, the leap in intelligence from models seems to be getting smaller with each release. Or for example, the jump from GPT 2 to GPT 3, the GPT 3 to GPT 3.5, 3.5 to 4, you know, seems to be getting slightly smaller with each iteration. So one, I'd love to see if you think that is correct and then second, do you anticipate another significant leap in intelligence soon? If so, what does that look like and what's driving it? So yeah, I'll pause here and let.

You react if we think in terms of order of magnitudes. So first thing is, and I elaborate on that soon, yes, I anticipate much bigger jump than what we had. I think it's easy to measure the gap of intelligence between or no age, not intelligence, let me say no age between a 5 years old kid and an 18 years old kid, rather than an 18 years old kid and a 30 year old.

Yeah, that's true.

I think we are seeing kind of the same thing with those models memorizing massive amount of knowledge and at some point they know almost everything. And so yeah, maybe the gap starts to be smaller and smaller in terms of knowledge, maybe beyond the long tail, but it's harder to measure. And so the perception from people is that the gap is smaller. I will add that in terms of intelligence reasoning, it might not be true. There's like all those tasks, for instance on math benchmark we had seen like immense progress from GPT 2 to GPT 3, 4 fair. And now like Gemini clock 3.5 and almost 90. I think with one of our recent panel that indicates some progress. I will also add that at some point what matters is not the knowledge. It's fine to not know something if you know how to find it by browsing or by looking at Internet.

Right.

And so I think we are reaching a level of intelligence where this kind of ageontric behavior will start to increase more and more. It's already the case. And for that you need a model that is able to format a structural output to reason, to write a plan to evaluate step by step, refine its own output. Ultimately I think it's one of the key feature self critique itself. Right now on the LHF stage everything is synthetic, yet we have a human at some point that says this output is better than this output. And we may see to me the early sign that the model is able to judge in some contexts it's still limited. And while we will not see this kind of skills improve on some main benchmarks which are very knowledge intensive anyway, like now maybe the gap of improvement on this kind of new skills and abilities will increase, which to me let me think that the gap is just not narrowing at each generation. The second thing is when you achieve this level of intelligence, you can compound with some additional order of magnitudes improvement like tool use. We did tool former like one year ago and now with instructed model in a single shot you can give a tool and the documentation in natural language to an API and then they will learn to leverage that. And what I was saying before, learn to search online information for instance if it doesn't know a calculator, code execution and all those tools, multiple steps and self refinement and say oh, now it's a dead end. Can I maybe explore another way to solve that? And all that will compound to orders of magnitude more intelligent behaviors and capabilities that may not be reflected with the pure LLM performance on some benchmarks, but with new generation enabling that.

That's a really great insight. And yeah, I couldn't agree more. I think the agentic path as well will definitely drive a lot more usefulness. Even if on the LLM benchmark you, you may not see that come out very directly. But maybe to expand on that, how much of that agentic workflow is enabled by software engineering with having LLMs with a layer of tools access and making sure that there's a recursive nature to using the tools and all of that engineering work versus the training of the model. So how do you balance those to create a really strong agent? Like you need both. But how strong? Where are we now in intelligence essentially to enable these set of agent use cases?

So we created a benchmark last year called General AI Assistant Benchmark GAIA. It was interesting because GPT4 had like 5% GPT 3.50%. It corresponds to tasks that will require a lot of different steps. So so complex tasks for which models are failing for which the answer is not present on the Internet. An example I can give you is for instance, oh, you know, like the NASA published a photo at some point on Twitter about this astronaut and they published a similar photo but with another one in the same condition a few months later. Can you tell me the number of days between the two? And you know that this information is not spent on Internet, arguably. So there's no memorization possible. So the model has to do the steps. This is a kind of question that doesn't require any expert skills. We didn't want to go to three agents or things like that that any human will solve easily given enough time. But that models are totally failing. And with the last year what we have seen is some agentic framework like Autogen, OS copilot, hugging face agents that did what you do, what you mentioned, connecting tools and some like self refinement loop and things like that. And this was critical. It moved performance from like 10% with GPT4 to now something like 30%. It's yet to be solved, but the progress is immense. It's strenuous. That being said, if you power Those agents with GPT4, that's the performance you get. If you power them with GPT 3.5, the performance is zero. That said that actually the gap between the two is crazy in terms of capabilities.

AGN capabilities. Okay. And you can imagine that GPT 5 or you know, llama 3 403B. Right. Like will. Will drive 405B. Sorry, will drive a lot of the of the progress there.

No, that's true. And we put a lot of effort. We emphasize that. I think in the paper about the tool used for post training. I think llama3 is very good at zero shot function coding which is a key feature that so many people ask about. So I recommend to test it for that.

Definitely the community is going to play a lot with Llama 405B and the Llama 3 family. So maybe one additional thing on the state of generative AI. Do you believe that scaling laws will continue to enhance performance of models as we just increase size of data and the training runs and the compute with this famous Twitter image of small circle GPT4, massive circle GPT5. Right. And it will have 1000x performance at this point became a meme. But do you see that growing like that scaling laws continuing on as we see today? Or are we reaching a plateau?

Yes and no. I think it's a bit monumental. I think that absolutely scaling will lead to improvement. Yes. That being said, what do you want to scale? And by that there's so many improvements in terms of algorithms, in terms of training with site that leads to a better scaling slope. That 2del model at 7b are often better than what we had like 2 years ago or like than GPT 3.5 which was much bigger things like that. So I think like Mistral just released a 7B model that is outperforming Minerva, which was something based on If I'm correct, 300 billion, I forget parameters. So 50 times bigger. And so there's so many improvements that comes from algorithmic optimization as well from data quality, from the training method and from additional external things like tool use, things like that that could end in a better loop with synthetic data which are augmented than expert iteration method. But yes, scaling is a way but not scaling anything. Like if we had to scale lstm, we will be nowhere today. So yes, scaling will keep playing the biggest factor role. But we can, if we, if we scale by one to three order of magnitudes, we can get maybe 10 or 20 orders of magnitude scaling by improving all the other aspects.

That's great. So quality of data, algorithmic optimization are also big levers outside of just additional data and compute. One thing that we saw as well kind of this year, especially with GPT 4.0 was like your multimodal capabilities. Right? That definitely multimodal capabilities improve a model's usefulness. But how crucial is it to train models on these different types of data to boost raw capabilities? Maybe intelligence is not the best word, but their ability to perform tasks. Walk me through the importance of having different types of data to have more performant models.

First thing is you want your model to be able to process ideally any input to any output. And that's just much more smoother scalable than a very complex orchestration. That being said, with respect to how it will improve a multimodal transfer to let's say text is a bit unclear. Not that it doesn't exist. My intuition is that either we didn't crack the problem yet, or another hypothesis is that text is so dense in information signal from what it contains, it's a log of the thoughts somehow that you need so many more images to basically map the signal of information you will have in the text. What else? Maybe in multilinguality there's an argument that a speech end to end model may with the accents in donations like that leads to improvement. There's a fake potency there which we haven't explored yet on our side.

And then maybe one last question on the kind of the state of generative AI and the technical landscape that we're living in that we are in today. You mentioned something, if you scaled LSTMs, we wouldn't be where we are today. Gary Marcus famously Repeats this notion that deep learning is hitting a Wall and that LLMs and generative AI will plateau as time goes by. Granted he's been saying this for the past at least five years and we've been seeing improvements all throughout. Do you think deep learning and the transformer paradigms are reaching their limits?

No. I mean anyone that said that in the past has been proven wrong in a very short term. No, what I can tell you is look, what I often say is one we know that we have, I was saying before exploration with respect to exploitation, I think if we had like zero more progress, zero more breakthrough in the next decade we will have still like five to 10 years to just explore, understand all these breakthroughs and revolution we just had in which will lead to incremental significant progress, but incremental only for the next 10 years. Integrating all the modalities together, managing everything, agentic behaviors, well done. Inference, quantization, better post training, where we are still at the prehistorical age there all those things together will lead to actually probably few dozen better models. And yet there's so many resources put in there. Not just money, but people that are working on that and more and more every day. It's likely that what the last decade of deep learning told us that a breakthrough will come again. What it will be, I don't know. We had seen like a post training method, we had seen transformers, we had seen self supervision and all those things progressively. It's likely that something else will happen and will again give us like 4 dozen magnitude more better deep learning stuff. So I think it will be very crazy to bet against that.

I personally agree with you. We should have Gary Marcus on the show, see his thoughts maybe as we shift a bit to discuss kind of the broader LLM landscape as well and other providers, the open source community, the private foundation models. With the announcement of GPT 4.0 by OpenAI OpenAI Claude 3.5 by Anthropic First I'd love to see from your perspective, how do you view these models, right, like you know, from a technical perspective what is impressive about these models and then what are your thoughts on their capabilities? And then I'd love to see kind of how you view the ecosystems of open source AI versus you know, closed source or you know, private foundation models, you know, emerging or like evolving over the next year.

I think all those models are pretty good. I guess we're at a stage where we are more or less at the same level of capabilities. Besides maybe GPT 4.0 being like end to end Multimodal, which is something kind of breakthrough compared to the others. That being said, progress will keep going, at least incrementally and then for the next generation of compute and scale, I'm really keen to see what will happen. I would say we have filled the gap between open source and closed source. Think about one year ago we were releasing llama 2. I think it was a few days about one year ago. Before that you had zero open source chat model. We had done a long way, we have come from a long way. Now there's more and more open source models that are better and better. Hopefully with this one we'll fill the gap even closer. Now the question is how to have open source being number one and not just number two, which we are working hard on that the trend is pretty good. I guess the question which is an open question to me is in a world where there's more and more companies still leading the space, but with some secret source like Gemini Anthropic OpenAI, will it mean for them that they cannot compete with the open source community? Because every time there's a new thing that they will work on, the open source community will fill the gap and will be able to get better and better models and like closed source is not the way for open science and will just limit progress or they will be able to make some breakthrough internally that will be very hard to approach externally, which leading to some very important advance for some of them and not others. There's an open question there. We'll do our best so that open source leads the game.

Okay. And this really segues really well to my next question because I think meta is the exception here in a lot of ways and a few other organizations such as Mistral, maybe. But do you believe private models have an inherent advantage over open source counterparts because of their compute the amount of money they have, their ability to hire researchers, the data that they're procuring. If so, how can you mitigate this as a community, as an ecosystem, rather than as meta?

Yeah, that's a question I often ask myself. I don't know to what extent. I mean, for instance, we don't train our model on user data. We don't have anything special about that besides working hard on high quality data. So we don't have a competitive advantage to be at meta right now for that kind of models. Do they have at Google or at OpenAI? Maybe, I don't know. Is it something? I'm not sure entirely. It may be helpful, but to what extent is not totally clear to me yet. So I'll say the main thing is OpenAI started that game and betting on scaling language model a long time ago. We had our first language models one year and a half ago. I was working on Galactica December it was one year and a half. Then my friends in Paris followed with Lama one pretraine model. And we are very, very early in the game on our side. So we move very fast I think to fill the gap. And the train is good. I think it's the main reason.

And the open source communities is super vibrant. Right. LLAMA has an incredible open source community and there's a lot of players today in the open source community. Right. Which LLM providers do you think will survive this wave? Or maybe what will be the defining feature of an LLM provider that will survive this? I don't want to name names here. And why do some models gain popularity while others don't? Right. This is such an interesting aspect of the ecosystem.

Yeah, that's interesting. I mean I think first of all the performance of the models is something that works. There's of course a bit of hype and marketing around that. Although most of the time I tend to think that the credibility of the organization that reads the models, you can release everything with putting all the numbers and overfitting benchmarks. If people know that and discover that once, they will have much more doubts the next release you will do. For instance, we have proven in the past to be very solid in our evaluation with lemma. So people trust us and others have done the same and that's good. So I would say then what I'm seeing right now, but that might change. Things are changing very fast, is a lot of small models that are either specialized or good generalized. And you want to take this one for code, this one for science, this one format, this one for general stuff and then for the bigger size there's less and less competitors though for the 5 like at GPT4 performance we are the only one. There's a question of what is your usage. There's probably a lot of platform actually that could take advantage of that and serve all the models together. I think that's good, that's good for the community, that's good for transparency. It's good to not have a monopoly there. So overall I'm in favor of having more and more open source companies raising data models.

Couldn't agree more. And then because something that you hinted that I'd love to also unpack a bit more, there's going to be kind of a generation of really small models for really specific use cases like Code Stroll is a good example of that. For, for example. Right. But there are also massive models that you can use across a variety of use cases. How do you see that aspect of the LLM ecosystem evolving? Maybe when should you use a massive model versus a small specialized model? I'd love to gain your thoughts on how you see the ecosystem evolving there.

I think it's always a question of price to pay and time to inference for your product needed and capability in the setup you need for your client. If you need something very, very good and you cannot accept any hallucination or like and the 405e model is giving you enough performance so that your product is useful for your client. But a smaller model is not reaching the threshold. Well, you don't have choice then. It's a balance with how much you are willing to Pay. A further 5B model will cost more than a smaller model. It will take also a bit longer to infer versus being better. So it's always case by case, project question rather than anything else. Now there's smart strategies that I'm seeing more and more leveraging high good models at the beginning to create synthetic data that you maybe distillate in smaller models and all that. So there are also all this stack that people are building.

Interesting. So a lot of it is depending on inference time and costs for serving these models and the use case, which definitely agree with you. And I think this segues as well to my next question, which is Europe's in a research environment that's actually shipping models that are being used in products today, which I think is quite a, you know, not necessarily common for all types of research environments. That it's a lot of the product, a lot of the research that you produce is being used in production almost immediately as soon as it's shipped. So what's it like working in a research environment that's, you know, also delivering, you know, real world value as soon as the research is done Right. And how do you succeed in this type of environment?

Yeah, it's very new and actually it's reflected with even our organization at Meta. We had, as you may know, split between. We were all working at fair. Exactly. It's a foundation AI model research and we split it in a small group at the beginning to a new org called Genai which is also responsible not just to create foundational models, llama 2, llama 3, but also products, meta, AI and all those things that are now shipped in our products. I think it's the right move. And it's very interesting because we are at a level of capability for those models where it's directly useful for some products and so driving the research with the project feedback. What is useful or not? What do you need where the model is bad? Just that evaluation of those models which are general by nature is extremely hard. Having some people actually using your model for direct products, having billions of users telling you your model is bad at that and that and then telling you yeah, I saw the progress now it's good at that. This is extremely useful.

You mentioned something like serving a model to billions of users. That's I think a challenge that only a handful of teams on the planet have to deal with. Yeah. How do you deal with such level of scale when serving like what are some of the best practices tricks you've learned along the way to be able to serve so many users with inference?

So I think one thing I can answer you technically is having super efficient small models which is why we kind of from the scaling laws went way beyond what is optimal compute at training time to actually like beyond the scaling the chinchilla trap. I can write if you want about that. But basically we train it on so many tokens that it makes small models very efficient. That's the first thing. Maybe a word about that just scaling. You have two dimensions basically the size of the model, the weight or the training tokens, the number of steps and thing is at inference time one of the two dimensions drops and not the other. Your model is big, it will remain big at inference time. But training number of tokens of steps is not a factor that impacts inference time. You can train on 1 billion token or 100 billion tokens, you will have the same inference time. And so we train them on a lot of tokens. That's one way first to tackle efficient models for inference serving a lot of users. Now the second thing I want to add is that's not my experts scale there. There's huge teams at meta and that's in our DNA to be able to scale. There's super good teams here. I saw they're able to sell that to start to work on that. And while we haven't saved yet on billions of user, we are starting to see the on meta AI. I don't know if you tested it, but it's pretty good. We've already integrated generation of images, flash images, freshmen like that. That's pretty cool.

Okay, that is definitely pretty cool. And then maybe as we cap off our chat, Thomas, what current research in the Field excites you the most at.

The moment Agentic behavior. I think pre training has still a lot to get but it's getting more major. There's yet a lot of different things to get post training we are now starting to understand in the beginning there's a lot of research to improve it incrementally. But agents is to me what will lead to the next many orders of magnitude more capabilities and it's the far west for now. So when I did tool former that's when I said to you before I realized we need language models aligned to follow instruction to move to the next stage. And my view is now that we have GPT4 level capabilities with llama 3. That's level of capability where you start to see the preliminary self refinement capabilities. There's this nice paper Eureka from Jim Fanhood Nvidia for instance. And from now on, while we will keep increasing and improving on post training and pre training, this is where you can start to do agentic behaviors.

Okay, that is excellent. And then maybe as we wrap up doma, what are some major trends that you expect this year to kind of happen in the LLM space and the generative AI space? And maybe if we can put a. When do you think that next breakthrough will happen? Like what's your time range for the next breakthrough if you were to put money on it?

Right. So about trends, I would say end to end multimodal integration that starting to. We had The Moshi the first model release. Actually GPT 4.0 is not really released in mixed model. So startup QT did the first release there and I expect to see more and more we will go there as well. Of course that's one, two agents. Of course we are seeing that more and more. Three. Compute optimal inference, a lot of strategy to improve compute and actually the fact that we release for the first time such a big good model. I expect the community to play with it and to find a lot of crazy ways to make it more inferent. Funny for I would say robotics it's something very very at the age at the beginning. But I expect to see more and more progress soon which is the same. It's very correlated to agenting behaviors. But I'm also seeing the cost of robots decrease which will make these models grounded in real physical work. Basically yeah. To answer your question about when is the next breakthrough, I don't come from the future. I don't know. I would say something big will happen in the next five years. I'll be surprised.

Okay, so that's the only prediction I was asking for. Next five years, something big will happen. We'll see what it will be. And I don't know. Yeah, we don't know what it is. And yeah. Thank you so much, Tama. I think that's great. To cap off today's episode, thank you so much, Tamau, for coming on Data Frame. I really, really appreciated the chat and thank you so much for coming and sharing all the great things the Llama team is working on.

Thank you for having me.