Ep 12: EleutherAI's Aran Komatsuzaki on Open-Source Models' Future and Thought Cloning

Jacob and Jordan sit down with EleutherAI's Aran Komatsuzaki to discuss the future of open-source models, thought cloning, his work on GPT-J and more.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  I'm Jordan Siegel and I'm joined by my co-host Jacob Effron. On today's episode, we have Aron Komatsuzaki, lead scientist for Eluther AI, where he was the chief architect behind the largest open source model in the world at the time it was released, GPT-J. He's also a member of Lion, one of the leading nonprofit AI research institutes, which contributed to Bloom, Stable Diffusion, Imagen, and Openfold. We talk about the latest in AI research.

• 0:23Speaker 01

  The area I'm most excited with, at least within the LLM field, is the idea of self-cloning.

• 0:28Speaker 00

  He discusses the incredible gap between top AI companies and everyone else.

• 0:32Speaker 01

  Even if, say, Elon Musk just spends $10 billion on model trading, but he still can't compete with OpenAI. These researchers and engineers, they do not really care about money. They simply want to work with other top researchers or engineers. So they want to go to, let's just say, Google OpenAI.

• 0:47Speaker 00

  And of course, we asked Aron about his predictions for AGI.

• 0:50Speaker 01

  Probably it's gonna take very little amount of time until we achieve actual AGI.

• 0:55Speaker 00

  Hope you enjoy the episode. Aron, we're big fans of your Twitter, so we're very excited to have you onboard today. Thank you so much for joining us.

• 1:02Speaker 01

  Thank you for having me.

• 1:04Speaker 00

  To start, you know, it'd be great, and I'm sure our audience would love to hear a little bit more about yourself and what got you into AI in the first place.

• 1:11Speaker 01

  Yes, I was, just like many other machine learning researchers, I was very fascinated with the idea of AGI. So when I was a child, I wanted to become a mathematician, physicist, a music composer, and biologist, and many different jobs. And the idea is that if I can build a machine learning model that can behave as a researcher or like artist of many different domains, then I don't have to be a mathematician myself. So I thought that was like the most straightforward path for me.

• 1:49Speaker 00

  Awesome. And then, you know, what attracted you to Eluther?

• 1:53Speaker 01

  When Eluther AI was founded, there weren't really any active machine learning Discord starters. So most of the discussions were happening in Twitter or Reddit. And

• 2:07Speaker 01

  The latency was not really good compared with chat. So I personally prefer chat format over Twitter. And I was just invited my friend to join this AI Discord server. And I was personally very interested in GPT three. So I just joined their project. And I didn't join their GPT Neo or prior project personally. But I was, I started becoming more serious.

• 2:43Speaker 01

  When DALL-E was released by OpenAI. And my project surrounding DALL-E led to Lion four hundred bit medium project and GPTJ.

• 2:55Jacob Effron?

  Yeah, how did you, I mean, it's such an interesting set of products you've worked on there. Like, how did you even originally decide to like work on GPTJ in the first place?

• 3:03Speaker 01

  So actually, GPTJ originally was just a DALL-E deprecation project.

• 3:08Jacob Effron?

  Really?

• 3:08Speaker 01

  The thing, yeah, the thing is that DALL-E, for producing DALL-E, we need a few things. Firstly, you need a huge image text dataset. And actually, that became Lion dataset later. So unfortunately, we could not replicate it in AI community. But our, my project sort of motivated the same similar project at Lion, this called Cyber. So their lead, Kristoff, was originally in AI Cyber. And he was following my project and we sort of paired.

• 3:50Speaker 01

  So we brought this project to Lion Cyber and continued. And that was for the Lion project and for GPTJ. And so.

• 4:01Speaker 01

  And so for reproducing DALL-E, there are two components. One is encoder-decoder model, and another one is transformer language model for generating the players. And this component is pretty much the same as GPT-3. So except for the difference in data. So we decided that we can build this library that can do both DALL-E player and GPT-3. Then we can do one stone with two bird. So then there are a few improvements with this library.

• 4:42Speaker 01

  Firstly, it was written in JAX, unlike TensorFlow, which was used for GPT-Neo. And inference was much faster with JAX than TensorFlow. So that was a major selling point for our library. And we also made sure that we also solved some of the issues with GPT-Neo, which was probably the training instability. And so we shuffled the dataset much better than GPT-Neo. And we added a few improvements, which led to the fact that GPT-J was the first model

• 5:22Speaker 01

  to replicate a smaller version of GPT-3 with a performance comparable to GPT-3's small 6 billion version. And GPT, because GPT-Neo, the performance was not really comparable to GPT-3's 3 billion or 2 billion models. So I think that was the measure.

• 5:46Speaker 00

  Super cool. I mean, we'd love to hear a little bit more on sort of how you went about gathering the data for Lyon and GPT-J. Obviously, you know, today the dataset is sort of the crown jewel and what we're seeing as differentiator for a lot of these models.

• 5:59Speaker 01

  Yeah, for GPTJ, we simply used the same dataset as GPT Neo, which was the the pile dataset. It was built during the first six month or four month of LFO AI and and organization and basically we had about a dozen of contributors for each subcomponent of the dataset. And the pile was to simply replicate the dataset used for training GPT three. With additional

• 6:35Speaker 01

  and components such as stack exchange and and many different smaller sub-databases to improve the diversity, which was slightly lacking with GPT three as training, original training dataset. And I was personally not involved with the development. It was led by Leo Gao, who was one of the co-founders of LFO AI and now he's in OpenAI. So, um, yeah. And and as for Lion dataset, and we just tried to simply replicate the dataset used for Dali training.

• 7:16Speaker 01

  And so it has about four hundred million data, image, text, pairs. As for how we built the dataset, it just like the dataset which was built for Dali training. We simply collected and links from Common Cloud. So Common Cloud is a set of

• 7:39Speaker 01

  set of webpages, clouds across layers, and it was used for the pile as well. And we just looks for the links for images, JPG or PNG. And we simply tug those links and and and accompanying text. And yeah, that was not, and that was there is, this process is kind of expensive if you just naively collect some of the all the images, but some of us came up with

• 8:14Speaker 01

  Some tricks which made it slightly more affordable and we could do it with a reasonable amount of budget.

• 8:21Jacob Effron?

  Yeah, super interesting. And you were doing all this like while you were in school?

• 8:24Speaker 01

  Oh, actually, I think back then I was taking a break from school because it was during COVID. So, yeah, I didn't typically I don't spend much time for my school project. And most of my projects are coming from externally like AI.

• 8:44Jacob Effron?

  Yeah, no, that makes that makes a ton of sense. Super interesting. And I guess, you know, it might be helpful for our listeners just to kind of contextualize. Obviously, Luther is one of the main, you know, open source efforts out there. Just kind of the history of Luther and some of, you know, and the open source AI community in general. You know, I think would be, you know, folks would be super interested in that.

• 9:02Speaker 01

  Yeah, AI is an open source resource, this called community of machine learning. And it was founded about three years ago, so 2020 summer. After GPT was released, the original objective of the this community was to reproduce and open source GPT-3. And even though it was not achieved historically, in the historic sense, we could

• 9:32Speaker 01

  open source and reproduce some of the smaller version of GPT-3 and other models. And for the first six months, it was spent mainly for developing the pile data set and GPT new model. And the next six months, some of the major milestones includes GPTJ. And, yeah, and Lion data set was not exactly the contribution from AI side. But some of us worked with Lion to build this data set, which, yeah. And after that.

• 10:09Speaker 01

  Um. Some of the uh people builds uh this model called GPT NeoX, which is a twenty billion GPT three like model. And uh they later developed this uh suit of models uh like uh. Uh it's called it's a Pether and it's basically a GPT three with many different sizes like uh from twelve billion to one hundred million uh parameter models. And uh this is uh quite useful for uh doing research with interpretability and uh many uh

• 10:49Speaker 01

  different kind of uh experiments with GPT like models without spending ridiculous amount of compute. And after that, I think most most of the focus of the uh LFO AI was uh was to uh interpretability and projects. Yes.

• 11:08Speaker 00

  Luther was sort of founded on the principle of open source models and you played a huge role in GPTJ and others you mentioned. Um. There's this, you know, sort of battle between open source models and closed source models right now, where the proprietary models are sort of in the lead from a performance standpoint. I mean, do you think open source models will catch up at some point? And, you know, how will that happen if if you think it will?

• 11:29Speaker 01

  Yeah, I think it's really difficult for open source models to catch up with closed source models. Uh primarily because I think this uh there's this general trend of the winner uh becomes more like um successful. And so um I think uh this this huge gap between open top players such as OpenAI, Anthropic and uh Google and uh other lesser players, including open source communities as well. So uh I not um uh for example, in the case of Pether AI or Lion, we've never like uh at the same level as uh OpenAI in terms of uh technical uh.

• 12:16Speaker 01

  Uh expertise uh for editing and so I I'd say um it is so far there is no evidence that we're gonna catch up with OpenAI anytime soon for Google. And and and not to mention that uh if you wanna uh build a model like G P T four uh you uh need at least uh for nearly one hundred billion dollars. Uh which is extremely difficult to um spend for open source models. Uh because open source models are much harder to uh uh capitalize compared with cloud source models.

• 12:57Speaker 01

  So yeah, um e- even though my I'm very passionate with open source uh uh open sourcing uh large language models uh. I'd say this um I'm I'm not super um I'd say the chance is very small.

• 13:13Jacob Effron?

  Yeah, it's interesting. I mean, it seems like there's there's, you know, certainly some folks that think you'll need massive amounts of compute to stay state of the art uh and in large language models. And then, you know, other folks that are like, well, look, these models get so much cheaper to train over time and, you know. Is is there actually that much returns to throwing more compute at these models? I think even, you know, uh it's still unclear how much like actual, you know, increased parameter size of G P T four was versus just like a better mixture of experts model. Um and so it I guess, do you kind of fall in the camp of, hey, you know, the cutting edge models are always gonna require compute that's kind of out of the out of the like spend range of open source?

• 13:47Speaker 01

  Oh, yeah. So I should also mention that that you need a lot of uh budget for other components. So uh compute is obviously uh one of the major factors. Another major factor, well, two of them. I'd say uh the uh researchers, uh how how many uh top researchers and H E S you had? And how much you can spend on uh the data set part? Uh more specifically, I'd say supervised fine tuning and R level H F component. So uh those uh this data set is uh is actually expensive and.

• 14:26Speaker 01

  Maybe not as expensive as the amount of money you spend on compute. But it is expensive, especially for supervisors to fine-tuning. And yeah, most importantly, the number of top researchers and engineers you have. So there's a clear gap between top players, they call it AI, or Google versus other smaller players. So for these top researchers and engineers can come up with like a good design for the pipeline, like training, data sets, like how to design good data and how to

• 15:06Speaker 01

  and nowadays, they don't even release all their research findings by AI. So and those kind of knowledges. So and so even even if, say, Elon Musk and he just spends like a ten billion dollar spend on model training because he can't have all these top researchers and engineers. So he still cannot compete with OpenAI. And these researchers and engineers, they do not really care about money. They probably simply want to work with other top researchers or engineers. So they want to go to places like Google or OpenAI.

• 15:48Speaker 01

  So yeah, that kind of, I think that's a major barrier, I'd say.

• 15:57Speaker 00

  So it sounds like if you're an investor, you would not bet against OpenAI.

• 16:00Speaker 01

  I I think so.

• 16:02Speaker 00

  And so if if OpenAI is sort of, you know, that far ahead of the curve, you know, maybe Anthropic and and Google as well. Do you think there's room for companies that are creating LLMs for very specific use cases? Whether it's verticalized use cases or cogeneration or things like that? Or do you think companies will just end up taking OpenAI or another, you know, LLM and and fine-tuning it?

• 16:25Speaker 01

  Yeah, I think that's a very good question. And if you look at the historical trend of machine learning, everything is that there's a definite trend of every model becoming more and more generous than specialist. So for example, before GPT two, many of the models were specifically trained for one or a few.

• 16:49Speaker 01

  But after G P T two and the major trend is uh multi task learning and generist uh trend. Even though uh uh until recently, this uh general generist models performance is pretty um. For I I compared with uh single uh specialized models. But nowadays, G P T four's performance is so much better than after so much better than special specialist models. For example, if you compare uh MedPalm, which is a um uh specialized on medicine medical uh domain. Uh if I recall correctly, their performance is comparable to G P T four, even though G P T four was not specialized on medicine.

• 17:33Speaker 01

  So the reason why this happens is is that first the uh G P G P T four's reasoning capability is uh pretty good. Even though it's uh it's data, high during data was probably not specifically on uh the domain of medicine. Um if you, as long as you have a good reasoning skills and a low enough pre-training data set, which uh necessary contains uh medical knowledge. Then you can um uh you can still perform well on the area of medicine. And because uh uh the training uh the pipeline is agnostic to any particular domain.

• 18:13Speaker 01

  So you can expect the same uh strong performance on many different domains, including code generation and uh. Yes, so um I'd say so. Another thing is that if you uh make your model uh generist, you can invest more on your model. And you can spend more uh time and uh budget on on your models and expect higher returns. Uh from this model because it can be used on many different tasks, so the customer base is probably bigger. So um.

• 18:49Speaker 01

  This all these factors

• 18:53Speaker 01

  so the put it to a stronger generous aiming for generous models rather than specialist models. Yeah, so that's my personal opinion.

• 19:03Jacob Effron?

  Yeah, no, it makes sense. I mean, I guess I'm curious then, obviously, you know, you're seeing people use, you know, your own open source models today. As you think about like then where folks will use open source models going forward versus like the state of the art closed source models. You know, both today and then in the future, like where do you think these these open source models fit in?

• 19:22Speaker 01

  Yeah, I think so.

• 19:25Speaker 01

  I'm also if this is like the my my response is going to be relevant to what you are trying to ask. But I think the the presence of open source model is gonna is going to force the companies with closed source models to be more like a I think they they want to reduce their price to be competitive, even though. So so if even though I say closed source models performs much better than open source models. If you apply these models on easier tasks, then the performance gap is not so big. Because if the task is easy, even the weaker models can perform well. So, for example,

• 20:12Speaker 01

  chat chat GPT is much cheaper, ten times cheaper than GPT four, if I recall correctly. And so and I think there are many tasks where you can simply use chat GPT and you don't need a GPT four. And so and I'd say that the existence of open source model adds another layer to it. And yeah, the best open source model sets the determines the like the price of the weaker, low tier version of the closed source models.

• 20:50Speaker 01

  So I think it may be the contributing factor for why OpenAI reduced chat GPT to to their current price. So, yeah, I think I like this.

• 21:03Speaker 01

  This kind of pressure.

• 21:05Jacob Effron?

  Yeah, so as long as open source keeps evolving, it puts the pricing pressure on the on the closed source players. And I guess over time, probably whatever that lowest tier of capabilities is expands and expands, right, as just the models get stronger and stronger.

• 21:17Speaker 01

  Exactly.

• 21:18Speaker 00

  Switching gears a little bit, I mean, as I mentioned at the beginning, we're big fans of your Twitter and you do an awesome job of sharing, you know, the latest research in the space. And we'd love to hear a little bit about some of the most interesting areas that you're excited about.

• 21:30Speaker 01

  Yeah, and the area I'm most excited with, at least within L L M field, is the idea of thought cloning. So there was this paper titled Thought Cloning released in the beginning of this month. And I think it was released from Jeff Cleer's group. And the idea is to basically, so the paper advocates for clones that stop steps that happens inside human brain. And their experiment was very simple, not like a, it's a very toy toy problem. So they they had like a very simple video game

• 22:16Speaker 01

  and they verbalized their future action into natural language. And so it's essentially verbalizing this thought step that happens, say, humans, before acting according to this thought. So there's not much to say about this paper, but I personally have this extended version of this idea. So, for example, when you write an article, you do not simply start an article like how G P T four does. G P T four simply starts from top to bottom in the page immediately.

• 22:59Speaker 01

  But that's not how humans do, humans write an article. We typically start with some brainstorming and some ideas to think about or write about. Then we continually refine and reflect on these ideas and elaborate any of them. And.

• 23:23Speaker 01

  Then we we may um possibly add or remove some of these ideas. Then um once we decide uh all these ideas, then we probably start writing. Maybe we uh we have to brainstorm how to um translate these ideas into the concrete uh format. Uh which uh how how we should start the open uh opening paragraph or ending paragraph. Or like uh should we change uh change the position like order presentation. All these uh things happen, which is why we spend um maybe hours before uh finishing the um article.

• 24:04Speaker 01

  Uh unlike G P T, which um most of uh G P T's way of uh writing an article is very immediate. And doesn't seem to happen uh doesn't seem to have much thought uh into it. So I think so the way how humans spend uh time uh for uh doing this kind of work. We spend most of the time um uh in the internal uh discussion, internal moment. I think this kind of uh so if we want to uh encode this kind of process into large language models. Then um I do have some ideas, for example, when we write an article.

• 24:46Speaker 01

  Uh let's say uh we uh let's say an annotator writes an article. And when she uh writes an article, she has to uh verbalize every uh thought steps that happens in the brain. And write it down um um one by one. And we it may not be quite accurate, but I think uh we have to do some experiments like that. And just uh writing code this uh thought step that happens. And uh all the outputs that human writes on the editor. And uh if we fine tune a lot uh large language models on this kind of data.

• 25:26Speaker 01

  Hopefully, uh the model um learn some of the internal processes that happens in the brain.

• 25:34Speaker 01

  Yeah, that's what I'm thinking about. And I think other people have similar ideas. And I think this kind of process may happen in the future language models.

• 25:47Speaker 00

  Awesome, super interesting. And we also understand that you're working on something in the video space, a video model of sorts. I mean, I think our audience would love to hear a little bit more about that.

• 25:56Speaker 01

  Yeah, thank you for mentioning that. So we are collecting

• 26:03Speaker 01

  we are collecting computer activity data set with many different modalities. So we are letting annotators to use computers and laptops and do many kinds of tasks, for example, playing video games, watching YouTube videos, or doing many different kinds of

• 26:25Speaker 01

  so like I'm browsing internet, finding information and using apps. And we are recording those processes using video screenshot, audio, mouse and keyboard input, and even eye movement tracking. So

• 26:45Speaker 01

  so eye movement is actually often overlooked in machine learning today. I think it's super important because the way humans

• 27:01Speaker 01

  the visual inputs for human eye is quite different from the visual input for camera or video camera. So in the case of video camera, each input frame is like 1,000 times 2,000 pixels, which is very expensive. But on the other hand, human eyes only process

• 27:24Speaker 01

  like a very small region around the center of the eye. And it's like probably like 100 times 100 pixels. And every region around that is processed in a very blurry manner. So I'd say you can down some of those regions into very small square. And that means you can process very small number of pixels per time frame. And I think this kind of work mission is needed for future video models.

• 27:59Speaker 01

  Otherwise, you have to spend a lot of compute on each frame, which is kind of a physical.

• 28:05Speaker 00

  Really interesting. You know, sort of thinking back to our earlier conversation on generalized models versus specific models and OpenAI, you know, they're moving toward a multimodal world. They already deal with image. You know, do you think there's room for a generalized perception model that would have the sort of same level of effect in the video space as LLMs did on text?

• 28:25Speaker 01

  Yeah, I think so. By the way, this is just a rumor, but I heard that OpenAI's next target, what next project is involved with like extending GPT's capability to video side. So, yeah, I think there are going to be many players working on video. And obviously, for the reason I stated, it's going to be quite useful for web tasks on web. And also, it is very much needed for playing video games better. And so, right now, there are some models that perform well on Minecraft,

• 29:07Speaker 01

  but obviously, it cannot complete the entire game from the beginning to the end. It's just for like some processes, like building, like, yeah. So I'm guessing that in a few years, there's going to be an agent that can complete Legend of Zelda or Dark Souls or any very long video games from beginning to the end.

• 29:32Jacob Effron?

  And do you think that like, I mean, obviously, some folks think, you know, the kind of combined power you get from having models that understand, you know, text, image, video together, you know, will kind of reinforce the strength of the underlying model versus having these standalone kind of like LLMs, you know, text, image, text to video. Like, do you think there ends up, you know, being kind of bespoke models in each kind of modality? Or you end up with one kind of overarching model that kind of understands all different sorts of inputs and outputs?

• 30:00Speaker 01

  I think by the general trend of generous to a generous model, I think there's going to be just this one model that can do everything. But one thing. Um, I I recently noticed is that um when you combine image generation model with like the um image representation model and that large language models, there's not much synergy uh happening. So, yeah, for some reason, so it doesn't necessarily improve the image generation capability. So maybe some of the uh tasks can be separated. Uh but uh in in general, I think uh it makes a lot of lot of sense to just combine everything together.

• 30:41Speaker 01

  And uh I see that uh there are some evidences that combining uh language model models with uh video models and improve the video model's capacity quite a lot. So I'd say it's gonna be something like uh what happened with G P T four. So G P T four has both text-only model and image-text model. So I think that kind of uh resource is actually, that kind of way of releasing models is gonna be popular. So um uh the the reason why this is is is good is because once you train, I saw training large language model is gonna be very, very expensive.

• 31:24Speaker 01

  But once you have this model, you can find this model combined with uh say image or video component. And that part is actually probably not as expensive. So, yeah, um then this model may not be, may not perform as well as uh purely text models on purely text tasks, text tasks. And uh having text-only model is quite flexible for many different kind of um customization. So I I think this kind of uh training is gonna continue.

• 31:55Speaker 00

  You know, for for years, sort of industry, the relationship between industry and academia. Um industry has sort of led in certain areas of A I research, such as computer vision and others. You know, and now, obviously, on L on the L L M side. You know, obviously, you just finished graduate school and would love to just hear your thoughts on sort of the role of academia going forward into the future, particularly in the L L M space.

• 32:16Speaker 01

  Yeah, so the influence of academia um in terms of large language models and state-of-the-art machine learning models. I think it's gonna further diminish. Um so it it was already in decline, and I think this uh this training is gonna continue further. But the one thing.

• 32:37Speaker 01

  But still, academia is like the major source of like young researchers in the form of PhD students. And many engineers are coming from PhD as well. And for example, in the case of Google, even though some of the engineers who just joined them and for as a machine learning engineers after getting a bachelor's or a master's degree, some of them decide to be become a machine learning researcher and they just go to Stanford or Berkeley or other PhD program.

• 33:20Speaker 01

  So even though there are some successful researchers who just

• 33:28Speaker 01

  directly, there are some engineers who directly become a machine learning researcher inside Google or Meta. But that's not exactly easy to do. So I believe that many of them just go to PhD and then later go back to the original company and become a researcher. So I think this pipeline is not going to change anytime soon. And but one thing I I don't like about is that the quality of faculty members and what their research is not particularly high outside of a few top labs, like if you back to your Stanford.

• 34:12Speaker 01

  So there aren't many, there isn't much opportunity in general. But there are some opportunities out there.

• 34:19Jacob Effron?

  Well, we always like to end our our interviews with kind of a quickfire round where we give you some some questions and get your get your quick takes on them. And so maybe to kick it off, would just be curious, you know, for one thing that you think is overhyped and one thing you think is underhyped in AI today?

• 34:35Speaker 01

  Yeah, there are many things that are overhyped in general. I'd say.

• 34:41Speaker 01

  Yeah, sometimes, well, speaking of open source models, some of the models are definitely overhyped. Like, for example, even though some of the models perform pretty well according to human evaluation on like

• 35:00Speaker 01

  like summarization or writing an article, but.

• 35:05Speaker 01

  I bet you you guys have seen this some of some of the reports on Twitter that these models, open source models are not very good at knowledge intensive tasks like um

• 35:19Speaker 01

  some much with choice academic questions. So yeah, open source models tend to be overhyped. And also, there was this overhype on like um, what was it? I think this like an agent type model based on GPT four that was like a overhyped a few months ago. And we don't hear about that anymore. I don't remember the name. So yeah, there's a lot of overhype recently in the in machine learning because all the influences from crypto and other areas are coming coming to

• 36:00Speaker 01

  machine learning. So unfortunately, that's happening and many of the conventional ML influences and tips and knowledgeable people are really complaining about that.

• 36:15Jacob Effron?

  Yeah, what about underhyped?

• 36:17Speaker 01

  Okay, um, this this may be a bit different, but I think the contributions from individual researchers tend to be often overlooked. And is this is this related or?

• 36:32Jacob Effron?

  No, that's great. Anyone in particular you want to shine a light on that you think has been underhyped?

• 36:37Speaker 01

  Yeah, yeah. So um, I often hear that, for example, in the case of stable diffusion, there's a

• 36:47Speaker 01

  Robin Lombard, this researcher named Robin Lombard, his contribution was often

• 36:53Speaker 01

  overlooked and many people attribute the success of this model to say like a CEO or like whoever was was the top of the company. So I see this generally, and maybe not limited to machine learning, but a lot of credit is assigned to CEO or whoever was the most famous among among the others, among the all the others in the in the paper. So I think that's a very unfortunate and

• 37:25Speaker 01

  Yeah, if you just forget about it and if you could properly attribute praises to the person who actually worked on the work, then you can probably, even as an investor, you can just, you know, look at this person and maybe if he leaves the company and funds a startup, then you can be more, you can trust these people more.

• 37:51Jacob Effron?

  That that happens in my work too. You know, Jordan always overshadows me and all the the stuff we do together. It's tough. But but I guess, you know, another one on the quickfire side, what are your thoughts on AGI? You know, is it possible and what timeframe?

• 38:06Speaker 01

  Yeah, I'm I'm guessing that it's gonna happen. So, okay, before speaking about AGI, I think what we should aim for is a machine learning model that can do machine learning research better than humans do. As long as we can build a model like that, then we don't have to do anything and we just have to wait for this model to just keep working on machine learning research. And probably it's gonna take very little amount of time until we achieve actual AGI. And so we want machine learning models to do research, so without any human

• 38:48Speaker 01

  assistance. For that, there are many, a few things we we have to solve. For example,

• 38:57Speaker 01

  that, for example, that the context of lengths of the machine, the current language models is very limited, unlike humans. We humans typically remember what happened even years back. And that's not the case for machine language models. So humans typically speak, hear, well, read like a hundred thousand of words. So if you extend it to an ear, then it's gonna be like millions of time steps. So we have to vaguely remember what happened millions of time steps ago, which is actually quite a bit difficult for language

• 39:39Speaker 01

  models to do, even though there's some approaches that can do something like that. So.

• 39:45Speaker 01

  Also, the point I uh I raised uh with the uh thought cloning. So this uh human-like thought step is not happening in the current language models. I think that's kind of thing should be incorporated. And um the uh video model um may be useful sometimes, maybe not for research, but for many of the processes we need uh vision input. You know, even though there are some blind people who cannot process any uh vision input and still do perform very well. So maybe that's not completely needed, but it's gonna be quite useful.

• 40:25Jacob Effron?

  Totally. And then obviously, you know, you're a source of news for so many people in the space. Besides our podcast, which I'm sure you listen to religiously, like what do you use for sources of news? And how do you kind of keep up the tabs on what's going on in space?

• 40:39Speaker 01

  Yeah, um my primary source is uh archive. So I think um most of my information is coming from them. And the thing is that most people don't carefully read these papers. And they just they just skim uh Twitter and um so often people uh get confused with like uh any research findings. Like uh um so, yeah, that's eighty percent of it. And um other than that, we I just talk with many different peoples. And um next thing is that when I uh reach out to any people, uh any machine learning researcher,

• 41:22Speaker 01

  they typically respond back to me because they probably know me already. Uh so that's a next thing. Yeah, and um also uh Discord is a nice way to uh immediately uh communicate with uh people across the world. And uh um yeah, that's about it, I think.

• 41:45Jacob Effron?

  Yeah. What's the coolest conversation you've been able to have, uh you know, given your Twitter presence?

• 41:51Speaker 01

  Oh, in um anywhere like Twitter or Discord.

• 41:54Jacob Effron?

  Oh, yeah, or through anything.

• 41:56Speaker 01

  I sometimes hear like uh leaks uh from companies like Google and OpenAI from these collections. And even though they are not completely uh reliable, so.

• 42:10Speaker 01

  Yeah, also, you know, even though I'm pretty active on Twitter, I often miss a lot of the news. But when I talk with people around me, then they just remind me that this happened. It's a, um, I think it's a nice way to just keep updated.

• 42:31Jacob Effron?

  Totally. Well, this has been a fascinating conversation. You know, I'm sure many of our listeners already do follow you on Twitter, but I guess, you know, would love, one, for those who don't, you know, and also just, you know, generally, if you want to point folks to, you know, any of the open source work you do, the floor is yours. Where can, where can folks go to learn more about you and the work you do?

• 42:48Speaker 01

  Oh, um, I'd say you can just go to my Twitter page. And I sometimes, oh, I even sometimes talk about my own project as well. Like, I recently started multi lingual project, which happens in my Discord server. So, yeah, um, um, well, please check my Twitter, I guess.

• 43:11Jacob Effron?

  Amazing. Well, thanks so much for coming on. This is a super interesting conversation.

• 43:14Speaker 01

  Thank you so much for having me.