Machine Studying（機械の勉強）：Jacob X. Li 氏と考える、コーパスだけから事前に専門性を獲得するAIエージェント

Jacob, welcome to the WeVA podcast. Thank you so much for joining. Thank you, Tyler. Thanks for reminding me to this wonderful podcast. Yeah, I've been following this for a super long time. Awesome. Well, yeah, I've been aware of machine studying for a long time, and I'm just so excited to be diving into it. Do you mind kicking it off for our audience who aren't already familiar with machine studying and your paper? What is machine studying? Sounds good. Sometimes it can be logs of your machines when you're trying to debunk your system. So basically a lot of this new field, I would say emerging field,

or sometimes field that is conditioned on your confidential data, is pretty much just a few documents, right? And if we're looking at these documents, there's a new problem, a new preparation problem showing up, where basically we wonder whether these agents can also look at these documents. and then decide what to learn from these documents when we don't have a downstream task in mind yet. For example, if I give it my coding repository, I will have a few different tasks to have it done. Maybe just rewrite my whole complete code base or maybe run a set of experiments for my next research paper. There's a few possibilities about what downstream questions can be given to these agents. Right. And we wonder whether they can actually prepare themselves to answer these questions better. And this is pretty much similar to how we study. If you think about how we study in school and we study a subject and then ultimately will be quizzed with an exam.

Right. We don't we don't get to see what the exam is, but we know it's pretty much conditioned on the materials we are taught in the class. So basically, machine studying asks whether given nothing but this corpus. right this uh there's all the definition about this field new field is like in the corpus right can these agents develop expertise in these underlying domain but i think just for now i will uh not use the word expertise but because i we can probably elaborate a little bit more but now i'll pretty much say that um can these agents develop arbitrarily deep understanding in these field. And if you think about it, like current agents, let's just say codecs or clock code, they're mostly relying on inference time compute, which means that if we give them our code base, they will be launching lots of queries after we give them our first questions, and it's going to explore the code base a little bit. So it's pretty much relying on these inference time compute.

And in a realistic you know, agentic settings. No matter what sort of studying or preparation we give these models doesn't really replace access to these source materials, to this corpus. So here we pretty much have everything you need to know about machine setting where we are given these agents a corpus of documents and they don't know about the downstream task yet, but they are pretty sure that these tasks are going to be conditioned on the corpus. And then the model, when they're actually handling the exam, they will still have access to the corpus. So this is pretty much the setup of machine studying. Yeah, amazing. I don't think there are a few topics in AI more exciting than the kind of always-on agent, 24-7 agent, especially when it's like multi-agent. So that's kind of the first thought I had with reading machine studying and this idea. It reminds me a lot of, Leda had this paper called Sleep Time Compute, which is generally this idea of you have the corpus of documents and preparing for questions,

just, I guess, expanding on the data. It's just such a captivating idea to think about our agents just working with our data, I guess, just to mine new knowledge. And so I guess kind of the first question, just sort of setting the stage, and particularly this language of machine studying, is I'm curious. My first thought after reading your paper is maybe just the difference between like studying compared to a lot of these works on discovery. Like we recently had a podcast with UC Berkeley Sky Discover team. I think it was kind of an interesting notion that the always on agent, the agent that works while you sleep, maybe going through your data to try to like discover something compared to maybe studying for an upcoming questions that you're going to ask it. I guess just maybe this particular use of studying, I think, is a really interesting way to think about what the 24-7 agent is doing while you're not actively using it. Yeah, that's a great question. I think for discovery, definitely there's this emphasis on revealing something new.

And there's this same spirit as in studying because, you know, a lot of time when we read textbook or we look at the code base, sometimes we have new ideas just conditioned on the code base, which is not necessarily written down in the code base. If you think of it, all these corpas, if you don't mind me using some weird languages, they kind of define the physics of the field, right? Because if you have a code base, you pretty much have everything you need to understand how to use this code base, like everything, how to manipulate this object, how to use it for practical tasks, right? But a lot of things that related to understanding might not be written down. For example, if there's a code base where a senior user might be like, okay, it is more reliable to call this API instead of that API.

Because this API may have a faster reaction time. A lot of these understanding or experiences may not be included in the documents, but are completely reachable if it's given to a human. For example, if I actually give you a code base and I have you to play around with it, you will probably find this pretty efficiently. So I think for machine studying, it's also, I would say, similar to this discovery from within this corpus. You know, we're not necessarily saying that, okay, by just studying this corpus, we expect completely new knowledge to be discovered. but we do expect some sort of like internal connections or things to be revealed by, by carefully prepared for this field. Yeah. So interesting. And I think that maybe transitions to start talking about some of the formal

definitions and experiments in the paper. I love the, like the definition of expertise. I think it's just a really cool experiment and really interesting way to think about it. Before I read too much into it, Do you mind explaining how you're defining expertise in machine setting? For sure, yeah. I think one, I'll just repeat this setup a little bit again. Basically, when you're considering any of these realistic, agentic setups nowadays, we have them interacting with the environment. We most of the time have the exact same documents that we want to give them to study and also later when we are quizzing them, when we're giving them this exam, we want them to search around, right? So for this definition of expertise, I think an expert is an agent that can effectively and efficiently turn this inference time search, inference time compute, into accurate work.

And this sense of turning or the efficiency of turning inference time compute into accuracy is measured by this weighted area under the performance versus compute curve. Because most of the time when people think about interest on compute, probably they will be thinking about the BrowseCamp paper where they have this log linear increase on performance when they're given more time, more and more budgets to search around, right? So this is, I think, the same curve, but we care about the area under the curve. So that basically if you think about it, if you have someone who doesn't understand the code base, and if you've given them something that's reasonably hard but not unsolvable, just by giving them a little bit more time, maybe not a little bit, but like a lot of time,

They can probably brute force exhaustively search through everything, every corner, and then answer them correctly. But an expert might have some heuristics, some better understanding of where to search, and how to better interpret the search results and just give you the same results much faster. so yeah i think like most of the uh uh the reason we choose machine studying and expertise is pretty much because like we have very good definition of these terms on us over like human learning theory right because like when i say like how do you uh uh define like expert over a human expert over human novice or human beginners uh we we might not like not maybe not everyone has a very formal definition of this, but we have a pretty good sense about, or a very good intuition about, like, what differentiates expert and beginners. I think we can pretty much borrow that intuition into this setup as well.

Yeah, super cool. And I'm super excited to dive back into the human intuition and what we think it means to study, especially with, I just love how the paper explores like jointly context augmentation as well as like weight optimization and like that kind of chi-chi versus the lore updates. And yeah, I'm so for that topic, but I guess just saying a little more on this definition of expertise. So I understand the definition is about, you know, how much test time compute do you need, and particularly in this sort of React agent harness where, you know, the agent has these turns of thinking tool calls, thinking tool calls. And so you have first answering with no tool calls, then you have up to five tool call loops and up to 20 loops, and then you force 20 loops, so you get four points on the graph. and then you have the accuracy. And I think that's just a really novel way to think about just measuring the intelligence of the sort of agent system, which is a combination of the model, the particular data that's in the database,

as well as how it's using the tools. And I guess I'm just curious a little further into that metric of like the area under the expertise curve and just how you came to it and just I guess what your current thinking is on it. Okay, so I think basically the X-axis you were describing where we allow models to have maybe search up to 20 times or we force them to search up to 20 times is just a knob on controlling the inference and compute. because we want to understand how these models can perform on their different inference budget. And usually in, I would say, a lot of the pre-agentic paper, like pre-agentic setup, most of the people are controlling the reasoning effort. You know, should we reason for... There isn't really a number.

Sometimes it's just X high, max, or should we do medium reasoning, right? So this is basically how they're controlling a reasoning effort. We basically port that knob and decide to control the inference time budget by how many times can you search, right? Can you search, like, the most basic form when it has to answer directly pretty much says, like, you have to remember everything. Although we kind of allow the model to take for a bit, so it's still allowed to reason. But in our about to be released version we have made a few changes to how we control this inference time budget but I think the overall intuition under this is that we want to check how well these models are performing on their different budget If we give them a little more time, can they search more effectively and actually just turn this huge inference time budget into better performance?

right so yeah the four points on the graphs are just like four different inference time budget just like low reasoning, medium reasoning high and x-line reasoning in the about to be released version we actually find a way to make more fine grained knobs of the inference time compute so you have like more than 10 points on the xx yeah I guess That first point about how you construct the points on the inference time axis, it's very interesting that we've got this kind of reasoning parameter of just medium, high, X, high, and it's just kind of like how many tokens are sampled before it answers. And then the agent harness gives you this nice number of loops abstraction for thinking about how much inference time your computer is spending. I'm also very interested in these DSPy-style harnesses you could design where you could have these really intricate inference patterns that kind of govern how much computers spend. But I'll be spending a little more time on that, on the React loop, the agent function calling loop.

And just sort of what you just said has inspired me thinking about another work from MIT called Peak, about, like, creating, like, a context map. I like this paper. It's called, like, Compass Rag or something about, like, using topic modeling to come up with the topics that are in the corpus. And I guess generally not without previewing too much about the cheat sheet approach when we're talking about the actual studying methods you test, but, like, it sounds a lot to me like kind of like learning to search and, like, how the agent is going through the corpus to search more effectively. I think how to search more effectively is definitely part of what you're studying. And in the blog post, we have this third example where we kind of touch on how well the language model can be used to write literature review for a new paper, let's say. So if you think of this task, it basically can be reduced into two steps. One is that can the model launch correct queries conditioned on some new contents?

Basically, if I give a title and an abstract of a new paper that the model has never seen, can the model launch correct queries to find related papers? And the second part of it is that out of all these papers that it retrieved, maybe like hundreds of them, can it decide which one is actually the useful one, the correct one that will be actually used by a real human writer, real human researchers, right? So just like if you think of studying or like doing research as a problem of search, it's pretty much like search is only part of research. We kind of have to search iteratively and we have to decide which one actually, like when we condition on the search results, which part of the search results is actually going to contribute to the actually downstream research. So one finding from this mini experiment was that even if you use a very outdated model, the model and given one of the worst retriever, BM25, compared to a better model, like a newer model, these two models will pretty much cover the same set of paper.

So basically, if I give it a new, like a title and some new content, the model will launch a lot of searches. And if you go through all the searches trajectories, both model will cover pretty much like a set of paper that's believed to be useful. Like both of them will see a lot of useful papers. And there isn't a clear advantage of a new model here. However, when the model are actually writing the literature review, conditioned on these search results, these retrieved papers from the past, these new models have a huge advantage over which paper is important to this potential new field because probably this new model has been trained on some of the later papers that's not seen by the older model. So you see that this old model and the new model might not have that much difference in its searching capability.

But based on its conditional search and retrieved results, there is indeed a difference in the capability of it using it more efficiently to answer, like, I'd say, like research type of questions. Right. For example, right. Let's show you. So, yeah. Yeah, there's so many interesting ideas in that. I guess, like, it firstly reminds me of that kind of, like, browse comp plus evaluation where you're using, you have the agent searching in a loop, and then you're using that to compare, say, you know, GPT-6 Luna with GPT-6 Sol with the idea that how many, you know, it takes Sol less search turns to find the correct document. And then you're saying, like, therefore it knows how to search more effectively. And I guess just that topic of like, how do you search more effectively? How do agents learn to search more effectively? What does it really mean to like write good search queries out? So find that topic to just be really interesting. And I guess the other one talking about like using BM25 as a search tool compared to say,

maybe vector search or late interaction models. I know we're maybe getting a little off topic, but just if you don't mind humoring me a little more, I love this kind of agentic search and just, you know, taking down this topic. What do you think about the kind of role of, you know, VM25 versus vector search in this particular context? I think it really depends. So basically there's this, so when you're talking about agentic searches, I always kind of think about this other work from our last, the Oblique Bench, right? So when it's basically like, can you, let's say if I'm giving you a math question that's solved by a specific theorem or a specific procedure, can I find a similar question that's solved in the same way? I feel like these type of questions will definitely not be so easy for a BM25 keyword search type of log retriever.

and maybe people would probably think about training better retriever just to have a better sense of I would say similarities, how do you compare the similarity of two questions definitely training a better retriever will help you to increase the performance but I think one premise of studying is that, because first of all, we never stop any setting algorithm from training a new retriever, right? If the model feels like, okay, in this new field, maybe what I need is just a new retriever, and the model autonomously derived that decision, then sure, just have it designed the new retriever by itself, right? The second thing is that even the model is only given a BM25, what we believe is that there should still be a way to effectively learn.

Because after all, I think human doesn't have something better than BM25, right? If I give you a textbook, all we have is probably read it page by page. And if you really want to find something, you probably go to the glossary and find, like, you do the keyword search yourself. we probably have something worse than PM25. But that doesn't stop us from effectively studying, right? So that's why I think in my blog, I made this claim that retrieval is, well, like basically studying is not just retrieval, right? Retrieval is definitely, like how to retrieve more efficiently is definitely a very important topic. but just like I said, even with the worst retriever, the model should still be able to study. And apparently, even humans are only given BM25, we can still navigate the textbook very effectively, right?

Yeah, it's a super interesting topic. And yeah, I guess just on that one, I think the strongest argument I've heard for BM25 is the retriever in agentic search is that it gives the agent really great feedback because if you search for a keyword and then it's like no matches found it's just like really great for that kind of iterative on the fly learning and then also when you incorporate this machine studying you get this sort of you know context map if you don't mind me using or cheat sheet where it gives you like kind of like some terms and it gives you that preliminary context that helps you then use BM25. And I've also heard that same argument for late interaction because it gives you these per token scores. And I think all that's really interesting. I guess on the single vector search side, I think there's maybe something to diversity ranking methods like maximal marginal relevance and how that can maybe help agents expose space. But yeah, generally, I do think the feedback it gets is a really strong argument for BM25 and agentic search. Yeah. So I think coming out, I think we

kind of expanded the scope of machine study, like you started to talk about the kind of methods, and I really want to dive into the methods because you mentioned it might be able to train a retriever, which maybe even before we come into that context, just do you mind explaining the sort of how do machines study? What are some of the different methods that we know to do this? I think before talking about how machines study, I kind of also want to talk a little bit about, like my understanding of machine study versus control learning, right? Because I would say these two topics are pretty much put together to discuss by a lot of people. And I think one thing that control learning really did great is to lay this foundation of, or like it kind of promised us that like agent can keep learning after they're deployed, right? but one thing that it didn't really talk about is like it really says where or what the model should learn from

or like what for example like when we are given a new core episode to the to the agent what should the model do like how should it learn how should it continue to learn this is not really clearly defined in um in the in the control learning literature and and to be honest i think this term is a little bit overloaded And in practice, even control learning can be, you know, can end up getting reduced unfairly to just, like, long context kind of processing. Like, can we just have, like, model to get incredibly long context? And then, you know, there's no, like, control learning just becomes putting more things in the context, right? So, but, like, there's one, you know, other difference I kind of want to highlight here is that control learning is usually framed around the sequence of tasks or benchmarks, right? So, like, for example, if I give it 10 benchmarks or 10 reinforcement learning environments, can the model learn the 10th one without forgetting the first one, right?

And a more ambitious version is that can a model learn the first one, the second one? And when it's learning the third one, can it take advantage from what it has learned from the first and second one? But I feel like for a mission setting, it's more like can we learn a new field, like the first environment when we don't have a clear benchmark or clear reinforcement learning environment yet? Can the model itself decide what actually should be learned or extracted from the corpus itself? where does the supervision come from at all, right? So then when it comes to machine studying, we kind of, this is also part of my ongoing research, but I feel like every existing approaches will make one of the three bets, right? So it's either doing some sort of self-supervised, like it following some sort of self objectives on the corpus For example conducting next token prediction loss or it it self some training data or some environment And when it updating itself it could update its own weights through let say LoRa

Or it can update a prefix tuning cache, right? I think there's a word called cartridges, which is like doing, like it's training the KB caches. Or you can update or train a retriever, right? For example, can you synthesize some questions to train your own retriever? That's also definitely a viable solution, right? And the third batch is like just amortized context management, like notes and teach-tees or preparing a context or a dynamic context for itself when answering the question. So I think even for future studying method, we should still look at the three batch and maybe a better setting algorithm can emerge if we just be smart enough and, you know, play around with all these three different directions. Yeah, I love that framing of machine setting versus continual. I guess it's like it's always kind of presented like this idea that you have sequential tasks presented is like what is definitely going to happen or the idea that the distribution.

I think my favorite understanding of continual learning is kind of the distribution of data could shift and that could cause the kind of catastrophic forgetting. And I think continual learning has come up a lot lately with, like, agent memory products, and it comes to mean, like, you know, if you come to me and say, Jacob's a student at MIT, but say you, I don't know, transfer to Berkeley, and then it's like, where is Jacob's study? And that's kind of, like, somehow sort of confounded with continual learning now. And, yeah, I think this machine-studying framework, yeah, I think it's a – like, I agree with the sort of framing of it. Yeah. definitely yeah sorry yeah so I think that would transition nicely then to talking about yeah just these methods for machine studying and yeah just the space of it yeah for for the memories and I would say a lot of the recent approaches on on

training models to let's say better taken into consideration of its interaction history with human or interaction history of, let's say, part of the HIDA exam is definitely very important. And I would say a lot of the time, even if we think of just having the model to be trained on some human-prepared data, And one thing I kind of want to highlight is that we should be aware of some sort of sample complexity there, no matter if it's RL or if it's any sorts of SFT. The thing is, one degenerate form of continual learning or machine studying even that we see every day is from these frontier labs.

They actually, they can just, let's say, retrain their whole model from scratch. just take all the new data and maybe some of the synthetic traces that they want the model to be specifically good on into their free training data set and train on these models. And then they just replay their mid-training and post-training pipelines. And then they probably get a model that's potentially good on a field that you want. Or at least they kind of solve the... Learning from literature kind of questions, like maybe if we just replay a lot of times the part of the literature we care about, it's free training data. The model can be really good on that certificate field. So that's also another question in machine study is that how effective can you study? And we did not elaborate in very detail because this will involve comparing different sorts of training method under different training budget, which is probably too extensive for me to conduct.

But a good thing is that in our framework, we keep that comparison possible by pushing one layer above, like by actually collecting the expertise with your different training budgets. Just like imagine how we are collecting performance versus inference time budget. Now we are collecting expertise versus like training time or studying time budget. we are able to compute this area under this curve, which we call setting intelligence. And this intelligence allows you to compare different sort of setting method. And here, a baseline is literally just training your model from scratch, or maybe training your model from scratch plus some of the synthetic data you want to include. And I think this also gives you a view of, like, how to see machine studying and future studying algorithm, or like how to compare future studying algorithm,

because like some algorithm will probably do better than, you know, pre-training from scratch. That's what we hope to have. And some method, if you actually see maybe, let's say if we're training on completely synthetic questions, then that sample complexity might even be lower than, or like the samples needed to train the model to be good or to achieve certain level of expertise might even be worse or be even more inefficient than training from scratch. Yeah, these studying algorithms, we recently had researchers from Humans and on the WeVA podcast who talked about the training of persimmon and how they started from mid-training to get the model to sound, to mimic human behavior in the data, and just that kind of thinking about like when do you need to use mid-training compared to post-training? Like how far do you kind of like rewind back up to the pre-training stack?

I find that topic to be super interesting. But yeah, I guess just maybe like unpacking these. I love maybe just like the high-level topic of when you can get away just with sort of a cheat sheet, like prompt basically like taking notes and having an augmented prompts compared to when you actually update the weights of the neural network and this kind of idea that you have this interplay between the two. And, yeah, just I guess maybe jumping ahead to the results that I think you're finding that you found in the first paper that's out now is that the cheat sheet seems to be just way more effective than the sort of LoRa updates and training on synthetic question-answer pairs. So I guess just maybe unpacking these, like what are the studying algorithms? yeah so cheat tree being I think one result that we can get from the block version is that cheat tree actually helps for DSPype but not so much for open claw because maybe open claw is just a newer code base or plus it's much larger

and a cheat tree might not be good enough plus the model itself might not be smart enough to actually understand what it needs or explore the harder side of the questions, like harder side of the repository. But one thing that we kind of don't want people to take is that we don't want, like, we didn't really want to show that maybe just relying on context instead of relying on weights is the right path to future studying algorithm. Soon we'll have some more experiments showing that, okay, if we do some sort of reinforcement learning over some synthetic data set, you can also get pretty good gain on expertise. It's just that at the time of the block, cheat sheet is still one of the best ways

because after all, the model is allowed to interact with the code base more, like, with more agency, I would say. And you would actually expect a smarter model to be able to do this smarter than, like, in a smarter way than a lot of the smaller QIN models, right? But you would also expect that these models will be, just think of this hypothetical case where we're giving this model a new domain, right, and the defining purpose. the model itself can synthesize a lot of question-answer pairs. And then there's no reward, but the model itself can verify the answer. So basically, the model synthesizes the question by itself, grade itself, and train itself. This is pretty much a degenerate form of reinforcement learning, and I think most of the time in the past literature, you will expect it to fail. But if we actually think about maybe having the best model to do this, it might not lead to the optimal expert, but it might actually take this model to somewhere that's better than the original self.

So I think we would like to leave the comparison between weight-based and context-based setting algorithm open. And we believe that those are going to be very important for future learning algorithms, like future studying algorithms. Yeah, I guess I'm also even curious just how like the gap of prompt optimization might interact with the cheat sheet. I've had this talk like hundreds of times in the podcast, it feels like where it's like how to think about weights versus prompt optimization. and I guess this kind of like cheat sheet memory adds sort of like a third layer to it. I think kind of, or I'm not sure it does, but I feel like it isn't quite the same way of thinking about prompt optimization as like a lot of the kind of DSP predecessors of like say OPRO or MIPRO or GEPA and just it's like it still feels like a slightly different way of thinking about it. I guess the difference between the sort of cheat sheet approach compared to if you were to like optimize a prompt,

like with a formal prompt optimization algorithm well I don't see them contradicting to each other maybe like if we condition on the actual downstream you know if we have some sorts of we can just have the model itself to decide for example the model itself write some questions right that treat them as their own held out because they have no access to the hidden exam. So they themselves could just write an exam. And then they have the condition on that exam score, right, that exam score. You can definitely optimize your prompt for writing the cheat sheet, right? And, of course, you can run all sorts of, like, prompt optimization from past literature, and then you can definitely have a better cheat sheet condition on the self-generated exam.

But again, I think this also goes back to the topics on whether the model itself can write these questions in the first place. Can it actually understand what are the questions that's worth practice on in the first place? This also goes back to this comparison to controlled learning because if you are talking about controlled learning or learning a specific task where there is a benchmark that exists. It's fairly easy to see, okay, which method is better than the other ones because the model are directly optimizing, directly help climbing the specific benchmarks, right? Let's say if we want the model just to be better on browse comp, then we are directly comparing, should we train a better retriever? Should we fine tune the model over all the browse comp literature? Should we do something else, right? Should we write a cheat sheet? should we prompt optimize it but here again the model are only given the corpus right so should the model itself write some questions and then do prompt optimization or directly train on these

questions or like what is um like how should the model even learn from this corpus is pretty much open and and i think this is why uh i think machine setting has been really exciting to us is because there are so many unknowns, and a lot of the previous method can directly be applied here if you actually figure out how or if the model feel like they the right way yeah it such a fascinating idea i think like we had like the needle in the haystack question generation philosophy where it like the prompt is write the question that this document uniquely contains the answer to and then it sounds like you're opening up the box of like how would we come up with other sort of synthetic tasks just given the corpus i don't even think there's a really great way to um synthesize deep research style questions they mostly come out as riddles kind of yeah I think yeah it's such an interesting aspect of this yeah it kind of ties me back to what we were talking about earlier with like

discovery because it sounds like it's kind of like the discovery of tasks I guess the framing of discovery is usually in this sort of like auto research framing where it's like discovering new knowledge or you dump all your notes into your database and it's going to Yeah, I guess like entertain you with what is discovered compared to like discovering tasks and studying tasks. Yeah, pretty much. I think you can think of a discovery of new tasks or discovery of a procedure that itself can follow. And after following the procedure, it just automatically gains so much better understanding and more expertise over potential downstream exam. Yeah, so interesting. and maybe coming back to the human intuition I'm curious if you think a lot about the meta of how you study things and if that influences your work me personally I like to just kind of transcribe things like I'll read a paper and I'll transcribe it and for some reason that just works for me definitely

I think thinking about how we study has definitely shaped a lot of my approaches especially like if you I think the three bats that we were talking about earlier is pretty much or very tightly connected to how we study, right? For example, if we read papers, we are doing nothing but just like, you know, like reading word by word. And that sounds like that just seems to be some sort of self-supervised learning. Like next to competition, maybe. sometimes we we practice a little bit right we might get that quiz from the textbook or we might just sometimes like we after reading a chapter i'll write down some questions for my tomorrow self right can i do these questions tomorrow and if i can i am sure that i gain a pretty good understanding of the first chapter right um and we can also transcribe write notes or even just

annotate the paper, right, so that when we are reading the book next time, so maybe we can read the book multiple times, right, when we have nodes on the side of the book, and the next time we write, we are not only reasoning over the book, we are reasoning over our past reasoning, right, so there might be some sort of interest and synergy where we can study more carefully from how to Miller, and maybe that's also very important to future machine studying algorithm. Yeah. Jacob, welcome to the Andrew Huberman podcast. Thank you. Yeah, I tried that, like, active recall method where you would, like, kind of like you mentioned, like, write some questions for yourself to answer the next day, and I would do that kind of thing of, before I started studying, I'm going to write down as much as I remember, and I thought that worked pretty well. I did a lot of like the I think everyone is doing this thing where you give what you're trying to study to AI and say hey

design me a test yeah I think that's actually I also use that way to learn too yeah I'll get a paper and be like okay test me on the paper or something yeah it's really interesting thinking just about how this can teach us how to learn better to be interesting to give a like to give a machine setting algorithm the thing I'm trying to learn and see how it figured out how to teach itself best and maybe if that would transfer to teaching me yeah awesome uh cool so yeah I guess um maybe just sort of talking about these algorithms I'm curious if there's sort of like a vision for database systems and AI that I find really interesting which is like the sort of creation of big data and more people working with big data because imagine like say you or I were studying, say like we have eight papers and that you chunk that up in your database and

that creates maybe like a hundred unique objects or it's just the 15 papers or objects. But then these studying algorithms are like creating more data. I'm curious what you think about that sort of vision of like people's personal knowledge bases becoming really massive because we have these sort of always on agents that are just working with our knowledge i think that when we are thinking about better understanding um users own corpus is a good machine study algorithm is definitely very important because like because of the scattered nature of like a lot of human information like of like our own personal information like the the scattered nature of it um and that's why i feel like um well i'd say that maybe we we just we need not only it's like i just feel like this this requires a lot of engineering and uh smarter algorithm

because maybe all we should do is that we have an agent controlling how these information are scattered in the first place. How do we index them more carefully? Maybe agents can help with this. But when we are given this whole corpus to a new agent, there seems to be this missing capability that we kind of described here, right, that the model can really learn or, like, try to understand about you deeply yet. And we do believe that if this is possible, there will be so many exciting downstream applications that can be used, not only for personalization, but also, just as you said, like, maybe education. Maybe the model can understand more about how you're, like, what is your current understanding on this topic in a faster way, right? Yeah.

Yeah, and I guess maybe like the sort of clod.md files are kind of like occupying. It's like definitely the first thought I just had when I think about this. It feels like it's doing a little bit of a machine setting with my code base when I give it a task and it starts keeping its notes in the clod.md file. And the kind of deep personalization, I think the counter argument to that is like, well, I just bring that clod.md file around with me from like project to project. and, like, you know, it's coming on the road with me. Yeah, I think these scales.md, cloud.md are very handy. Like, I use them as well. For some reason, they contradicted with my own design philosophy a little bit, but I'd love to use them. Like, they're doing great. But, yeah. So I wouldn't be coming too much on this, but apparently they work for me too awesome well Jacob I think this work is already so forward looking but

I'd love to just know what directions for AI excite you the most well one very I think one very related area to machine is studying is evaluation as a science, right? Because if you think about all of these tasks we were just discussing, for example, how well your language model can or your agent can understand you, how well can it do deep research? It requires a lot of evaluations, and apparently not all of them are so easily defined. And let's say most of the time we've observed that as soon as there's a very good benchmark in a field, people will find a way. We are stealing human labor, right? We can find a way to help find these benchmarks. But there will always be a new field. Let's say how well does your model understand you instead of me, right?

you might need a benchmark yourself rather than using my benchmarks. And how these benchmarks are created are very interesting questions, right? Because to be honest, there's nobody understanding you better than yourself. But a lot of the time, or like when we are dealing with these types of questions, even though, let's say me, I know myself more than anyone, but I might not be capable enough to actually design a great benchmark. You know, the hard question for me might not be hard for agents. And this is such an exciting direction of how do we actually create benchmarks for these emerging fields, for this more nuanced domain, right? because it seems like we are talking a lot about, okay, recursive self-improvement.

But when we are actually talking about recursive self-improvement on understanding of a new domain where a benchmark doesn't exist, how are you even monitoring the progress? How do you evaluate studying algorithm at all? It is a very, very interesting question. and I think we'll briefly touch on this in our about-to-release paper. So, but like in general, evaluation as a science is such a fascinating topic said that interests me. Yeah, I think, yeah, just kind of what we're talking about with like the studying and just what questions to ask, how to particularly study. I guess I'm a little curious further into this idea of like knowing myself and just, I guess, your view on personalized agents and maybe just further into what you mean by that. Oh, well, for example, like, let me give you an example. Like, for example, if I'm teaching myself on a new topic,

let's say I'm studying quantum mechanics today, I just want to learn it more efficiently and more effectively. Why can't an agent teach me? But I feel like the first step is to have the agent study me, like, where I, like, where my understanding is at. Am I still capable of doing anything like high school physics or am I just like incapable at all? Right. So can these agents actually try to study me from, let's say, maybe my own history or maybe give me a short quiz and look at the answer. And then maybe just how how well can it actually understand me? And this sort of things can, I think, definitely fall under, like, machine setting questions, right? And while we were talking about this, how do we even measure this? Like, do you actually want to scale me as, like, let's have agents teach me 10 different fields

and actually let me learn them first and then see how I perform and to conduct some sort of RLL with them to teach some model better at teaching me. I think that will work, but I might be exhausted learning, like my or 10 unrelated field just to have the agent better teaching me about quantum physics. Yeah, that inspires me thinking about like instead of writing a short question, and writes an open-ended question to realize, oh, Connor actually is missing several concepts. Yeah, that's fascinating. Yeah, amazing. Well, Jacob, thank you so much for joining the WeVade podcast. I think this framing of studying intelligence, just this field, particularly I think like when you just read this paper and you're contrasting it with continual learning, the highlighting the missing gaps. I mean, who doesn't – like this framing is awesome. Machine studying, studying intelligence is so captivating. And congratulations on this work, and thanks again.

Thank you so much. Thanks for having me.