Managing Machines vs. Managing Humans. What's The Difference? | Simon Shaolei Du, Apodex

Large language model is still less random compared to human. Large language model listens to you, follows your instruction, but human are not. So it's still easier to organize.

We know that LLMs are non-deterministic. Wouldn't so many agents cause error compounding? Just throwing human bodies at a problem may not be the right solution.

100 agents than 100 people.

Tell me about Apodix. You moved from academia to the industry. What kind of product is it?

We're not just building a chatbot which can have a nice conversation with you. What we really What we want to do is to solve very hard problems, scientific problems, economic problems, something that can create a lot of value. Usually it is like an unsolved problem. So we are trying to use our heavy-duty server to solve the hardest problem in developing large language models. They are, you know, creating a loop. They are self-evolving to become a stronger and stronger model.

There are a lot of fresh graduates or people who want to transition their careers. They are very confused. How do you think prepare us towards a positive outcome?

Accelerate this work 10x or 100x. If you have some open source projects, especially agent, it's very hard to gather resource to train a large model. But for writing agent framework, you only need APIs. You can still build a very comprehensive, strong agent framework just using API. You don't need that many people to do the same kind of work.

Today I have the chief scientist from Apodex, Simon Zhu. He was a computer science professor at University of Washington.. And now he's building deep reasoning models and research agents that can beat the frontier labs. Simon, welcome.

Hi. Hello everyone. Very happy to be here.

Simon, when I was doing research about your background and your company, my research said that you're one of the most decorated young theorists in machine learning. And the honors you stack include Sloan Fellowship, Samsung AI Researcher of the Year, NSF Career Award, NVIDIA Pioneer Award, Intel Rising Star, MIT Tech Review Innovators Under 35, an AI 2050 Fellow, and so many. You know, this is so amazing, but I'm curious, how do you introduce yourself to a new friend?

Well, I'm just like a faculty in University of Washington, so I like to work with young people, especially smart young people, to work on hard problems. That's what I'm doing right now.

I love it. Tell me about UpPodX. You are now, you moved from academia to the industry and you're the chief scientist at UpPodX. What kind of product is it?

Yeah, so our vision is to build heavy-duty software. So we are building a large language model, but it's kind of different from like normal large language model, for example, GPT, Gemini, or like Claude. What we are building is like We are very focused on solving hard problems. That's why we call our model Heavy Duty Solver. So we're not just building, for example, a chatbot which can, you know, have some nice conversation with you. But what we really want to do is to solve very hard problems like scientific problem or economic problem or something that can create a lot of value. It can be usually it is like a unsolved problem. So that really requires a very strong solver to solve it. So that's the general vision of our model. And I would say, generally speaking, to solve a hard problem, the first step is to do thorough research. So our current product, or the first product we just released, really focuses on deep research. So deep research is a mode in actually several large language model products like ChatGPT, Gemini. They all have their deep research mode. But this is our first focus. So we are building, we will say, even better deep research product which can give you more complete and accurate research report than those like deep research output from GPT or Gemini. So that's the first product we have. Yeah, so this is a general introduction of what we have.

What do you think is the problem this model that you're building is really solving?

So we are trying to, so we are building a general heavy-duty software. So we are not exactly targeting one specific, very specific domain, but we do have some target domain that we are considering. One is the, I would say, drug discovery, drug repurpose, healthcare, this kind of medical domains. So we believe there's a lot of unsolved problem there, of course. And also it can create a lot of values from there. So that's one targeting domain we are seriously investigating. So in our team, we have a sub-team which we call Heavy Duty Discovery. So this team tries to identify the hardest and challenging and most valuable problem that we need to solve using the strongest AI model. And for now, they are researching the hardest problem we see in this biomedical domain. So that's the particular domain we are very interested in. There are also another domain is basically the large language model or AI development itself. That's why we always focusing on self-evolving, which is one of the goal of our company. So we are trying to use DeepResearch or our heavy-duty solver to solve the hardest problem in developing large language model. And so in this In this way, we are creating a loop. So we are self-evolving to become a stronger and stronger model. So we hope we can exponentially expedite the development process through all of the AI development. Yeah. And so those are two, I think, domains that we will definitely spend a lot of resource on, but we are also considering other domains. For example, material science, we're also considering maybe making new materials. Also generally, we are trying to predict the future. Which most relevant there is like finance. For example, you can predict the stock price. Of course you can create value.

Yeah.

So that's, yeah, that's another domain that we are interested in.

Yeah, that's amazing. So you have, you basically have different, you focus on different domains and these are deep research heavy domains like finance and biotech, material science. And you also have, you also are building the model in such a way that it's self-evolving. So maybe let's separate these two. First, let's talk about domains, right? What do you mean by, you know, you're focusing on deep research in specific domains? Does it mean that your models are specific?

Not exactly. So we are building a general heavy-duty solver. For now, it's a general deep research agent. So not only about the model itself, also there's an agent system around it. I may talk about later, but in training this model, we are not targeting on a specific domain. I mean, we do have, for example, training data from all kinds of domain, including biology, biology, physics, chemistry, engineering, finance, and so on. But we are not, say, you're not like, say, 50% from biology, not like that. It's kind of evenly distributed data across all domains, but also about the knowledge. So that's important. But a more important, I would say, capability of this model for deep research, there are, okay, there are several key capabilities. One is planning. So once you want to do some research, you first need to make a research plan. So that's called planning step. So we do have specific data for this planning step, like how do you decompose the problem into subproblems and you can attack each subproblem individually or separately. And then, so how do you decompose a very hard problem into subproblem and even you can decompose subproblem into sub-subproblem. So that's the planning and/or decomposition capability.

So is your planning phase, you mentioned you have data for, in your specific planning and decomposing a harder problem into small pieces, is your data about specifically doing scientific research?

So actually there are common patterns across domains. Once you want to decompose a hard problem, it's like once you, so first of all, for example, you ask, so first of all, it'll consider whether this problem is a specific problem or general open-ended problem. And based on this judgment, you will decompose problem, whether you will first decide whether I need to search more information and then based on information, further make a plan, or I already have enough knowledge in my model, I can already make some decision to decompose problem. So it's really like how scientists attack a problem. So we do gather this kind of data to help the model learn how serious scientists attack this problem or decompose this problem. So that's the first step. And yeah, we kind of gather those data to change the model's behavior to think like a scientist.

That makes sense.

Yeah. Yeah. Yeah. And yeah, so that's the first key capability. The second key capability is really about search. So search is, you know, nowadays we think search is a very common I know everyone use it, but actually it's still a hard problem once you want to search for some hard to identify information. For example, is the query you use in Google search matters a lot because, you know, different query, Google will return different answers, right? Like, or different order of the pages. And you hope that your query is good enough to gather the most useful information in the first page. That's how human, when you do search, it's like. So we train our model to do more efficient and thorough search to write better query so that we can get the most important useful information we want. Oftentimes it's about maybe some of the key information from some paper, from some report. So you need to first, so it's called the multi-hop problem. Basically, you need to first search for some keywords to gather the information, gather the report, and then maybe you further do some search on that report, or from that webpage, you further do some search to search for the next key information. So this is a general search capability, and there are some benchmarks to measure this. And on those benchmarks, we can get probably the strongest results. So this measure, that's a key, we call it atomic capability, which is search to gather the hardest or hard to find information. So that's the second key capability. Yeah.

When you say for search you beat the top of the benchmark, do you mean that you actually beat the deep research across the board, like OpenAI or Claude, you know, those deep—

Yeah. Yeah. So there's a benchmark. There are several benchmarks. One is called BROS Comp. It's developed by OpenAI. So basically they have, I think, 1,000-ish problems. Those are very hard to find, basically asking you to find some information that is really hard just by purely, say, prompting a model or purely do Google research.

Can you give an example? Like, what is a hard search problem?

For example, something like find a book name by 1990s and with the first word start with ZY, something like this.

That doesn't sound like it can be easily solved by Google search.

Yeah, exactly. So you need to have multiple steps usually. And so you may First search for books by some missing some time range, and the missing time range, you do another search. So it's like usually it's a multi-step search. We call it agentic search. So it's not like a single Google query search problem. We are building a system to solve, to improve the search capability. So on this benchmark, there are around 1,000-ish problems. There are easy problems and harder problems, and on harder problems, Sometimes when we can solve, we can find the answer, but they cannot. So there, I think an overall score, we are better. So that's why we are, we say, you know, we are, we are the state of the art or SOTA. We have the SOTA performance. So the reason we can get this strong performance is really about, so there are two things. One is about the data. Like we are building certain datasets so that we can improve the general capability of search. And secondly, we are building an agent system around our model. So we call it AgentTeam. So it's a multi-agent system to do not only about search, but overall this deep, the whole deep research process from the planning to search. And then there's also a last step, which is called summary. Basically, you summary all the information you gathered. So for now, for this system, it's a multi-agent system, which is a hard system or agent system right now. Because now I think every LLM company, they have something like this, like a multi-agent system. Basically, a single agent is what we called REACT mode, reasoning and then act. That's called REACT mode. It has its limitation because one of the hardest or bottleneck of a single large language model is context length. Now people often say, no, we have a context window of 1 million or maybe slightly more. However, 1 million is still not enough. So you need multiple agents so that they can process more information. That's why you need this multi-agent system. Okay. But this is not like super normal in the mid of 2026 because it's like, maybe this is normal maybe half a year ago, but it's now, I think now every company has a version of it. What differentiates from us is like, we also in this multi-agent system building, we also have agent to do verification. Okay, so if you think about it, when you do search, you know, now you can write anything on the website, right? So some of the results can be, you know, reliable, but some are not. For example, generally speaking, you know, the information from Wiki is like, you know, someone at least check it. So it's more reliable than a random post from, say, Reddit. So we also have specific verification called verification agent. To judge whether this information source is reliable or not. So, but this verification concept is quite general. It's not only about judging a source is like reliable or not, but this is very important, of course, but also about, for example, whether the reasoning from the source to a claim is correct or not. So this is like judging another agent's reasoning process is is correct or not. Okay. So that's another kind of verification. There are other kinds of verification, which also, we also do, for example, if it's about some coding task, you can actually write some tests to verify this program you write is correct. So we, but what we emphasize here is that we do have an explicit verification agent. It can be judging the source, judging the reasoning process, judging the code you write is correct or not., but we have an explicit sub verification agent to do verification, which actually is very important to improve the accuracy of the results we get either from the search, either for search or for final deep research report.

But what you're saying is you have different ways of judging whether something is correct or not.

The most common one is we call it LM as a judge. Basically we use basically just using our own model as a verifier to look at another, or the output of another agent. But the agent itself is still our model. It's just, you know, how we, the role are different. One, for example, one agent to process all the information gathered and then generates a claim. And another agent is a verifier to judge this reasoning process is correct. Or not. The difference is that their inputs are different. For example, the first agent's reasoning agent's input is like the search result. For example, another agent, this verifier agent, is not only about the search result, but also the reasoning process and final claim the first agent generated. So the second agent kind of needs more information so that it can judge whether the whole process is correct or not. So that's what basically what differentiates two kinds of agents. But you do need the second agent so that you can have this verification step and the final outcome will be more accurate.

How do you handle the large amount of context?

The current, I would say, technology of large language models, one of the key bottlenecks is that a single agent or a single large language model cannot hold a very, very long context, say more than 1 million Okay. But if you basically do divide and conquer, then you can decompose a very hard problem into subtasks. Okay. So that's why it goes back to the first point I make, like the decomposition or the planning is very, very important. So once you decompose a problem, you don't need a very long context for a single subagent. Okay. So for example, the original problem, it may take a very, very long time context or a lot of tokens, more than say 10 million tokens to solve. Okay. However, what you actually need as input to those subagents, to a search agent, it might be much simpler. Okay. For example, a single query to a subagent for some search problem might just be find a book or say find a math book in 1980 about calculus. So this is a very concrete one. Something like this. So for this specific problem, you don't need very, very long context. It's like, it's a reasonably context should be, you should doing a pretty good job. Okay. So basically the whole system is about decomposing a hard complex problem into small sub-problem, but each individual problem only require a relatively small context so that as a whole, you still have, you still need maybe more than 10 million tokens, but each a single agent doesn't need a very long context to solve the subproblem it needs to solve.

We know that LLMs are non-deterministic. Wouldn't the sheer volume of, you know, so many agents cause error compounding or having friction when coordinating them? Because if we use the analogy of thinking about human organizations, right? Oftentimes if you're solving a problem within a human organization, just throwing human bodies at a problem may not be the right solution, right?

Yeah, exactly. Exactly. So I think there are two things as we develop or emphasize that are really important for this kind of, you know, how do you basically, how do you organize a large group of agents or people? First of all, I would say that I think large language model is still less random compared to human. Their outcome is more predictable than general. Yeah, yeah. I mean, generally speaking, you know, the large language model listens to you. Follows your instruction. At least that's what they're trained by the human or not. So I think it's still easier to organize, like for example, 100 agents than 100 people. Yeah. Okay. Okay. So there are two things I think that are important. First of all is verification. Okay. Maybe I can talk a bit more detail. So verification agents, one single verification agent, it may require more than one, I would say reasoning agent or search agents. Output. Okay, so basically, although you may assign a lot of tasks to different agents, but we have several intermediate steps so that a certain step basically is like the aligning or the sync. It's kind of aligning meeting or sync meeting, just like when you're in a company. Those verification agents basically acting as organizer of this alignment meeting in a company. Basically, it gather all the information and then it will verify which one is correct and maybe discard those wrong ones. And then it will kind of reassign the task for the next step.

Okay.

So we just like how you organize a company, you know, why you have weekly meeting, because you need to sync all the information from this week's output. We also have some similar basis like this verification agent will verify all the information and then we'll reassign some tasks to the next step. So that's one very important step. Like verification is not only about verifying the correctness, but also about gathering all the information and then you can decide what to do in the next step. And then secondly, it's about redundancy.

Okay.

So in a company or human organization, usually we will not assign the same task to, for example, two or more people. Although I know some companies do to make some competition, you know, but we're not doing like this. So in spirit, maybe it's similar, but you know, RM agent is still cheaper than a real human, right? So we can let same agent, or sorry, we can assign same task to multiple agents. And because the LLM, they have some randomness in it, so we will have more coverage about the same task. We'll gather kind of more information than a single agent.

Okay.

Especially for complex, more open-ended problems. Even if you use the same prompt to the same model, because of the randomness, within logical model, we'll still generate maybe different answers. And this kind of diversity is also important because you may gather more complete answers or coverage or complete information about the same problem. So it's like, well, it sounds to me like statistical sampling. Yeah, exactly. You sample more. Exactly. You sample more. So it's not really like the, you know, about competition. It's really about, I would say, cooperation in some sense, even though their tasks are the same, but even you assign the same task to the same, sorry, different agents, although the different agents are in essence, they are the same model, but they can still generate more complete, diverse answers or information to you. Okay. So this redundancy is not new in computer science, you know, in system design, cluster, you know, you often have this kind of redundancy. Redundancy. But here is, it's more about gathering more diverse data or diverse information so that you can process, you can get a more complete picture about the same problem. So yeah, so basically these two things, one is verification, one is this redundancy. These two things are what make, you know, this, you can get a very complete and accurate answer to your deep research question.

Are you using this teams of agents for your team's work?

Yeah, I'm using this model, of course, kind of every day. I think this, especially when we do some deep research problem, I think it's very useful. And I think it's super useful actually, especially now we are, it's a quite generic problem. You can search it for how you organize your team, this kind of problem, or some what's the right research problem, or for example, what's the right biomedical problem we should study. And you can ask follow-up question to it as well. And usually our Apodex model can give you a very complete, comprehensive answer. So yeah, we are using it every day, actually.

Can you share an example question that you use for your day-to-day?

Building a larger model, one important aspect is about data, right? But how do you process the data? How do you label data? Or how do you add some tags? It's actually what kind of tags you need to add is a research problem. Like how do you organize it in the right way?

Way.

So we often search this kind of problems. So this is just one example. And also, you know, you can also do some research on building large model architectures, optimizers. You know, you can also do this kind of research as well.

So if I were to compare this research question and answering from a human perspective, a human only equipped with ChatGPT or Claude versus a human researcher equipped with Opodex, How's the workflow different? Or, you know, I have one question, I write a prompt, I pass it to either ChatGPT Cloud and our PodEx, and then what's going to happen next? Is a PodEx running longer?

It depends on the problem. Some problems may do run longer, some problems may run quicker time. But usually I would say if you have a complex problem, okay. Then our system will assign more subagents to make sure that the information we get is correct. So we'll cross-check a lot of evidences, claims. So for that, because our, this whole verification and redundancy agent-based redundancy system, sometimes especially for long complex problem, we can run longer. The interface is very, I would say it's very similar to deep research mode of ChatGPT and Gemini. Basically you ask some question, it will generate a report. So this interface is very similar. It's just that in the process we emphasize more on, we do a lot more verification, search, or do some research with some redundancy there.

It's like science fiction, right? You're asking basically an AI to predict the future.

People always want to predict future, either for money reason or whatever, right? So predicting future, I would say it's not the entire, especially in finance, it's definitely not a new problem. So I also work in a quant company before, quantitative finance, where they are basically trying to predict the future, like every, even every second, if you are doing high frequency trading. So, but it's very different from what we do.

Okay.

So, so for finance, when you try to predict future, usually the most informative information is about, is from the market data, basically the price and volume, the bar chart, you know, the chart you see on those finance, like Robinhood. Whatever.

Yeah.

So, so, but those are very informative. It does have predictive power. For example, you can predict the next second, next minute, next hour, or next day. You know, the longer this horizon is, the less accurate it is.

Okay.

So this is like, I would say that more or less a traditional or yeah, or the classical quant research is about, which trying to use the market data to predict the future price. And you can predict better than 50% accuracy, then you can, generally speaking, you can make some money. Okay. Yeah. So, but what we are doing is different. We are more, our difference is more like the fundamental analysis of finance.

Okay.

Where you analyze this company's outcome based on, for example, its upstream downstream sector or company industries. And so you have a like holistic view of this companies, for example, next month's, next year's earnings, whatever. So that's what you are trying to predict. So this process is really— so now, you know, different financial institutes, they all have their, their researchers and, you know, fund managers doing this kind of research, trying to gather more information about a company, about sector, about the whole industry, so that can predict in the future what's it like, what the company's will, will, you know, stock price go up or, or go down. Okay. So our, basically, if you think about it, this is exactly what deep research is about. You are trying to gather all the information about this company and then you're trying to predict the future of this company. Yeah. So this is, so I would say, you know, deep research, this is predicting future in, but this kind of predicting future is really about deep research. So, but this only goes beyond this finance. You can also predict predict, for example, the outcome of a match of a World Cup, right? Because you can predict the outcome of a sports game is also about first gathering all the relevant information for even about, for example, whether the player is strong or not, whether how good is their cooperation, how good is the coach, whether some players are injured or not. You know, you need to gather all the information to make the right bet.

When you say you have data, do you mean that you actually have the data of how, you know, the expert's knowledge?

Is very different for quant research. Quant model is more like a classical machine learning model. Basically you are predicting a number. In quant, generally you predict the return of a stock, okay, for one day or something like that. And the input is usually, or the most important input is market data. It's like price, volume, this kind of data. Quant companies, Citadel, what I heard is like they definitely use some news data as well, but they first transform the natural language news into some features as still numerical. Yeah, yeah, exactly. Into numerical numbers and then use those as an input to this model itself. That's how, you know, traditional quant predicts the future or the stock price. But here, Large Language Model, it's a generative model. It generates, say, natural language with reasons. Okay. So that's more like a fundamental analyst, those researchers to analyze the stock, right? Basically you gather all the information. It may not be even quantitative. Like it can be some kind of news, some words from the CEO, things like that. And you make some judge, human judgment. That's how the, I think non-quant financial institute analyze stock, right? So basically, a large language model is more like this because large language model now it can reason, it can think. So it just behaves like a researcher in this kind of non-quant institute, financial institute to analyze a stock. However, you know, large language model, it can, I would say just in terms of search capability, it's much stronger than human right now because it can, like I said, it can do multi-hop search, things like that. And you know, it can run 24 hours, so we can analyze a lot of information than a human. So I would say this kind of deep research capability is generally stronger than a single human, but of course you need to train it to make it happen. So I think there's a, that's how the main difference from the quant, like it's more like the more traditional financial institute, how they analyze a stock. Yeah. The data for quant and non-quant, quant and non-quant model, or we say the quant quant model and the large language model are very different. So for quant model, usually it's very structured. You structured, you try to structure everything into a numerical format, like what do you call a feature vector, things like that. But here, you know, large language model just can deal with natural language directly. So that's very like what a normal human does to process this kind of information. So that's the main difference. And of course you need to train, how you train it is like you need to train the model to have, for logical model, you need to train the model to have this judgment, reasoning, analysis capability. And you need to have good, either using, for example, good analysis data, either from usually like from real humans, expert, how they analyze. So they gave this kind of data so you can mimic their behavior, or you can use, say, reinforcement learning so that you can, when you do correct prediction, you get some reward. So will somehow change the model to get better prediction capability. Yeah. So there are several training methodology to enhance this kind of capability.

And which ones are you using?

So we use all, like we use all kinds of training methodology. Yeah. Yeah. So, yeah. So I think that's standard. I'm not claiming this is something super novel. It's like standard. People use supervised fine-tuning and reinforcement learning. Although the details are very different, but at the high level, those are standard steps.

Who should be using Opodex?

I think as long as you want to do some serious research on anything, you should use Opodex. Like we have discussed, for example, if you are a scientist, you want to research something about your own domain's problem, like biology, physics, whatever, you can use it. It can do very thorough research to give you very— it basically can generate a paper about what's the question you want to ask.

Task.

Okay. Or, you know, like another example we talk about, like about finance, or just like you want to predict some future event, you should use AlphaDex.

Do you think a domain expert team should think about using a commercial solution like our products, or should consider building a system themselves, like multi-agent system themselves?

So first of all, I think there are common patterns about reasoning and deep research, it can generalize with different domain. Okay. We specifically gather this kind of data. So our, I would say, search, deep research, verification capability is super strong compared to other model. And we also support you upload your own data just like as an information source. So we can also, our model can also use this information source as well. And so generally, I say I think it's highly non-trivial to build the agent system around like a raw model. So I mean, if they have super, super strong scientists and engineering team and resources like GPU, I mean, they can build their strong model, but that will take a very long time. Usually it will take a very long time and a lot of resources. I think the easier way is to use Apple DeX.

How do you think that that can affect our lives? In the future if we outsource all these research to agents?

If you look at the deep research process, right, there's a lot of, like, if you ask a human to do the kind of deep research, there's a lot of labor tasks, like do some search, you do some editing, gather information. So basically our deep research agent can automate this altogether. And because we have verification, the outcome is very accurate. So now it basically gives you more freedom to think about what is the most important problem and to do risk, like what kind of risk, what problem you need to do research. So this is still up to, at least for her, I think for now it's still up to human. Maybe in the future, like AI can also tell you, you should do some research. Yeah. So that's more like a sci-fi situation. For now it's still based on human judgment to do what kind of research or what kind of question you want to answer.

Right.

I think this is still a more like subjective thing because because different people may have different opinions. There's no single answer that you can only rely on, at least for now. Yeah. So I think it gives people more freedom to explore more, to think more about what is the right question instead of doing all this kind of laboring work during this deep research process.

Is there a part of that world that would worry us or worry you?

I think different people may have different opinions. I'm generally optimistic. I think every new technology will create more more opportunity and more kinds of work people can do. It's maybe beyond our imagination. So, yeah, so I'm still optimistic about, you know, people can do more creative, interesting work using AI.

CEO Eric was like booed by the students when he was talking about AI. There seems like a little bit sentiment about like people are confused about the future. How do you prepare us towards a positive outcome?

I think we are at a transition point, right? Because now the work we have still, the work we had before large language models came out, for example, software engineering, like frontend, backend, many kinds of office work, things like that. Those are, I mean, those are, you know, AI can accelerate this work quite a bit, like 10x or 100x. Yeah, I mean, we should admit like you don't need that many people to do the same kind of work for this kind of work. But I believe there are new work or new kind of jobs people will create. It's maybe even beyond my imagination. But generally speaking, for new grad or undergrad or high school students, I think the fundamentals are always important, like the reasoning, like human reasoning, like other, you know, the common sense. Those are always important because they're always a step, you know, human can make some judgment. And for people who want to transit, or for example, software engineers who want to transit to large language models, I would say, so now this job market is like this. I don't think large language model job is super hard. It requires a lot of IQ, intelligence. I would say it's really about you should be very responsible and you need to be hardworking. So those are more important characteristics of a good candidate we are looking for. However, it's very hard to identify those signals. So if you look at all this frontier lab, how they hire, referral is probably the most common way to get a job in those frontier lab or any place that are building the large language models. Because whether you are hardworking, whether you are responsible, whether you are keen to detail, those are characteristics that are hard to be identified on your CV. So the most reliable way is just asking some referrals. However, I think there's one way that maybe is more common now is like if you have some open source project, especially on agent. Well, so why I say agent, because it's very hard to get resource to train a large model, but for writing some agent framework, you only need API. So you can still build a pretty, actually very comprehensive, strong agent framework just using API. So I think that's probably the near-term kind of direction people can do to, if you want to transit to some doing some large model related work.

But you are building something to replace this human reasoning process, right? How's it different?

To me, it's like human is always like a verifier for large model because you ask language model problem questions and then will generate some answer, it will interact with In the past, you may need to do reasoning on your own, but now you let the Larka model generate answer with its own reasoning. But you need to be able to reason in order to verify or look at the answers from Larka model, whether it is correct or not. I think so. See, that's why, I mean, because you need to make the final call or you need to judge whether the response from Larka model is good or not. To have this judgment capability, you still need to learn to reason unless you put everything, you rely everything on large-scale model, then sure, you don't need reason. But as long as you are still interact with large-scale model, you still need reasoning, the judgment to even more, to use large-scale model more effectively.

You know, it's really interesting. I have a, you know, this is a thought process I'm thinking. The how do someone like get trained if they don't get exposed to such opportunities to train that human judgment by doing it, you know, going through the process of the quant finance analyst, the whole process themselves. Because what I can do is this, okay, you have 150 agents who did all the analysis for me and I'm going to be the final verifier as a person. What I'm going to do is I'm going to put the results to another large language model and have it to judge it for me. I will, I will have a result. I mean, if that's the end outcome I'm looking for, but it didn't pass through my brain at all. I've done it myself. I literally, you know, we probably all have done it right from one AI to another AI, right? I think we still grew up in an era that we still get trained from scratch, but I don't know about the next generation, right? What are they going to do?

Hard to predict. That's a hard problem, like, because I think they, from when they started, school, they already have this, all the AI tools and all the reasoning process is done by AI, then I don't know what their brain will be like after they graduate. I generally believe learn by doing something, even though with AI's help, you still learn the whole process. I think that's still very helpful and really can improve your understanding about Larka model itself. Also the reasoning or general system building capability. Generally speaking, just like you work with some people, if you know these people better, then you can ask the right question, you can cooperate better. So I think maybe the next generation students are trying to have a better, how to better interact with large language models. This is a very important training subject.

Are you concerned at all about OpenAI or Claude, their deep research progress?

Not really. I would say they are for-profit companies and I know, I mean, they're focusing on coding agents. So I'd say actually the deep research is not their focus right now. And as far as I know, actually they put even more resource, people or compute, maybe compute as well for deep research. Than them, so I'm not too worried about they will catch up.

What's on your roadmap?

For now, we are still a deep research agent, so we are building a heavy-duty solver.

You're saying heavy-duty solver?

Yeah. So heavy-duty solver is our, the vision of our company, right? So heavy-duty solver is not only about deep research, although deep research might be the the first important step. So after you do this research, you may have a concrete plan to execute, but you still need to execute. For some problem, execution means coding. But here, maybe coding is not about software coding, it's about scientific coding. For example, you need to use some biology tools, just like computational biology or bioinformatics, those kind of tools. And you write some system to solve a computational biology problem. So we still need to, coding is very important, of course, like you need to have coding to really execute something. Also, so that's one way. Another thing is after we have some deep research plan, we also need to collaborate with a wet lab to do real experiments. Yeah. So that's another thing we plan to do in the future. So yeah. So by having Unity software, we really mean end-to-end solving this problem, not only about planning, which is basically what deep research does, but we are trying to really solve the problem.

If you were to teach me one thing to help me become a better founder or a better AI user or researcher, what would that be?

Yeah, so I think nowadays this system-level thinking is very important. Like we call our agent system Agent OS operating system. I really use the idea from operating system. It's like a classical, it's a classic in computer science, like Windows, Mac, they're all operating system. But if you understand how this operating system works, how was the structure, how they organize, that's really helpful nowadays to understand all this, the whole large model system. I think it becomes more and more important nowadays to think about in a system system way.

Could you recommend a book to our audience? Maybe fresh graduate, younger version of yourself.

I think the operating system textbook might be the most useful one. I would say, yeah, yeah. I think this, I mean, system, I think operating system is not only about computer science. You can think of more broadly. It's about organizing a system, although operating system is like how basically instantiating this idea into a computer system, but I say it goes beyond computer science. So I think learning this structure, how it works, is generally helpful. Yeah.

Thank you so much, Simon.