Author: Jason, Founder of NDV

This article discusses macroeconomic mechanisms and public information, does not involve any fund products, and does not constitute investment advice or promise of returns.

In the past, doing "20 Minutes of Non-Consensus", most of the time I was alone, spending twenty minutes to clarify a judgment.

This time I want to try a different approach.

It is not a brand new program, nor is it turning interviews into resume Q&A. It is still "20 Minutes of Non-Consensus", but I will invite some friends periodically to sit down and chat about a topic that I am currently thinking about but have not fully figured out.

I increasingly feel that when a person outputs for a long time, the answers become skilled while the questions may become more fixed. When two people have a genuine back-and-forth conversation, the advantage is not the doubling of information, but that the other person can pull you out of your habitual path.

In this episode, I invited Jason Kam. We recorded for nearly two hours, discussing how two fund managers, who were not originally programmers, started using AI; we also talked about what AI can do for people, what it cannot do, and what truly scarce resources exist in investing when information becomes cheaper.

1. Why Jason Kam

Kam is the founder of Folius Ventures.

He studied in the United States in his early years and later worked for many years in research on Wall Street and hedge funds, before transitioning from traditional finance to digital assets. The paths we have taken are somewhat similar: we have both worked in traditional finance, both later entered new asset classes, and both face the same very direct report card—judgment is ultimately determined by results.

However, our approaches to work are quite different.

Kam likes to build systems. He breaks down a question into several roles, allows different AIs to research separately and refute each other, and then has another role summarize the findings. In one of his public processes, there are 6 AI analysts, going through 5 rounds of debate, finally handing it over to 1 "Research Director" for consolidation.

I used to be more accustomed to treating AI as a very capable assistant: breaking tasks down, clarifying boundaries, and handing them one thing at a time. I found it easy to accept which data to look for, which public materials to organize, and which repetitive work could be automated; but putting the entire research chain into a black box still instinctively made me uneasy.

Both of us, who are not programmers, eventually developed two very different usages. That is why I wanted to talk to him.

Actually, before the recording, we had dinner together. What started as an ordinary gathering turned into a three or four-hour conversation. I later realized that there was a rather interesting chemistry between us: Kam easily pushes a question further into systems and mechanisms, while I often interrupt him halfway—hold on, what practical use does this have for an ordinary person?

In this episode, I really did interrupt him like that.

2. What exactly did we do with AI

At first, Kam talked about a big change: advances in energy, chips, and models are rapidly lowering the cost of "intelligence."

I said, I have to interrupt for the audience. We have all heard this story to some extent; let's talk about something more concrete.

You haven't seriously written code for over a decade, and I have never written code. So what exactly are we doing with AI now?

Kam provided three very specific answers.

First, information scanning. He has the system continuously monitor public social media, forums, research materials, and a selection of podcasts, not to generate a pile of summaries every day, but to find out if there is a question worth pursuing during that time.

Second, building a research map. Previously, researching a company might start with Googling, then reviewing financial reports, earnings calls, interviews, and industry materials. Now AI can lay out the materials, indicating where there is evidence, where it is merely secondhand reports, and where something is still missing.

Third, automated repeated tracking. Company updates, changes following financial reports, and both sides of the same question can be run through by machines first. People do not need to start from scratch but still must return to the original text.

He also created BidClub, turning a range of investment podcasts from China and the U.S. into a searchable spoken database. When visiting the official website on August 16, 2026, the page listed 29 programs and 2,251 episodes. In the past, you needed to listen to each episode one by one; now you can first find out which segments are worth listening to and then go back to hear the original words.

My initial reaction was: what AI first replicates is not fund managers, but the workload of fund managers.

In the past, a person who couldn't code would often get stuck in execution even if they knew what they wanted to see. You cannot re-scan dozens of sources daily, nor can you build a toolset for every thought. Now, that barrier has suddenly lowered. Many work habits that previously existed only in one's mind can be broken down, tested, and even run repeatedly.

This is certainly useful. But the next question is more important: when everyone can read more materials, why aren't judgments becoming easier as well?

3. The more information there is, the harder it is to find the main contradiction

Kam is not blindly optimistic about the system he built.

He said something that struck me: if you draw a circle for AI, it becomes difficult for it to jump out of it itself.

It can dig deeply into a question. The issue is that the question may have been wrong to begin with. The longer the dialogue continues, the more likely it is to work very hard in an unimportant direction, ultimately providing you with a complete, smooth, yet useless answer.

The hardest part of investing has never been listing all the variables.

Interest rates, competition, products, management, valuation, policies, funding situations—each is important and could fill dozens of pages. The real difficulty lies in sifting through the sea of information to find three or four contradictions that will change the conclusion and then judging which one the market is currently trading on.

AI is now very good at answering questions, but not so good at deciding what the question is.

This also explains why, having more information does not equate to deeper understanding. If a person doesn’t have a framework in their mind, AI merely allows more material to pass through them at a faster rate. Kam’s original words are direct: without a framework, all information is noise.

I have similar feelings myself.

AI can tell me when a certain relationship reaches a historically rare position; it can also pull out previous similar situations. However, whether this time it will revert to an old path, or because the environment has changed, making past comparisons irrelevant, it cannot decide for me.

It can widen the search radius and create a solid opposing viewpoint. Ultimately, which piece of evidence is truly important and which merely seems fresh remains a human task.

Therefore, as we talked, we actually formed a very simple consensus: AI can help you find material, but it cannot decide what you should believe; it can replicate processes, but it cannot bear the consequences.

4. What I took away from this conversation

The first thing I took away is: first replicate the work, do not rush to outsource judgment.

If something is repetitive, time-consuming, and has clear sources, and you can quickly check its correctness, it is suitable to hand over to AI initially. Finding information, listing differences, organizing original texts, and establishing timelines all fall into this category. Conversely, anything involving "which variable is most important" or "whether this is really similar to the past" should be retained for human handling first.

The second thing is: the first step for ordinary people is not to pursue more information, but to build a very rough framework by themselves.

I asked Kam in the show, if an ordinary person wants to use AI for research, what should they do first?

His answer was not to learn prompt words or to subscribe to more tools, but to personally create the simplest unit economic model.

In layman's terms, it means writing down the smallest operational formula of a business: who pays, why they pay, how often they pay; where the revenue comes from, what the largest cost is, and where the real bottleneck for growth lies.

This model can start off very roughly, even with many blanks. But once you have written it down, each new piece of information will have a place to fit. Otherwise, the difference between reading ten reports and reading one hundred reports may just be that you have collected more sentences that cannot be judged for weight.

The third thing is: results and process should use two separate report cards.

We talked about something not easy to discuss openly: short-term results can inadvertently train a person's actions. When performance is good, it is easy to mistake luck for ability; when performance is poor, all fluctuations are interpreted as personal errors.

Kam’s approach is to simplify, stretching out the report card to avoid being shackled by overly frequent scoring. My approach is more additive: engage in new projects that do not have a direct relationship with the existing report card but can provide real feedback. One removes negative signs, while the other adds a few positive signs.

The directions are different, but they both address the same issue: do not let a short-term result monopolize all your explanations about yourself.

5. If you want to try it out immediately

You do not need to set up a complex system first. Find a company, industry, or even a small business you are interested in, and do an initial screening of research for an hour:

1. Write down in one sentence how it makes money: who pays, how much, how often, and what the largest cost is.

2. Leave blanks for the unknown parts, and let AI only find the original sources, providing the source name, publication date, and link, and categorize the content into "known facts, estimates, and unknowns". Do not let it give conclusions yet.

3. Reduce the truly judgment-altering variables to a maximum of three, and then have AI find the strongest supporting and opposing arguments for each variable.

4. Finally, write your own sentence: if any of the facts are proven false, how will I change my current views?

These four steps will not directly give you an investment answer nor should they be used to make any buying or selling decisions directly. What they truly train is something else: slowly transforming "I have looked at a lot" into "I know why I think this way, and I know what could prove me wrong."

At the end of the recording, Kam expressed a very honest concern: AI can propel a person very quickly and very far in a new direction, so quickly that after a long while, they may realize that is not at all where they wanted to go.

My reaction is somewhat different. I feel that many directions are originally discovered by collision. If one stands still optimizing the route out of fear of making a mistake, they may never encounter anything truly interesting.

This is probably the most valuable aspect of talking with friends. We may not reach the same answer in the end, but the other person helps me see what personality my answer relies on and what costs it overlooks.

AI makes starting easier, but also makes taking the wrong path quicker.

Thus, the question this episode leaves me with is not "Will AI replace fund managers?" but rather a more specific question: as both information and execution become cheaper, do we have the capacity to recognize earlier that we are answering a wrong question?

This is also why I want to continue inviting friends to chat.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。