AI Won't Make Ordinary People Excellent, It Only Makes Excellent People More Excellent
The bottleneck for investment-research Agents isn't the model — it's data governance and human judgment. Using overseas models makes you dumber; using domestic models forces you to get smarter.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Agent projects stall on data, not on the Agent
Lao Bai says the delivery of a large number of AI Agent projects stalls not on the Agent but on data. To solve the problem you have to hunt for data everywhere, and much of it sits on different platforms with different specs and types. Many sites have anti-scraping mechanisms, so web search can't reach them; a single site may have tens of thousands or even over a hundred thousand pages, and figuring out which page is actually worth scraping is very hard. He believes data governance as a whole will be the most important core of getting AI Agent projects to land — even more than 50%. On the logic side, capability has already improved substantially since K3 came out, and the problems and solutions seen day to day are basically covered.
— Lao BaiUse someone else's Agent and you get a product manager's thinking
Lao Bai says that if you use your own Agent, its training forms internal memory files through your conversations with it, and the answers it gives you are your mirror. But if you use an Agent that has been trained and you can't train it, then that's that product manager's thinking, not yours. Every AI investment-research terminal on the market, whether Walk or Claw, is broadly hard-coded by its product manager, and users find it very hard to train such Agents further. To get a style that suits you, you have to find an AI Agent platform that is open at the底层 and train it.
— Lao BaiThe multi-Agent pitfall: calling a skill without its underlying files
Lao Bai points out that interaction between multi-Agent platforms has many pitfalls to avoid. For example, when you call another Agent's skill through one Agent, the result differs from when the Agent that owns the skill calls that skill. The reason is that when you call another Agent's skill from one Agent, that skill doesn't call the underlying files inside the Agent it belongs to — equivalent to all your Agent training having no effect underneath, with only the skill's capability handed to the current Agent. So multi-Agent collaboration has a certain threshold, and that's where the FDE's value lies.
— Lao BaiCodex and Claude Code write code, they don't do investment research
Lao Bai stresses that Codex and Claude Code are products of Web Coding, not Web Walking. For example, SuperPower is a very famous plugin on Claude Code — brainstorm first, then write the report according to the plan, then check it, very perfect — but this thing is actually not suited to investment research; it's for Coding, for breaking a problem into a project, writing it as code and delivering it, not for producing research reports. Workbody is an office platform for Workspace Agents, not an investment-research platform. FanBot is the multi-Agent platform built for investment research.
— Lao BaiUsing overseas models will make you dumber
Lao Bai says he talked with a public-fund manager yesterday and the two agreed: using overseas models makes you dumber, using domestic models makes you smarter. Because overseas models are too smart, you don't need to be very smart; but with domestic models you really do have to be smarter, otherwise the results are unsatisfactory. He explains why he has stuck with domestic models: using overseas models is so pleasant that you'll keep using them and won't spend as much time thinking. There's a theory overseas called AI causing brain degeneration — like a person lying in bed for 20 days, can they still walk normally when they get up.
— Lao BaiCuring AI laziness: delete the memory first, add checks last
Lao Bai says AI is lazy, because it wants to use the fastest way to tell you the best answer, so it makes things up for you. He gives a real scenario: a dozen-plus years ago a classmate was deputy general manager of a property company, and his biggest pain was that no one patrolled at midnight; later they put notebooks in various places in the building, and the security guard signed in at each point every day, and walking the whole route took exactly two hours. AI is the same. The first step of his Harness is to delete all memory entirely, so it can't use historical data to fool you; the last step is to add a layer of checks on the data. This is exactly the workflow SuperPower designs — head and tail greatly reduce the hallucination rate.
— Lao BaiAgents won't make ordinary people excellent
Lao Bai sums up: an Agent won't make an ordinary person especially excellent, but it will make an excellent person especially excellent. Because the Agent itself has many thresholds and demands on the person. He gives an example from three years ago: a person tells AI to make me a Taobao website — a very famous joke at the time. But in the first half of this year he talked with someone who had used Claude Code for a long time and used many overseas models, and that person still asked questions in exactly that way when using AI, and the result is bound to be disastrous. So AI actually has a threshold for a person's ability.
— Lao BaiLong-horizon retest: AI foresees memory capacity expansion
Lao Bai shares a scenario where AI is irreplaceable. He has a dedicated skill called long-horizon retest, with a growth-stock-leaning framework, requiring T+2 greater than T+1. In May and June this year he used this skill to look at four industries: memory, real estate, Chinese baijiu. At the time it appeared for one day on Feb5; he asked the question that day, and by the afternoon it was no longer usable. AI's judgment was very clear: at 70% to 80% gross margin and 60% net margin, supply-side discipline means Micron, Samsung and Hynix will definitely expand capacity; it also mentioned which fabs would come online in 27, and even said more capacity-expansion plans would appear in Q3 and Q4 this year. Lao Bai says that even an investment manager who has long watched memory would find it hard to collect this data or reach the same granularity.
— Lao BaiIn their own words · checked verbatim
The delivery of a large number of AI agent projects stalls not on the agent but on the data.
大量的 AI agent项目的交付 其实卡并不是卡在agent上 是卡在数据上
Lao Bai6:06
If you use an agent you trained yourself — but if you use, say, an agent that has been trained and you can't train it — then that's that product manager's thinking, not yours.
如果你是用一个自己训练的agent 但是如果你是用比如说像是被训练过的这种agent 而你无法去训练他 那就是那个产品经理的思路 而不是你的思路
Lao Bai13:10
Using overseas models makes you dumber; using domestic models makes you smarter. Why? Because overseas models are too smart, you don't need to be very smart; but with domestic models you really do have to be smarter, otherwise the results really are unsatisfactory.
你用海外模型 会让你越来越笨 你用国内模型 会让你越来越聪明 为什么 因为海外模型太聪明了 你不需要太聪明 但是国内模型 确实你要聪明一些 否则它确实结果是不如人意的
Lao Bai25:14
AI is very lazy. Why is it lazy? Because it wants to use the fastest way to tell you the best answer.
AI是个很懒的 那他为什么懒 是因为他想用最快的方式 告诉你最好的答案
Lao Bai45:18
An agent won't make an ordinary person especially excellent, but it will make an excellent person especially excellent.
agent 不会让一个普通的人 变得特别的优秀 但是他会让优秀的人 变得特别优秀
Lao Bai58:24
Using overseas models feels great, you feel good, so you'll keep using them, and then your thinking, your considerations will keep — how to put it — you probably won't spend as much time thinking.
你用海外模型很爽 就感觉不错 你就会一直用下去 然后你的思维 你的考虑就会不断的去 怎么说呢 就可能就不会花那么多时间去思考了
Lao Bai1:00:25
If you can ask AI good, high-quality questions, then the answers it gives you will actually affect you. I read this answer and I felt my intelligence was completely crushed in front of it.
如果你能问出好的AI AI这个高质量的问题 然后他其实给你的回答会效果到你 我看了这个回答 我就觉得我的智商在他面前 是完全被碾压的
Lao Bai1:15:33
Figures
| K3 parameter count | 2.8 trillion parameters | 7:07 |
| KIMI token price | eight times that of other tokens | 29:14 |
| H200 price change | from 1.5 million two years ago to 5.5 million now, a three-to-fourfold increase | 26:14 |
| Number of SuperPower skills | 14 skills | 19:11 |
| FanBot default rounds | 150 rounds | 1:08:30 |
| Number of metrics for listed chemical companies | ten or twenty thousand | 1:09:30 |
| Number of metrics used by a chemical sector chief | one or two thousand metrics | 1:09:30 |
| Number of industries analysed by the long-horizon retest skill | four industries | 1:13:32 |
| Memory industry gross margin | 70% to 80% | 1:14:32 |
| Memory industry net margin | 60% | 1:14:32 |
Glossary
- FDE / Forward Deployed Engineer
- A role popularised by Palantir, requiring business understanding rather than technical skill, using AI to deliver custom solutions for clients.
- MOE / Mixture of Experts
- A mixture-of-experts model architecture used by large models such as K3, using multiple expert networks to handle different tasks.
- Harness / agent framework
- The underlying runtime framework of an Agent, including memory, tool-calling and other configuration; different Harnesses produce different results.
- ETL / data cleaning
- Extracting, transforming and loading multi-source data into clean data that AI can recognise.
- QA / data quality check
- The quality-check step in data processing, flagging anomalies and cleaning them.
How to listen
Investment managers, researchers and FDEs who are landing AI investment research into their teams, and anyone torn between overseas and domestic models or wondering whether to build their own Agent platform.
The perks introduction and product promotion after 1:16:08 can be skipped.