The AI Barrel Paradox — The Era of Shoring Up Weaknesses Is Over

Start with an experiment. Harvard Business School and BCG recruited over 700 consultants — the kind of professionals who advise big companies for a living — had them do work with AI, then scored and evaluated the results. The finding: the group that had been at the bottom improved their overall scores by more than 40%. The most elite expert consultants improved by under 20%.

Which is to say, as long as you clear a certain bar, AI levels out the quality of what gets delivered, and the gap keeps shrinking. So the question is: are experts still worth anything?

The weakest-link theory needs a total upgrade

How much water a barrel holds is determined by its shortest plank — which is why our generation, from school through work, has been taught to shore up weaknesses. But as I see it, in the AI era this theory needs a total upgrade. How long the planks are no longer matters much. What actually matters is whether you have enough of them.

The people who got the dividend earliest in the AI era (possibly including me) got it for one real reason: crossing domains. On top of knowing how to use AI, they understand marketing and product, and maybe data and finance too. As long as a dimension isn’t a total blind spot for you — even if you’re only at 30 points on it — AI can pull you up to 70 fast. I call that the “breadth dividend.”

But that dividend won’t last long. Further out, whether you’re at 20 points, 10 points, or just aware that a thing exists, AI will take you straight to 80 or 85. At that point nobody’s output is meaningfully different, huge numbers of knowledge workers get compressed into the same tier — and their value depreciates as a matter of course.

What’s genuinely valuable is a strongest suit above 90

Renaissance Technologies on Wall Street runs the fund widely acknowledged as the most profitable in history. Put 10,000 yuan into it in 1988 and you’d have 200 million today, an annualized return approaching 40% — Buffett is only around 20%.

What did they get right? Back in the 1980s the entire market only recorded opening and closing prices. They saved every detail of every single trade and cleaned it up bit by bit — and that pile of data nobody wanted later became a gold mine nobody else could get. They didn’t hire Wall Street analysts. They hired physicists, astronomers, mathematicians, even philosophers. Their above-90 capability: seeing signal in what everyone else writes off as noise.

Proprietary data plus expert discernment — that’s the capability that can genuinely command a premium in the AI era.

And 90 points alone isn’t enough. To work well with AI, the capability that matters most is communication. If an expert can’t describe their whole body of craft clearly, AI can’t execute it automatically or at scale — no communication means zero. And AI fundamentally exists to serve human needs, so without it you can’t read what a client actually wants, you can’t market, and you certainly can’t close. The biggest problem most one-person companies and solo operators run into right now is precisely this — reading genuine demand, and selling.

How small and mid-sized companies should hire: three test dimensions

One, in a world without AI, what did they once accomplish that nobody else could — this tests the unique above-90 capability. Two, have them explain the thing they’re best at to someone who knows nothing about it — this tests expressive ability. Three, test their discernment on deliverables. Given two pieces of copy both written by AI, which is better and why? People often comment under my WeChat Channels videos that there’s no real difference between Codex and Claude Code — that’s a textbook lack of discernment.

At this point a lot of bosses will say these are all elite people and a small or mid-sized company like theirs can’t hire them. Correct — because you are that person. Stop thinking about hiring someone to use AI. The biggest lever AI can put under your company has you as its fulcrum. And in the era of the solo operator, you don’t have to bring all the talent in-house. You can partner — retained advisors, project-based outside brains. You don’t have to fight the big companies for experts.

A question I don’t have an answer to either

Over 40 years ago the economist Rosen described the “superstar effect.” Once technology lets the output of the very top be copied at near-zero cost, income stops being distributed by amount of labor and concentrates in a tiny few at the top. It shifts from “who works more, who works better” to “who owns the scarcest resource” — and that’s not even compute and capital anymore. How hard is it right now for capital to get into a good company? They’ve been profitable since day one. Why would they take your money?

Knowledge labor really is depreciating fast. So when we go and learn, go and improve ourselves — what’s the point? That traditional, gather-the-facts style of learning is really just a way of managing your own anxiety. I still don’t have a clear answer to this question today. Let’s talk about it in the comments.

Source: 7月24日.mp4 on-camera original recording · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] Today let’s talk about one topic. In the AI era, are expert consultants — those traditional knowledge workers — still worth anything? There was an experiment. Harvard Business School and Boston Consulting Group took over 700 consultants, the kind who normally consult for big companies, had them do their work with AI, and then scored and evaluated them.

[00:14] The experts who were at the bottom to begin with saw their overall scores go up by more than 40%. And the ones who were already top-tier improved by less than 20%. Which means: as long as you clear a certain bar, AI levels out the quality of what gets delivered. The gaps keep getting smaller.

[00:25] Everyone knows the weakest-link theory, the barrel. The old line is that how much water a barrel holds is determined by its shortest stave. And our generation, from school all the way into work, has been taught to shore up that weakness. In my view, in the AI era, this barrel theory needs a complete upgrade.

[00:38] How long a stave is doesn’t matter that much anymore. What actually matters is whether you have enough staves. The people who got the earliest dividends in the AI era — myself possibly included — the real reason is crossing domains. Vision and capability across multiple dimensions. On top of knowing how to use AI, you understand marketing and you understand product, and maybe you understand data and finance too.

[00:53] As long as I have no blind spot in a given dimension, as long as I’m at around a 30 out of 100, AI can quickly bring me up to a 70. I call that the breadth dividend. And this dividend won’t last very long either.

[01:02] So what happens after that? You won’t even need a 30 in that dimension. You have a 20, a 10, or you just know that such a thing exists, and AI can take you straight to an 80, even an 85. At that point there’s no fundamental difference between what anyone delivers. Huge numbers of knowledge workers get squeezed onto the same level. And what naturally follows is the devaluation of that value.

[01:20] But that doesn’t mean we should pour all our time into adding more staves. If anything, taking the long view, we should be growing a genuine strongest suit — something above 90.

[01:30] Take Renaissance Technologies on Wall Street. It runs a fund that’s widely considered the most profitable fund in history. If you’d put 10,000 in back in 1988, it would be 200 million now. Annualized return of nearly 40%. Keep in mind Buffett’s company only does about 20% annualized.

[01:42] So what did this company get right? Back in the 1980s, the whole market was only recording opening and closing prices. And they saved the detailed data on every single trade, organizing and cleaning it up bit by bit. That pile of data nobody wanted turned into a gold mine nobody else could get to. They didn’t hire Wall Street analysts. They hired a batch of physicists, astronomers, mathematicians, even philosophers.

[02:01] And what’s the above-90 capability those people have? Seeing signal in what everyone else sees as noise. Exclusive data plus expert discrimination — that’s the capability that actually commands a premium in the AI era.

[02:10] But having that above-90 capability alone isn’t enough. If you want to collaborate perfectly with AI, the most central skill is actually communication. All of an expert’s skill — if there’s no way to describe it clearly, AI can’t execute it automatically and at scale. No communication equals zero.

[02:24] And fundamentally, AI serves human needs. Without the ability to communicate, we can’t see into what customers really need, we can’t market, and we certainly can’t close. The biggest problem most of today’s OPCs — so-called one-person companies, solo operators — run into is really insight into genuine needs, and selling.

[02:40] Let’s flip the angle. If you’re a small or mid-sized company, what kind of talent should you be hiring next? You can test on three dimensions. Dimension one: without AI, what have they done that other people couldn’t do? That one’s about the unique above-90 capability.

[02:51] The second point is asking them to explain the thing they’re best at to someone who knows nothing about it. That tests their ability to give instructions and to express themselves. And finally you can test their discrimination on deliverables. Two drafts, both written by AI — which one is better, why is it better, and what exactly is the difference?

[03:05] People often come to my WeChat Channels posts and comment that Codex and Claude Code, the tools and the models, aren’t any different. That’s a textbook case of no discrimination. Someone like that can’t even judge which kind of deliverable is more valuable to our specific group of customers.

[03:15] At this point a lot of bosses will ask: everything you just described is top-tier talent, and as a small or mid-sized company I can’t hire people like that. That’s right. Because you are that person. Stop thinking about hiring someone to use AI. The fulcrum of the biggest lever AI can apply to our companies is you yourself.

[03:28] And in an era like this, the AI era, the era of OPCs and solo operators, don’t think about hiring every bit of talent into the company. We can partner instead. Retained advisors. Project-based outside brains. We don’t have to fight the big tech firms over experts. There are plenty of flexible ways to work together, and it’s a lot easier than it was a few years ago.

[03:43] Lately I’ve also been thinking about a deeper question. Over 40 years ago the economist Rosen put forward an idea called the superstar effect. Once technology lets the output of the very top be copied at nearly zero cost, income stops being distributed according to how much labor or delivery there is, and instead concentrates in a tiny handful at the top.

[03:58] It goes from who works the most and who works the best, to whoever holds the scarcest resource. And that resource isn’t even compute or capital anymore. Do you know how hard it is right now for capital to get into a good company? They’ve been profitable since day one — why would they take your money?

[04:09] Knowledge labor really is devaluing fast. So in a reality like that, where’s the point in studying, in improving ourselves? That traditional, picking-things-off-the-shelf kind of knowledge learning is just filling in our own anxiety.

[04:20] And to this day I don’t have a clear answer to that question either. So let’s talk about it together in the comments. That’s it for this episode. See you next time. Bye.