The central point of “Why one AI is better than four” is quite counterintuitive: a group of AI agents can perform worse than one AI that simply has all the relevant information.
Azeem Azhar is discussing what amounts to an AI version of the classic hidden-profile problem in group decision-making. In Anthropic’s experiment, four agents each possess only part of the evidence. The evidence shared by everyone points toward the wrong answer, while the decisive evidence is held privately by only one or a few agents. To succeed, the minority agent has to insist on its private information and the others have to believe it rather than follow the apparent consensus. Most model families managed this in only about 17–36% of runs, whereas one agent given the entire evidence base was correct nearly all the time.
That leads to a much more important conclusion than simply “four agents are bad.”
1. Multi-agent AI does not automatically give you diversity
We naturally imagine:
one AI = one opinion
four AIs = four perspectives.
But that assumption may be false.
Azhar notes that current LLMs are surprisingly low-variance: ask many instances of essentially the same model to solve something, and they often converge on remarkably similar approaches. He gives the amusing example that, when thirty agents were assigned the same coding task, eighteen independently chose the same Git branch name.
So four agents may really be something closer to:
one cognitive style × four copies.
This is important for all the recent enthusiasm about agent swarms. Multiplying agents does not necessarily multiply intelligence.
It can multiply conformity.
2. Discussion can actually destroy information
This is the most interesting part.
Suppose A, B, C and D all know facts 1–5.
But D alone knows fact 6.
And fact 6 changes the conclusion.
What often happens during discussion is not:
“Excellent, D has unique information; let’s update.”
Instead, facts 1–5 are repeated and reinforced because everybody possesses them. Fact 6 appears anomalous, receives less conversational weight, and gets effectively washed out.
Earlier experiments cited by Azhar found that combining several models’ responses preserved only about one-quarter of the useful ideas that had appeared in individual answers.
In other words:
aggregation can be subtraction.
That is the really important finding.
It reminds me of your recent language of fragments: occasionally the valuable thing is precisely the orphan fragment that does not fit the consensus.
A committee—human or artificial—has a tendency to grind that fragment smooth.
3. Human groups have institutions; AI agents largely don’t
Azhar identifies a second problem that I think is even deeper.
Healthy human institutions have mechanisms such as:
reputation, accountability, appeals, protected dissent, expertise and minority rights.
If the junior epidemiologist says, “Everyone else is wrong; look at this datum,” a well-designed scientific institution at least has procedures by which that claim can survive long enough to be examined.
Four LLM agents chatting with one another don’t really have equivalents.
There is no tenured dissenter.
No whistleblower protection.
No reputation accumulated from having been right twenty times before.
No institutional rule saying:
“Before reaching consensus, explicitly examine every piece of unique evidence.”
So multi-agent AI currently has something resembling society without institutions.
And society without institutions does not necessarily become collective intelligence.
It may become collective stupidity.
4. The answer therefore isn’t necessarily “one AI forever”
Azhar is not really arguing that we should abandon multi-agent systems.
His stronger argument is:
Multiplicity only helps if the multiplicity is real.
He approvingly cites Thinking Machines’ idea of developing an ecosystem of AIs raised differently, trained for different purposes and possessing different values—a way of “keeping the weirdness alive.”
That phrase is excellent.
Because what you actually want from four AIs isn’t:
GPT-A, GPT-B, GPT-C, GPT-D agreeing efficiently.
You want something more like:
Darwin + Freud + Nagarjuna + Wittgenstein sitting at the same table.
They should initially misunderstand one another.
Otherwise there isn’t much point in having four.
5. This has a surprisingly important implication for AI alignment
There is a hidden political philosophy here.
Much alignment discussion worries about AI systems becoming too autonomous, eccentric or unpredictable.
But this research exposes the opposite danger:
AI may become too consensual.
If models share similar training corpora, RLHF regimes, benchmark incentives and institutional norms, then even apparently independent agents may occupy approximately the same attractor basin.
Thirty agents then do not produce thirty perspectives.
They produce something like:
thirty samples from one civilization’s central tendency.
And the odd idea—the exception, the outlier, the thing that initially looks wrong—gets suppressed.
That is dangerous because genuine discovery very often begins as an outlier.
6. There is also a second argument in the newsletter: Jevons paradox
The issue has another interesting observation. AI inference is getting cheaper, yet usage isn’t exploding as dramatically as a simple Jevons-paradox argument might predict.
Azhar reports that a 10% fall in token price produces only about a 12–18% increase in token usage—enough that total spending rises somewhat, but hardly an explosion.
His explanation is very good:
the economically meaningful unit isn’t “a token.”
It’s “a useful unit of completed work.”
Nobody really wants cheap tokens.
They want:
- a bug fixed,
- a report finished,
- a drug candidate screened,
- a contract reviewed,
- a working program shipped.
And knowledge work is notoriously difficult to divide into cleanly priced units.
That’s why, he notes, AI expenditure is extraordinarily uneven: since October 2023 the top 1% of firms increased AI spending per employee by about $6,542, while the median firm increased it by only $9.63. Software appears especially advanced because useful output is comparatively measurable—commits, pull requests, releases.
So both parts of the essay actually contain the same deeper idea:
more computation ≠ more useful intelligence.
Four agents aren’t necessarily better than one.
More tokens aren’t necessarily more economically valuable.
The missing variable is organization.
And this connects unexpectedly well with what we were discussing yesterday about one—two—three—zero.
One AI with all the evidence can sometimes work cleanly.
Two or four agents can become 關係糾結: recursive attention to one another, reinforcement of shared information, consensus formation, extra coordination cost.
A genuinely productive three, however, would require something else: each participant oriented toward the third thing—the world, the evidence, the problem itself—rather than toward maintaining agreement with one another.
So I would translate the lesson into your vocabulary this way:
一：一個 intelligence，直接面對世界。
二：intelligences 開始彼此建模。
偽三：大家形成 consensus。
真三：各自保留差異，共同朝向那個不屬於任何人的世界。
And perhaps the most valuable future AI will not be the AI that agrees best with the other AIs.
It may be the one that can say:
「等一下，你們都漏掉了地上的這一塊碎片。」
That, I think, is the real point of #598.
I can also keep an eye on Exponential View for similar pieces about multi-agent AI and AI economics.
