System.Xml.XmlElement
Ramez Naam is, in my opinion, one of the best futurists alive.
在我看来，拉梅兹·纳姆是当今最优秀的未来学家之一。
But although he has a Substack (which you should definitely follow), he doesn’t write nearly enough.
但是，尽管他有一个子堆栈（你绝对应该遵循），但他写的还不够。
So I’ve been encouraging him to write more by cross-posting his posts on my blog.
因此，我一直在我的博客上交叉发布他的帖子，鼓励他写更多文章。
Somewhat to my surprise, Ramez has turned out to be a lot more of an AI skeptic than I would have though.
令我有些惊讶的是，拉梅兹比我想象的更加怀疑人工智能。
He’s certainly not dismissive of the technology, but he disagrees with many of the more expansive and dramatic claims coming out of the big labs (especially Anthropic).
他当然不是对这项技术不屑一顾，但他不同意大型实验室（尤其是人类实验室）提出的许多更广泛和戏剧性的主张。
In his first post in this series, he expressed skepticism that recursive self-improvement (RSI) would lead to a fast takeoff of AI capabilities in the near future:
在本系列的第一篇文章中，他对递归自我改进（RSI）是否会导致人工智能能力在不久的将来快速起飞表示怀疑：
In this post, he expresses skepticism about the potential for agent swarms — large numbers of AI agents working in parallel — to offer dramatic increases in AI capabilities.
在这篇文章中，他对代理群（大量并行工作的人工智能代理）能否显着提高人工智能能力表示怀疑。
Agent swarms have been in the news a lot, both because of the Hugging Face attack , and OpenAI’s solution of the famous Navier-Stokes problem is also widely attributed to a large swarm.
代理群经常出现在新闻中，这既是因为 Hugging Face 攻击，而且 OpenAI 对著名的纳维-斯托克斯问题的解决方案也被广泛归因于大型代理群。
But Ramez argues that parallelization is just another kind of input that suffers from diminishing returns, and that we should therefore temper our expectations about what agents swarms are able to achieve.
但拉梅兹认为，并行化只是另一种收益递减的输入，因此我们应该调整对代理群能够实现的目标的期望。
As was the case with his previous post, I’m not sure how much this matters, when the capabilities of the underlying models — swarm or no swarm — are still improving at such a steady and impressive rate.
与他之前的文章一样，我不确定这有多重要，因为底层模型的功能（群体或非群体）仍在以如此稳定和令人印象深刻的速度提高。
But it’s a very intellectually interesting argument, and Ramez can be very persuasive.
但这是一个在智力上非常有趣的论点，拉梅兹非常有说服力。
In my previous post , we asked whether AIs could improve themselves all the way to absurdly high levels of superintelligence.
在我之前的文章中，我们询问人工智能是否可以将自己一路提升到高得离谱的超级智能水平。
My conclusion: today’s feedback loop is probably too weak, because improving AI faces diminishing returns at every step.
我的结论是：今天的反馈循环可能太弱了，因为改进人工智能的每一步都面临着回报递减。
Even fully autonomous AIs doing all the research and coding don’t make those diminishing returns disappear.
即使完全自主的人工智能完成所有的研究和编码也不会让这些收益递减消失。
We should keep updating that conclusion as real-world parameters and data change.
随着现实世界参数和数据的变化，我们应该不断更新该结论。
But we didn’t look closely at another possibility: running millions or billions of AI minds in parallel.
但我们没有仔细研究另一种可能性：并行运行数百万或数十亿个人工智能思维。
How much do they help?
他们有多少帮助？
That’s what we’ll look at here.
这就是我们将在这里看到的内容。
Plenty of people have modeled how more AI agents working in parallel could fuel recursive self-improvement.
很多人已经模拟了更多并行工作的人工智能代理如何促进递归自我改进。
But the data we have today tell a different story.
但我们今天掌握的数据讲述了一个不同的故事。
More agents are useful for speed : they can solve many problems faster.
更多代理对于速度很有用：它们可以更快地解决许多问题。
They’re less useful for boosting AI intelligence .
它们对于提升人工智能智能的作用不大。
Multi-agent gains in solving harder problems often diminish incredibly quickly and may even saturate.
多智能体在解决更困难问题上的收益通常会以极快的速度减少，甚至可能饱和。
Whether multi-agent systems can produce the conceptual breakthroughs RSI might require to become self-sustaining is the harder question.
多智能体系统是否能够产生 RSI 可能需要的概念突破以实现自我维持，这是一个更难的问题。
Without breakthroughs in how well agents parallelize problems - something that may depend on the structure of machine learning problems themselves - it’s hard to see even millions or billions of copies of today’s models making repeated breakthroughs on the hardest problems.
如果代理并行处理问题的能力没有突破（这可能取决于机器学习问题本身的结构），就很难看到当今模型的数百万或数十亿副本在最困难的问题上不断取得突破。
These are among the strongest diminishing returns we’ve looked at.
这些是我们所见过的最强的收益递减之一。
Of course, these diminishing returns may change.
当然，这些收益递减的情况可能会改变。
And we have less experimental data to go off of today than we’d like.
今天我们能得到的实验数据比我们想要的要少。
We’ll need to recalibrate as new evidence comes in.
随着新证据的出现，我们需要重新校准。
I’ll end with the things we should track and my predictions about a natural experiment from newly released Anthropic data .
最后我将介绍我们应该跟踪的事情以及我根据新发布的人类数据对自然实验的预测。
Below is the two-image summary: The experimental observations of how large numbers of agents scale to solve harder problems (what I’m referring to as intelligence) are far far lower than the assumptions in studies that predict that AI recursive self-improvement will lead to an intelligence explosion.
下面是两张图片的摘要：关于如何扩展大量代理来解决更困难的问题（我所说的智能）的实验观察结果远远低于预测人工智能递归自我改进将导致智能爆炸的研究假设。
The first image is at the scale of 100 parallel AI agents.
第一张图片的规模为 100 个并行 AI 代理。
The second is at the scale of a billion AI parallel agents.
第二个是十亿个人工智能并行代理的规模。
Real world experiments show dramatically lower intelligence gains from parallel agents than hypothetical AI models.
现实世界的实验表明，与假设的人工智能模型相比，并行代理的智能增益要低得多。
Extrapolated out to a billion agents, the intelligence gain in hypothetical models is orders of magnitude beyond real world experiment gain curves.
推断到十亿个代理，假设模型中的智力增益比现实世界的实验增益曲线高出几个数量级。
To recap the last post : recursive self-improvement (RSI) means AI improvements feeding back into the next round of AI research.
回顾一下上一篇文章：递归自我改进（RSI）意味着人工智能的改进反馈到下一轮人工智能研究中。
Four findings stood out:
有四个突出的发现：
Real AI research is harder than benchmark scores suggest.
真正的人工智能研究比基准分数显示的更难。
OpenAI’s models needed human help on much shorter research tasks than the coding benchmarks led us to expect.
OpenAI 的模型需要人工帮助来完成比我们预期的编码基准短得多的研究任务。
More AI activity produces much smaller gains in useful work.
更多的人工智能活动在有用工作中产生的收益要小得多。
OpenAI used 124 times as many tokens per researcher but ran only 1.6 times as many experiments per researcher.
OpenAI 每个研究人员使用了 124 倍的代币，但每个研究人员只运行了 1.6 倍的实验。
Maintaining progress requires ever more resources.
保持进步需要更多的资源。
Training and thinking time face steep diminishing returns.
培训和思考时间面临着急剧的回报递减。
Anthropic reports that more powerful AI has been key to maintaining the current pace of AI progress .
Anthropic 报告称，更强大的人工智能是维持当前人工智能进步步伐的关键。
The measured software feedback loop doesn’t yet look self-sustaining.
测量到的软件反馈循环看起来还不能自我维持。
Our tentative calibration put it roughly five to ten times below the strength needed.
我们的初步校准使其比所需强度低大约五到十倍。
I want to reiterate a distinction here: fully automated AI self-improvement doesn’t at all guarantee an intelligence explosion.
我想在这里重申一个区别：完全自动化的人工智能自我完善并不能保证智能爆炸。
Diminishing returns in AI capabilities dampen that feedback loop.
人工智能能力的回报递减会抑制反馈循环。
Parallel agents can help automate a search across AI parameters, but don’t look nearly as useful in making new conceptual breakthroughs.
并行代理可以帮助自动搜索人工智能参数，但在实现新的概念突破方面看起来并没有那么有用。
My expectation is that we’ll need those sorts of breakthroughs for self-improvement to become self-sustaining, let alone to accelerate rapidly and reach truly profound levels of superintelligence.
我的期望是，我们需要这些突破来实现自我完善，实现自我维持，更不用说快速加速并达到真正深刻的超级智能水平。
Today , let’s examine in much more detail what it would mean to run thousands, millions, or billions of AI researchers in parallel.
今天，让我们更详细地研究一下并行运行数千、数百万或数十亿人工智能研究人员意味着什么。
In this post I’m focusing only on software: AI researchers improving the algorithms that make AI better or cheaper to run.
在这篇文章中，我只关注软件：人工智能研究人员改进算法，使人工智能运行得更好或更便宜。
Hardware, investment, and economic growth have their own feedback loops, which we discussed briefly in the last post .
硬件、投资和经济增长都有自己的反馈循环，我们在上一篇文章中对此进行了简要讨论。
Here, we’re asking how much more the AI minds themselves can discover.
在这里，我们要问的是人工智能思维本身还能发现多少东西。
Millions of AI minds, each more capable than a Nobel Prize winner across many fields, working at much higher speed.
数以百万计的人工智能头脑，每个人都比许多领域的诺贝尔奖获得者更有能力，并且工作速度要快得多。
That would be extraordinary.
那将是非凡的。
The intelligence of each mind is a crucial part of the premise.
每个头脑的智慧是前提的关键部分。
Tom Davidson and Daniel Eth’s Forethought paper describes the equivalent of “millions of top-notch human researchers.” Benjamin Todd’s explainer points to “millions of competent engineers working on AI research.” Dwarkesh Patel emphasizes AI’s “massive population size and perfect knowledge transfer.”
汤姆·戴维森 (Tom Davidson) 和丹尼尔·埃斯 (Daniel Eth) 的《远见》论文描述了相当于“数百万顶尖人类研究人员”的情况。本杰明·托德的解释者指出“数百万有能力的工程师致力于人工智能研究”。 Dwarkesh Patel 强调人工智能的“庞大的人口规模和完美的知识转移”。
How does the number of AI agents running in parallel enter these arguments?
并行运行的人工智能代理的数量如何输入这些参数？
In Eth and Davidson’s illustrative model, twice as many researchers produce twice as many papers.
在 Eth 和 Davidson 的说明性模型中，两倍的研究人员发表了两倍的论文。
Cheaper AI lets the same hardware run more researchers, while each software improvement gets harder to find.
更便宜的人工智能可以让相同的硬件运行更多的研究人员，而每项软件改进都变得更难找到。
Alan Chan and colleagues’ 2026 paper , including AI pioneers Geoffrey Hinton and Yoshua Bengio, assumes an even larger effect from adding researchers: doubling effective research labor increases the rate of software improvement by about 2.6 times , holding the software level fixed.
Alan Chan 及其同事（包括人工智能先驱 Geoffrey Hinton 和 Yoshua Bengio）在 2026 年发表的论文假设，增加研究人员会产生更大的影响：在保持软件水平固定的情况下，有效研究劳动力加倍，软件改进率提高约 2.6 倍。
Those are assumptions about what more researchers produce.
这些是关于更多研究人员产生的结果的假设。
The growing number of AI researchers must keep finding improvements fast enough to outweigh the increasing difficulty of further progress.
越来越多的人工智能研究人员必须不断地以足够快的速度寻找改进，以克服进一步进展所面临的日益增加的困难。
We’ll come back to those assumptions after looking at what larger AI teams actually achieve.
在研究了大型人工智能团队实际取得的成就后，我们将回到这些假设。
Meanwhile, we had OpenAI’s Navier–Stokes announcement : around 10,000 agents working concurrently during an 88-hour run solved a mathematics Millenium Problem.
与此同时，我们收到了 OpenAI 的纳维-斯托克斯公告：大约 10,000 个智能体在 88 小时的运行中同时工作，解决了数学千年难题。
It made the idea vivid.
它使这个想法变得生动。
A datacenter full of AI mathematicians, tackling a famously hard problem.
一个充满人工智能数学家的数据中心，正在解决一个著名的难题。
I understand the excitement.
我理解这种兴奋。
But the number of agents tells us how many AI researchers we ran.
但代理的数量告诉我们我们有多少人工智能研究人员。
We also need to measure what those additional researchers contributed.
我们还需要衡量这些额外研究人员的贡献。
Super-linear: Twice the input produces more than twice the output.
超线性：两倍的输入产生两倍以上的输出。
Accelerating returns.
加速回报。
Linear: Twice the input produces twice the output.
线性：两倍的输入产生两倍的输出。
Independent jobs can look like this.
独立工作可以是这样的。
A power-law exponent of 1 is the same case.
幂律指数 1 也是同样的情况。
Power-law diminishing returns: Output grows as input raised to a power below 1.
幂律收益递减：当输入提高到低于 1 的幂时，输出会增加。
At exponent 0.5, doubling output takes four times the input.
在指数 0.5 时，输出加倍需要输入的四倍。
At 0.2, it takes 32 times.
当为 0.2 时，需要 32 次。
At 0.1, it takes 1,024 times.
如果为 0.1，则需要 1,024 次。
Logarithmic: Each multiplication of input buys the same additional gain.
对数：输入的每次乘法都会购买相同的额外增益。
Another five points might require another tenfold increase in compute.
另外五个点可能需要将计算量再增加十倍。
Sub-logarithmic: Even successive tenfold increases buy smaller gains.
次对数：即使连续十倍的涨幅也会带来较小的收益。
Those gains can keep shrinking, or saturate at a ceiling.
这些收益可能会持续萎缩，或者达到上限。
A slow rise doesn’t require a hard plateau.
缓慢的上升并不需要一个艰难的平台期。
The classic pre-training scaling laws describe power-law reductions in prediction loss.
经典的预训练缩放定律描述了预测损失的幂律减少。
Many test-time score curves look roughly logarithmic over the measured range.
许多测试时间分数曲线在测量范围内看起来大致呈对数。
Some multi-agent results flatten much more sharply.
一些多智能体结果的平坦度要大得多。
Let’s illustrate this with numbers.
让我们用数字来说明这一点。
In the figure below, every curve earns the same gain from the first doubling of input.
在下图中，每条曲线从输入的第一次加倍中获得相同的增益。
The later differences then become visible.
后来的差异就变得明显了。
The hardest curves share a vertical scale, and each panel states its endpoint.
最难的曲线共享一个垂直比例，每个面板都标明其端点。
Different diminishing returns scaling patterns lead to very different long term results.
不同的收益递减缩放模式会导致截然不同的长期结果。
For runaway RSI, improvements have to make subsequent improvements arrive faster and faster.
对于失控的 RSI，改进必须使后续的改进到来得越来越快。
Logarithmic returns to the number of AI researchers are too weak to sustain that loop if every further doubling of software efficiency requires a fixed percentage more research.
如果软件效率每进一步翻一番都需要固定百分比的研究，那么人工智能研究人员数量的对数回报就太弱而无法维持这种循环。
Sub-logarithmic scaling or a plateau are weaker still.
次对数标度或平台期更弱。
The experiments I’ll summarize in the rest of this post aren’t perfect, and I wish we had more data.
我将在这篇文章的其余部分总结的实验并不完美，我希望我们有更多的数据。
But the clearest larger-team evidence we have today points to logarithmic or sub-logarithmic intelligence gains.
但我们今天拥有的最清晰的大团队证据表明，智力的增长是对数或亚对数的。
Those look too weak to make RSI self-sustaining, let alone runaway, under the assumptions we’re examining.
根据我们正在研究的假设，这些看起来太弱了，无法使 RSI 自我维持，更不用说失控了。
There are at least three reasons a larger team might be limited in its ability to produce better answers, and probably more.
大型团队提供更好答案的能力可能会受到限制，原因至少有三个，甚至可能更多。
Let’s start with these:
让我们从这些开始：
First, coordination skills.
第一，协调能力。
Can the agents divide the problem, manage the work, check each other, and combine their results?
代理人能否分工、管理工作、互相检查、综合结果？
Can they communicate what matters without spending all their time communicating?
他们能在不花费所有时间沟通的情况下传达重要的事情吗？
A team needs good project management as well as good individual contributors.
团队需要良好的项目管理以及良好的个人贡献者。
Second, homogeneity of thought.
二是思想同质化。
Are the agents contributing useful new ideas, or repeating the same approaches and mistakes?
代理人是否贡献了有用的新想法，或者重复了相同的方法和错误？
A hundred copies can share the same blind spots.
一百份副本可以共享相同的盲点。
More varied answers help only if some of that variety is right.
只有其中一些答案是正确的，更多样的答案才会有帮助。
Third, the structure of the problem.
第三，问题的结构。
How much can happen in parallel, and how much has to wait for earlier results?
有多少事情可以并行发生，有多少事情需要等待早期结果？
A million independent searches can run at once.
可以同时运行一百万个独立搜索。
A sequential investigation can’t.
顺序调查不能。
Even figuring out how to divide a conceptual problem may require an insight the team hasn’t found.
即使弄清楚如何划分概念性问题也可能需要团队尚未找到的洞察力。
These limits interact.
这些限制相互作用。
A better model can supply an idea its copies previously lacked, find a better division of labor, or recognize a solution the team had been overlooking.
更好的模型可以提供其副本以前缺乏的想法，找到更好的分工，或者识别团队一直忽视的解决方案。
A difficult task may become easier to parallelize once somebody discovers the right approach.
一旦有人发现了正确的方法，一项困难的任务可能会变得更容易并行化。
That’s why the shape of the curve matters, and why it can change.
这就是曲线的形状很重要以及它可以改变的原因。
Before we go further, let’s separate two things: speed , or how quickly agents solve a problem, and collective intelligence , or whether they can solve a harder problem and reach a new level of skill.
在我们进一步讨论之前，让我们先区分两件事：速度，即智能体解决问题的速度；以及集体智慧，即他们是否能够解决更难的问题并达到新的技能水平。
We’ll use verified benchmark scores as evidence of that intelligence.
我们将使用经过验证的基准分数作为该情报的证据。
They aren’t a universal IQ scale.
它们不是通用的智商量表。
So far, adding agents seems much more effective at speeding up execution than at raising the level of problem a team can solve.
到目前为止，增加代理在加快执行速度方面似乎比提高团队可以解决的问题水平更有效。
Let’s start with the best case.
让我们从最好的情况开始。
A million independent jobs.
一百万个独立的工作岗位。
A million candidates to test.
有一百万名候选人接受测试。
A million places to search.
一百万个地方可供搜索。
Each agent gets useful work it can do without waiting for the others, and a reliable way to check whether it succeeded.
每个智能体都可以获得有用的工作，而无需等待其他智能体，并且有一种可靠的方法来检查它是否成功。
That can scale extremely well.
这可以很好地扩展。
Ten workers may finish ten jobs in the time one worker finishes one.
一名工人完成一项工作时，十名工人可以完成十项工作。
A huge search may find a rare solution that a smaller search misses entirely.
大规模搜索可能会找到一个罕见的解决方案，而较小的搜索则完全找不到。
One concerning use of parallel agents is offensive cybersecurity.
与并行代理的使用有关的一个问题是进攻性网络安全。
Lisan al-Ghaib emphasizes this possibility.
Lisan al-Ghaib 强调了这种可能性。
I disagree with his broader argument about swarm intelligence, but many attacks running in parallel could be extremely powerful.
我不同意他关于群体智能的更广泛的论点，但许多并行运行的攻击可能非常强大。
A swarm may not need a conceptual breakthrough to search many targets or try many possible vulnerabilities at once.
群体可能不需要概念上的突破来搜索许多目标或一次尝试许多可能的漏洞。
That said, penetrating the most secure targets, such as operating systems and browsers, can require long attack chains.
也就是说，渗透最安全的目标（例如操作系统和浏览器）可能需要很长的攻击链。
Each step depends on earlier discoveries to get through another layer of defenses.
每一步都依赖于早期的发现来突破另一层防御。
Here again, we might find diminishing returns or eventual saturation, even with respect to speed.
在这里，我们可能会再次发现收益递减或最终饱和，即使就速度而言也是如此。
Indeed, one of the experiments below suggests that.
事实上，下面的一项实验表明了这一点。
Toby Ord analyzed how well GPT-5.6 Sol increases the speed of solving problems on different benchmarks, and at what cost.
Toby Ord 分析了 GPT-5.6 Sol 在不同基准上提高解决问题速度的效果以及成本。
This gives us a useful way to measure the tradeoff.
这为我们提供了一种衡量权衡的有用方法。
In general, using more agents can help get tasks done more quickly, but at higher total cost than if you’d used fewer agents.
一般来说，使用更多的代理可以帮助更快地完成任务，但总成本比使用更少的代理更高。
On top of that, the tradeoff varies by the type of task.
最重要的是，权衡因任务类型而异。
On command-line tasks (Terminal-Bench), Ord’s analysis finds that four agents can reach the same result about 1.9 times faster , at about 2.1 times the token cost.
在命令行任务（Terminal-Bench）上，Ord 的分析发现，四个代理达到相同结果的速度大约是 1.9 倍，而代币成本大约是 2.1 倍。
On web research (BrowseComp), four agents can reach the same result about 2.6 times faster , at about 1.6 times the cost.
在网络研究 (BrowseComp) 上，四个代理可以以大约 2.6 倍的速度达到相同的结果，而成本大约为 1.6 倍。
I’m using the figure below to show how different tasks differ in that speed vs cost tradeoff.
我使用下图来展示不同任务在速度与成本权衡方面的差异。
More speed costs more, and the tradeoff varies by task.
更快的速度成本更高，并且权衡因任务而异。
Based on Toby Ord’s analysis.
基于托比·奥德的分析。
Source.
来源。
Why the difference?
为什么有区别？
Web research can often split into independent searches.
网络研究通常可以分为独立的搜索。
A command-line task may have to change a shared environment in sequence.
命令行任务可能必须按顺序更改共享环境。
Some problems are simply easier to divide among agents than others.
有些问题比其他问题更容易在代理之间划分。
Ord uses lambda, λ , for how speed scales with the number of agents.
Ord 使用 lambda 来表示速度如何随代理数量变化。
We’ll come back to it.
我们稍后会回来讨论这个问题。
At λ = 0.5, speed and total compute cost rise by the same factor.
当 λ = 0.5 时，速度和总计算成本增加相同的系数。
A higher λ means speed rises faster than cost.
较高的 λ 意味着速度的上升速度快于成本的上升速度。
Inspired by Toby Ord’s analysis, I applied the same approach to Anthropic’s published results on program reconstruction, or ProgramBench.
受到托比·奥德 (Toby Ord) 分析的启发，我对 Anthropic 发布的程序重建（或 ProgramBench）结果应用了相同的方法。
Here again, more speed comes with more cost.
同样，更快的速度意味着更高的成本。
But we also see that models within the Claude family are getting better at the speed-cost tradeoff.
但我们也看到 Claude 系列中的模型在速度与成本权衡方面做得越来越好。
The newest model, Opus 5.5, gets more speedup for a smaller cost increase than Opus 5 or Fable 5.1.
最新型号 Opus 5.5 与 Opus 5 或 Fable 5.1 相比，速度更快，成本增加更小。
With five agents, Opus 5.5 gets about 2.7 times its solo speed, compared with 2.2 times for Opus 5, and does so at lower cost.
通过五个代理，Opus 5.5 的单独速度提高了约 2.7 倍，而 Opus 5 为 2.2 倍，而且成本更低。
[Note: We’re using total agent runtime as a proxy for cost.
[注意：我们使用总代理运行时间作为成本代理。
Token usage and dollar costs could differ.]
代币使用和美元成本可能有所不同。]
Opus 5.5 gets more speedup for a smaller cost increase than older AI models.
与旧版 AI 模型相比，Opus 5.5 的加速速度更快，成本增加更小。
Source.
来源。
Will these speed-versus-cost gains hold up at hundreds, thousands, or millions of agents?
这些速度与成本的收益对数百、数千或数百万代理来说是否有效？
Probably not completely, and certainly not for all tasks.
可能不完全，当然也不适合所有任务。
Some problems are easy to split ten ways, harder to split a hundred ways, and harder still to split among ten thousand agents.
有些问题很容易以十种方式进行拆分，以一百种方式进行拆分则较困难，而在一万个代理中进行拆分则更难。
More of the work may have to wait for earlier results, and coordinating all those pieces takes work too.
更多的工作可能需要等待更早的结果，并且协调所有这些工作也需要工作。
Smarter models can get better at this.
更聪明的模型可以在这方面做得更好。
But my best guess is that, on many problems, each further expansion will buy less speed for the added cost.
但我最好的猜测是，在许多问题上，每一次进一步的扩展都会以增加的成本换取更少的速度。
On the Lean formal math task, we can compare how quickly teams reach the same score.
在精益正式数学任务中，我们可以比较团队达到相同分数的速度。
Using Anthropic’s published checkpoints , we estimate the speed and cost of reaching 0.50.
使用 Anthropic 发布的检查点，我们估计达到 0.50 的速度和成本。
The chart starts with ten agents as the baseline.
该图表以十个代理作为基线开始。
On a math proof benchmark, the speed advantage of adding additional agents starts to taper with scale.
在数学证明基准上，添加额外代理的速度优势开始随着规模的扩大而逐渐缩小。
Source.
来源。
From ten to thirty agents, speed improves faster than cost.
从十个代理到三十个代理，速度的提高速度快于成本的提高。
From thirty to a hundred, cost rises faster than speed.
从三十到一百，成本的增长速度快于速度的增长。
The local speed exponent, λ, falls from about 0.75 to 0.33 .
局部速度指数 λ 从大约 0.75 降至 0.33。
The extra agents still save time, but the tradeoff gets worse.
额外的代理仍然可以节省时间，但权衡会变得更糟。
This suggests that even speed gains can taper as we split the work among more agents, let alone gains in intelligence.
这表明，当我们将工作分配给更多代理时，即使速度的提升也会逐渐减弱，更不用说智力的提升了。
More agents may add less and less as we approach the limits of how easily a problem divides into subproblems.
当我们接近问题划分为子问题的容易程度的极限时，更多的代理可能会添加越来越少的内容。
Better coordination may push those limits outward; the task itself may still impose them.
更好的协调可能会突破这些限制；任务本身仍然可能强加它们。
So in speed , more agents can buy us time at some additional cost.
因此，在速度方面，更多的代理商可以以一些额外费用为我们赢得时间。
Newer models are improving, but there are probably limits.
新型号正在改进，但可能存在局限性。
What about intelligence?
那么智力呢？
Most arguments for superintelligence go beyond running huge numbers of today’s agents.
大多数支持超级智能的论点都超出了当今运行大量智能体的范围。
They posit individual models becoming vastly smarter.
他们认为个体模型会变得更加聪明。
There’s a chicken-and-egg problem here.
这里有一个先有鸡还是先有蛋的问题。
Getting there may require conceptual breakthroughs in machine learning itself, perhaps as profound as the ideas that made modern machine learning possible, or at least as important as the Transformer and reasoning models.
要实现这一目标可能需要机器学习本身的概念突破，也许与使现代机器学习成为可能的思想一样深刻，或者至少与 Transformer 和推理模型一样重要。
Parallel search may not be enough to find them.
并行搜索可能不足以找到它们。
Anthropic has identified ambitious research ideas as an area where current models lag behind humans .
Anthropic 已将雄心勃勃的研究想法确定为当前模型落后于人类的领域。
In the previous post , I argued that today’s measured gains from smarter models are roughly five to ten times too weak to produce a self-sustaining feedback loop.
在上一篇文章中，我认为当今从更智能的模型中测得的收益大约要弱五到十倍，无法产生自我维持的反馈循环。
But what about running enormous numbers of our best models in parallel?
但是并行运行大量最好的模型又如何呢？
Can a massive team of frontier agents that aren’t yet superintelligent invent a new architecture, a new learning method, or the idea needed to build a much smarter successor?
一个由尚未超级智能的前沿代理组成的庞大团队能否发明一种新的架构、一种新的学习方法，或者构建一个更聪明的继任者所需的想法？
That’s the heart of the question.
这就是问题的核心。
For runaway RSI, those discoveries have to overcome the diminishing returns we discussed in the last post.
对于失控的 RSI，这些发现必须克服我们在上一篇文章中讨论的收益递减问题。
Here are three places they appear.
这是它们出现的三个地方。
The pre-training example comes from the Chinchilla scaling fit .
预训练示例来自 Chinchillascalingfit 。
The reinforcement-learning example comes from ScaleRL .
强化学习示例来自 ScaleRL 。
The inference row describes the test-time math curves we discussed previously.
推理行描述了我们之前讨论的测试时间数学曲线。
The goal is better returns to data and compute, and gains that continue longer before flattening.
目标是更好地回报数据和计算，并在趋于平缓之前持续更长时间。
My expectation is that we’ll get easy AI improvements through work we can automate, parallelize, and search.
我的期望是，我们将通过自动化、并行化和搜索的工作轻松实现人工智能改进。
But to get very large benefits, and runaway RSI, I expect we’ll need conceptual breakthroughs.
但为了获得巨大的收益和失控的 RSI，我预计我们需要概念上的突破。
That’s why the intelligence of each model matters so much.
这就是为什么每个模型的智能如此重要。
Bear in mind that we’ve observed these sorts of diminishing returns in machine learning for more than thirty years.
请记住，三十多年来，我们已经观察到机器学习领域的这种收益递减现象。
Power-law learning curves were measured in 1993 .
幂律学习曲线于 1993 年测量。
They appear to be a fundamental feature of machine learning as we know it.
正如我们所知，它们似乎是机器学习的基本特征。
We shouldn’t assume they’ll be trivially changed, or that we’ll suddenly find a way to improve AI without diminishing returns.
我们不应该假设它们会发生微不足道的改变，或者我们会突然找到一种在不减少回报的情况下改进人工智能的方法。
That would be a revolution.
那将是一场革命。
Discovering such a route might itself require far smarter minds, and we don’t know whether the mathematics of learning would permit it across the problems we care about.
发现这样一条路线本身可能需要更聪明的头脑，而且我们不知道学习的数学是否允许它解决我们关心的问题。
That doesn’t mean there haven’t been big improvements.
这并不意味着没有大的改进。
The transformer and reasoning models have moved the field forward enormously.
变压器和推理模型极大地推动了该领域的发展。
But, as we saw in the previous post , a runaway feedback loop needs enough acceleration to keep outweighing the obstacles, again and again.
但是，正如我们在上一篇文章中所看到的，失控的反馈循环需要足够的加速度才能一次又一次地超越障碍。
As ideas get harder to find, we may need a faster research pace just to maintain today’s rate of progress.
随着想法变得越来越难找到，我们可能需要更快的研究速度才能保持今天的进展速度。
Anthropic’s assessment suggests that better AI tools are becoming necessary just to maintain the current pace of AI progress .
Anthropic 的评估表明，为了维持当前的人工智能进步步伐，更好的人工智能工具变得越来越必要。
OpenAI’s GPT-5.6 Sol results provide a useful example on cybersecurity.
OpenAI 的 GPT-5.6 Sol 结果提供了网络安全方面的有用示例。
SEC-Bench Pro tests the generation of proof-of-concept demonstrations of software vulnerabilities.
SEC-Bench Pro 测试软件漏洞概念验证演示的生成。
This is a particular security task, not a measure of coding ability in general.
这是一项特定的安全任务，而不是一般编码能力的衡量标准。
The reported peak scores are approximately 71.2 with one agent, 74.3 with four, and 76.4 with sixteen.
报告的最高分数约为 1 名特工的 71.2、4 名特工的 74.3 和 16 名特工的 76.4。
On this cybersecurity test, sixteen times as many agents buy about 5.2 score points.
在这次网络安全测试中，购买大约 5.2 分的代理商数量是其 16 倍。
The first fourfold increase adds 3.1 points; the next adds only 2.1.
前四倍增加 3.1 点；接下来仅添加 2.1。
There’s still plenty of room below a perfect score, but the gain on this test is shrinking, even with multiplication of the number of agents.
距离完美分数还有很大的空间，但即使代理人数量成倍增加，这项测试的收益也在缩小。
This is sub-logarithmic scaling.
这是次对数缩放。
On SEC-Bench Pro, each expansion buys fewer score points, showing sub-log scaling.
在 SEC-Bench Pro 上，每次扩展购买的分数更少，显示子日志缩放。
Source.
来源。
These agents could coordinate through OpenAI’s ultra setup.
这些代理可以通过 OpenAI 的超级设置进行协调。
This is already a test of collaboration, not just a stack of independent guesses.
这已经是对协作的考验，而不仅仅是独立猜测的堆叠。
The points are the best reported results across different budgets, rather than an equal-compute comparison.
这些点是不同预算的最佳报告结果，而不是平等计算的比较。
When OpenAI announced its Navier–Stokes result , the 10,000 agents attracted enormous excitement.
当 OpenAI 宣布其 Navier-Stokes 结果时，10,000 名智能体引起了巨大的兴奋。
Interviewing Noam Brown, Dwarkesh Patel said he was “shocked by the scale of cognitive effort” concentrated into such a short run.
德瓦克什·帕特尔在采访诺姆·布朗时表示，他对集中在如此短期内的认知努力的规模感到震惊。
But how much did those extra agents contribute?
但这些额外的特工贡献了多少呢？
Brown, an OpenAI researcher who helped develop o1 and now works on multi-agent systems, was emphatic:
Brown 是一名 OpenAI 研究员，曾帮助开发了 o1，现在致力于多智能体系统，他强调：
He credits the smarter underlying model.
他相信更智能的基础模型。
OpenAI hadn’t run the single-agent experiment needed to isolate the swarm’s contribution.
OpenAI 没有运行隔离群体贡献所需的单代理实验。
Brown’s estimate is a judgment, but it directly challenges the idea that the sheer number of agents was the decisive ingredient.
布朗的估计是一种判断，但它直接挑战了“代理人的绝对数量是决定性因素”的观点。
Formal mathematical work seems especially analogous to the conceptual breakthroughs AI research may need.
正式的数学工作似乎与人工智能研究可能需要的概念突破特别相似。
It demands a chain of valid reasoning, sometimes with a new insight that makes the whole problem tractable.
它需要一系列有效的推理，有时还需要新的见解，使整个问题变得易于处理。
Formalizing existing mathematics isn’t the same as inventing a new learning algorithm, but it’s a useful place to look.
形式化现有数学与发明新的学习算法不同，但这是一个有用的地方。
And here we find almost no extra gain when Opus 5.5 grows from ten parallel agents to a hundred.
在这里，当 Opus 5.5 从 10 个并行代理增长到 100 个时，我们几乎没有发现额外的增益。
Anthropic’s system card tests teams on 49 weighted theorem statements, checked in Lean.
Anthropic 的系统卡对团队进行了 49 个加权定理陈述的测试，并在 Lean 中进行了检查。
Each team gets 24 hours.
每个团队都有 24 小时的时间。
The first tenfold expansion adds 27 score points; the next adds just two.
第一次十倍扩展增加 27 分；接下来只添加两个。
On these math proofs, most of the gain arrives by ten agents.
在这些数学证明中，大部分收益来自十个代理。
This is steeply sub-logarithmic scaling, and possibly saturation Source.
这是陡峭的亚对数缩放，并且可能是饱和源。
These agents could communicate, share a machine and Git repository, and organize their work.
这些代理可以进行通信、共享机器和 Git 存储库并组织他们的工作。
The hundred-agent team even developed a hierarchy with twelve sub-leads.
一百名座席团队甚至制定了包含十二名副主管的等级制度。
It could also use ten times as much aggregate agent runtime as the ten-agent team over the same day.
它还可以在同一天使用十倍于十个代理团队的总代理运行时间。
That isn’t a measured tenfold increase in tokens, but it gives the larger team much more room to work.
这并不是代币的十倍增长，但它为更大的团队提供了更多的工作空间。
We still didn’t get over the hump.
我们仍然没有克服困难。
This won’t necessarily stay true, but the diminishing returns are incredibly steep here, possibly pointing toward a plateau.
这不一定是真的，但这里的收益递减非常陡峭，可能会趋于平稳。
Opus 5.5 starts smarter than the older models and gets more speedup from parallel agents on ProgramBench.
Opus 5.5 的启动比旧型号更智能，并且通过 ProgramBench 上的并行代理获得更多加速。
It clearly can put additional agents to useful work.
显然，它可以让更多的代理从事有用的工作。
So why do intelligence gains nearly stop on these math proofs?
那么，为什么智力的增长几乎停止在这些数学证明上呢？
Two possibilities stand out:
有两种可能性很突出：
The remaining work may be hard to parallelize.
剩下的工作可能很难并行化。
Later steps depend on understanding earlier results.
后续步骤取决于对早期结果的理解。
The remaining problems may need insights the model can’t find, however many copies we run.
剩下的问题可能需要模型无法找到的见解，无论我们运行多少个副本。
Intelligence may matter more than agent population or speed.
情报可能比特工数量或速度更重要。
We haven’t isolated the cause.
我们还没有找出原因。
But a thousand copies don’t automatically supply the insight one copy is missing.
但一千本并不能自动提供一本所缺少的洞察力。
Looking at the same math results, we see something striking.
看看相同的数学结果，我们看到了一些惊人的东西。
Ten Opus 5.5 agents get a higher score than a hundred Fable 5.1 or Opus 5.0 agents.
十个 Opus 5.5 代理比一百个 Fable 5.1 或 Opus 5.0 代理获得更高的分数。
Opus 5.5 starts much higher than the older models.
Opus 5.5 的起步比旧型号要高得多。
It improves substantially with ten agents, then barely improves with a hundred.
使用十个代理时它会显着改善，然后使用一百个代理时几乎没有改善。
But at 10 it’s already better than Fable or Opus 5’s results with a hundred agents.
但在 10 个代理的情况下，它已经比《神鬼寓言》或《作品 5》的结果更好了。
Ten Opus 5.5 agents beat a hundred Opus 5 agents on these math proofs.
在这些数学证明中，十个 Opus 5.5 智能体击败了一百个 Opus 5 智能体。
They beat a hundred Fable 5.1 agents too.
他们还击败了一百名《神鬼寓言 5.1》特工。
The stronger model does more with one-tenth as many agents.
更强大的模型只需十分之一的代理即可完成更多任务。
Ten Opus 5.5 agents are able to solve problems a hundred older-model agents can’t.
十个 Opus 5.5 代理能够解决一百个旧型号代理无法解决的问题。
Source.
来源。
We saw an analogous pattern with test-time compute in the previous post .
我们在上一篇文章中看到了测试时计算的类似模式。
OpenAI’s internal model starts far above Astra on the curated math problems below.
在下面精选的数学问题上，OpenAI 的内部模型的起点远远高于 Astra。
Even its lowest displayed result beats Astra’s highest.
即使它的最低显示结果也超过了 Astra 的最高结果。
Changing the model gives a much bigger boost than spending more compute on Astra over the tested range.
在测试范围内，与在 Astra 上花费更多计算量相比，更改模型可以带来更大的提升。
A stronger model makes the bigger difference than more test-time compute in solving previously-unsolved math problems.
在解决以前未解决的数学问题方面，更强大的模型比更多的测试时间计算产生更大的差异。
Source.
来源。
A rough extrapolation of Astra’s line suggests it would need about 300 times as much compute as the internal model to match the internal model’s lowest plotted score.
对 Astra 线的粗略推断表明，它需要内部模型大约 300 倍的计算量才能匹配内部模型的最低绘制分数。
And that extrapolation itself is an unknown.
这种推断本身是一个未知数。
It’s quite possible that Astra would just plateau before ever matching the OpenAI internal model’s results with far less compute.
在用少得多的计算量匹配 OpenAI 内部模型的结果之前，Astra 很可能会处于稳定状态。
Does adding different minds help?
添加不同的想法有帮助吗？
A study of model and persona diversity found substantial gains as diverse teams grew to eight agents.
一项对模型和人物角色多样性的研究发现，随着多元化团队发展到八名代理人，收益显着。
After that, the average across its tasks barely moved.
此后，其任务的平均值几乎没有变化。
Diversity of AI models helps, but the gains flatten after eight to twelve agents.
人工智能模型的多样性有所帮助，但在八到十二个智能体之后，收益就会趋于平缓。
Source.
来源。
Other approaches, including X-MAS and Aitchison and colleagues’ forecasting study , find useful gains from combining different models.
其他方法，包括 X-MAS 和 Aitchison 及其同事的预测研究，通过结合不同的模型发现了有用的收益。
But none of these results demonstrates sustained improvement over many expansions of the team.
但这些结果都没有表明团队的多次扩张带来了持续的改进。
The first study directly shows flattening; the others test smaller ranges.
第一项研究直接显示了扁平化；其他人则测试较小的范围。
It’s plausible that today’s models simply aren’t diverse enough.
当今的模型可能不够多样化。
LLMs often produce strikingly similar ideas .
法学硕士经常产生惊人相似的想法。
Perhaps we need bigger changes in how these minds are trained.
也许我们需要对这些思维的训练方式进行更大的改变。
We’ll return to whether more deliberately different agents could change the curves.
我们将回到更刻意的不同代理人是否可以改变曲线的问题。
All of the above examples show sub-logarithmic scaling , roughly the hardest kind.
上面所有的例子都显示了次对数缩放，大致是最难的一种。
That said, some experiments show slightly easier diminishing returns, indicating roughly the same gain from each multiplication of parallel work.
也就是说，一些实验显示收益递减稍微容易一些，这表明并行工作的每次乘法获得的收益大致相同。
That’s logarithmic scaling .
这就是对数缩放。
It’s better, but still steep diminishing returns.
它更好，但回报仍然急剧递减。
If that’s how useful research grows while each further AI improvement gets harder to find, even billions of agents don’t make a self-sustaining RSI loop look plausible.
如果这就是有用的研究不断增长的方式，而每次进一步的人工智能改进都变得越来越难找到，那么即使有数十亿个智能体也无法让自我维持的 RSI 循环看起来合理。
In Ryan Greenblatt’s ARC experiment , a model generated and checked many candidate programs.
在 Ryan Greenblatt 的 ARC 实验中，模型生成并检查了许多候选程序。
More samples kept improving its score, by roughly the same amount per doubling.
更多的样本不断提高其得分，每翻一番，提高的幅度大致相同。
This is parallel search rather than a communicating research team, but it gives us a useful comparison.
这是并行搜索而不是交流研究团队，但它为我们提供了有用的比较。
A molecular-exploration study found a similar logarithmic pattern as it grew from one to 128 agents.
一项分子探索研究发现，随着药物从 1 种增加到 128 种，也出现了类似的对数模式。
Multiple agents explored chemical targets more effectively than one agent given the same total evaluation budget.
在总评估预算相同的情况下，多个代理比单个代理更有效地探索化学目标。
Each agent learned as it searched.
每个智能体在搜索过程中都在学习。
More AI agents explore more chemical targets, with logarithmic diminishing returns.
更多的人工智能代理探索更多的化学目标，回报呈对数递减。
Source.
来源。
There are encouraging results from communicating teams too, including Argus on web research and Microsoft and Berkeley’s work on ARC-AGI-3 .
沟通团队也取得了令人鼓舞的成果，包括 Argus 在网络研究方面的工作以及微软和伯克利在 ARC-AGI-3 方面的工作。
The latter finds unusually strong gains from one to five agents.
后者发现一到五名代理人获得了异常强劲的收益。
Those results deserve attention; we still need to see what happens with much larger teams.
这些结果值得关注；我们仍然需要看看更大的团队会发生什么。
Logarithmic gains are better than the sub-logarithmic curves we saw earlier.
对数增益比我们之前看到的次对数曲线要好。
Each tenfold expansion keeps buying the same extra increment, instead of less each time.
每次十倍扩张都会持续购买相同的额外增量，而不是每次都减少。
But both shapes rise drastically more slowly than the research gains assumed in hypothetical models of self-sustaining or runaway RSI.
但这两种形状的上升速度都比自我维持或失控 RSI 假设模型中假设的研究成果要慢得多。
Let’s put them side by side.
让我们把它们并排放置。
Some work can be split into independent searches or experiments.
有些工作可以分为独立的搜索或实验。
Other work needs the previous result before it can proceed.
其他工作需要之前的结果才能继续进行。
And some problems may need a higher level of intelligence or a conceptual breakthrough before we even know how to divide them.
有些问题可能需要更高水平的智力或概念上的突破，我们才能知道如何划分它们。
Here’s a very tentative and relatively subjective stab at which problems will scale best and worse with multiple agents.
这是一个非常试探性且相对主观的尝试，通过多个代理，问题会变得最好和更糟。
The chart runs from work that seems easiest to parallelize at the top to the hardest at the bottom.
该图表从顶部似乎最容易并行化的工作到底部最难并行化的工作。
I’m sure others have many thoughts here, and that this can be much improved upon.
我相信其他人对此有很多想法，并且这可以得到很大的改进。
Will a million or a billion parallel AIs get us to self-sustaining or runaway RSI?
一百万或十亿个并行人工智能会让我们实现自我维持还是失控的 RSI？
Let’s do the math.
让我们算一下。
Hypothetical software-RSI models assume that a growing number of AI researchers produces a much larger stream of discoveries.
假设的软件 RSI 模型假设越来越多的人工智能研究人员产生了更多的发现。
Those discoveries make AI cheaper or more productive, allowing the next round of research to grow again.
这些发现使人工智能变得更便宜或更高效，从而使下一轮研究再次增长。
There’s a strong case for gains in speed and for searching more of a well-defined space.
有充分理由提高速度并搜索更多明确定义的空间。
Even speed gains, as we’ve seen, can taper.
正如我们所看到的，即使速度的增益也会逐渐减弱。
Intelligence is much harder to boost through parallel agents.
通过并行智能体来提升智力要困难得多。
In several experiments, peak problem-solving performance (what I’m referring to here as intelligence) rises dramatically more slowly than speed.
在几个实验中，解决问题的峰值表现（我在这里指的是智力）的提升速度明显慢于速度。
Chan and colleagues (2026) : doubling effective research labor produces about 2.6 times the rate of software improvement , holding the software level fixed.
Chan 及其同事（2026）：在软件水平固定的情况下，有效研究劳动加倍可以使软件改进率提高约 2.6 倍。
I find that implausible as a general assumption about adding copies of today’s AI models.
我发现作为添加当今人工智能模型副本的一般假设是难以置信的。
The clearest large-team intelligence experiments look nothing like that.
最清晰的大型团队情报实验看起来根本不是这样的。
Eth and Davidson (2025) : their illustrative team produces twice as many papers from twice as many researchers .
Eth 和 Davidson (2025)：他们的说明性团队从两倍的研究人员那里产出了两倍的论文。
The accelerating example assumes that the number of researchers grows faster than the work required for each further improvement.
加速示例假设研究人员数量的增长速度快于每次进一步改进所需的工作量。
Tom Davidson and Tom Houlden (2025) : their central exponent for parallel research effort is 0.6 .
Tom Davidson 和 Tom Houlden (2025)：他们的并行研究工作的中心指数是 0.6 。
Ten times the labor gives about four times the research pace.
十倍的劳动力带来了大约四倍的研究速度。
They cut that exponent in half, to 0.3, when experimental compute is held fixed and becomes a bottleneck.
当实验计算保持固定并成为瓶颈时，他们将该指数减半至 0.3。
The benchmark reports above don’t appear to have been limited by experimental compute, so we’re using 0.6 here.
上面的基准测试报告似乎不受实验计算的限制，因此我们在这里使用 0.6。
The top half of the next chart shows research-output relationships used in hypothetical RSI models where AI self-improvement can sustain itself or accelerate.
下一张图表的上半部分显示了假设的 RSI 模型中使用的研究与产出关系，其中人工智能的自我改进可以维持或加速。
The bottom half shows real-world intelligence shapes from the benchmark experiments.
下半部分显示了基准实验中的真实世界智能形状。
Those experiments show dramatically less gain from running parallel agents than the hypothetical RSI models assume.
这些实验表明，运行并行代理所带来的收益比假设的 RSI 模型所假设的要少得多。
Let’s put these onto a single chart.
让我们把它们放在一张图表上。
We give every curve the same gain from its first expansion from one to ten agents or researchers, so we can compare their shapes.
我们为每条曲线提供相同的增益，从第一次扩展到十个代理或研究人员，这样我们就可以比较它们的形状。
Now let’s zoom in on the gap.
现在让我们放大差距。
The real-world shapes sit in a very different zone from the gains assumed in those RSI models, even when we let the math-proof curve keep rising slowly.
即使我们让数学证明曲线继续缓慢上升，现实世界的形状与这些 RSI 模型中假设的收益处于非常不同的区域。
Zooming in makes the gap clearer.
放大可以让差距更加清晰。
Hypothetical RSI models show much faster gains from adding more agents than empirical models based on real-world experiments.
与基于现实世界实验的经验模型相比，假设的 RSI 模型显示出通过添加更多代理获得的收益要快得多。
Today’s hypothetical RSI models assume gains from additional AI researchers that we haven’t seen in these real-world experiments.
今天假设的 RSI 模型假设来自其他人工智能研究人员的收益，而我们在这些现实世界的实验中还没有看到过。
Removing humans from the loop doesn’t remove shared blind spots or sequential dependencies.
将人类从循环中移除并不能消除共享盲点或顺序依赖性。
Adding more copies of today’s models hasn’t shown the sustained intelligence gains that may be needed for repeated conceptual breakthroughs.
添加更多当今模型的副本并没有显示出重复概念突破可能需要的持续情报增益。
I do expect better AI models to get better at spreading work across agents: dividing the problem, managing collaborators, and combining what they find.
我确实希望更好的人工智能模型能够更好地在代理之间分配工作：划分问题、管理协作者以及组合他们发现的内容。
We’ve already seen progress in speed on program reconstruction.
我们已经看到了程序重建速度的进步。
We haven’t seen a clear trend toward better multi-agent intelligence scaling in the math results.
在数学结果中，我们还没有看到更好的多智能体智能扩展的明显趋势。
More on that in a moment.
稍后会详细介绍。
First, an encouraging analogy: single-agent thinking-time scaling .
首先，一个令人鼓舞的类比：单智能体思考时间缩放。
Toby Ord’s analysis suggests that humans get more relative benefit from additional time than these AIs get from additional reasoning tokens.
托比·奥德（Toby Ord）的分析表明，人类从额外的时间中获得的相对收益比这些人工智能从额外的推理标记中获得的收益更多。
That may surprise people.
这可能会让人们感到惊讶。
But newer models are getting closer to the human reference on these cybersecurity tasks.
但较新的模型在这些网络安全任务上越来越接近人类参考。
AI agents can get stuck in cognitive ruts, repeating or refining an approach that isn’t getting them far.
人工智能代理可能会陷入认知困境，重复或改进一种无法让他们走得太远的方法。
Newer models seem to be improving at how they use extended thinking.
新模型似乎在使用扩展思维方面有所改进。
Newer models make better use of extra thinking tokens on these cyber tasks.
较新的模型在这些网络任务中更好地利用了额外的思维标记。
Source.
来源。
This isn’t multi-agent scaling, but it’s encouraging.
这不是多代理扩展，但令人鼓舞。
AIs were poor at turning more thinking into harder solved tasks, and that scaling parameter improved.
人工智能不擅长将更多的思考转化为更难解决的任务，而缩放参数得到了改善。
They might eventually pass the human reference.
它们最终可能会通过人类参考。
That’s part of why I don’t see runaway RSI as imminent, yet think it could become more plausible as these capabilities change.
这就是为什么我不认为 RSI 失控迫在眉睫，但认为随着这些能力的变化，它可能会变得更加合理。
We’ve also seen new findings on helping agents escape their ruts and use more thinking time productively.
我们还看到了一些新的发现，可以帮助特工摆脱陈规，有效地利用更多的思考时间。
The new Fork-and-Flush study periodically creates parallel branches that keep the accumulated work but start with fresh chat contexts, then continues from the best branch.
新的 Fork-and-Flush 研究定期创建并行分支，保留累积的工作，但从新的聊天上下文开始，然后从最佳分支继续。
It improved results across thirteen research and engineering tasks at the same compute budget.
它在相同的计算预算下改进了十三个研究和工程任务的结果。
That’s a concrete research direction for using additional work more effectively.
这是更有效地利用额外工作的具体研究方向。
So, have we seen a similar improvement in multi-agent scaling?
那么，我们在多代理扩展方面是否看到了类似的改进？
For speed, sometimes .
为了速度，有时。
For intelligence, there’s no clear direction , so far, at least on formal math proofs.
对于智力来说，到目前为止，至少在正式的数学证明上还没有明确的方向。
Opus 5.5 gets more speedup for a smaller cost increase on program reconstruction.
Opus 5.5 获得了更大的加速，并且程序重建的成本增加更小。
But the score gains in the chart below are noisy, and newer models don’t consistently get a larger boost from additional agents.
但下图中的分数增益很嘈杂，而且较新的模型并不能始终从额外的代理中获得更大的提升。
Newer models don’t get consistently large gains from larger math-proof teams.
较新的模型无法从规模较大的数学团队那里获得持续的巨大收益。
Source.
来源。
That doesn’t mean better multi-agent scaling won’t happen.
这并不意味着更好的多代理扩展不会发生。
I expect the labs to work hard on coordination.
希望各实验室在协调上下功夫。
Improving diversity of thought is a more open research question, but I’m sure it’s at least somewhat tractable given sufficient effort.
提高思想多样性是一个更加开放的研究问题，但我确信只要付出足够的努力，它至少在某种程度上是容易处理的。
The structure of the problem may impose the hardest limits: some discoveries have to happen before the next useful step becomes possible.
问题的结构可能会施加最严格的限制：在下一个有用的步骤成为可能之前必须有一些发现。
We don’t yet know how much each of these constraints contributes.
我们还不知道这些限制因素的影响有多大。
Summarizing what we’ve seen, here’s an update on the conceptual limits to multi-agent scaling.
总结我们所看到的，这里是多代理扩展概念限制的更新。
Frontier models share many of the same broad training ingredients: overlapping public data, related algorithms, and similar reasoning and coding exercises.
前沿模型共享许多相同的广泛训练要素：重叠的公共数据、相关算法以及类似的推理和编码练习。
So how different are these minds really?
那么这些想法到底有多大不同呢？
The Artificial Hivemind study found striking similarities in models’ open-ended responses.
人工蜂巢思维研究发现模型的开放式反应有惊人的相似之处。
A creativity study involving 467 people likewise found that adding AI responses eventually contributed less diversity than adding human responses.
一项涉及 467 人的创造力研究同样发现，添加人工智能反应最终带来的多样性低于添加人类反应的多样性。
These experiments suggest that AI agents can bring less diversity of thought than an equal number of humans on these tasks, which may limit how much we get from parallel agents.
这些实验表明，与同等数量的人类相比，人工智能代理在这些任务上带来的思维多样性较少，这可能会限制我们从并行代理中获得的收益。
Can we make agents more different from each other?
我们能让代理之间变得更加不同吗？
Very likely.
很有可能。
Here are four ideas.
这里有四个想法。
I’m sure there are others:
我确信还有其他人：
Different training experiences: vary data mixtures, curricula, reinforcement-learning environments, and reward objectives to develop strong agents that favor different useful approaches.
不同的训练体验：改变数据混合、课程、强化学习环境和奖励目标，以开发支持不同有用方法的强大代理。
Complementary specializations: combine agents trained for conjecturing, finding lemmas, searching for counterexamples, and checking proofs.
互补的专业：结合训练有素的智能体进行猜想、寻找引理、搜索反例和检查证明。
Test whether their strengths combine into results no member reliably achieves alone.
测试他们的优势是否能够结合成没有任何成员能够单独可靠地实现的结果。
Independent proposals before discussion: let agents develop approaches before seeing one another’s answers, then exchange evidence and criticism.
讨论前的独立提案：让代理人在看到彼此的答案之前制定方法，然后交换证据和批评。
Test whether this reduces early convergence on the same mistake.
测试这是否会减少同一错误的早期收敛。
Teams chosen for complementary errors: select members whose successes and failures overlap less, rather than relying only on individual benchmark rank.
选择互补错误的团队：选择成功和失败重叠较少的成员，而不是仅仅依赖个人基准排名。
Measure whether the team finds more verified solutions as it grows.
衡量团队在成长过程中是否找到更多经过验证的解决方案。
As we discussed in the previous post, we’re going off of scant data here.
正如我们在上一篇文章中讨论的那样，我们这里的数据很少。
This model could be wrong.
这个模型可能是错误的。
Or newer AI versions could get sufficiently better at doing parallel work that these diminishing returns would change.
或者，较新的人工智能版本可以在并行工作方面做得足够好，从而改变这些收益递减的情况。
If that happens, we need to be aware of it.
如果发生这种情况，我们需要意识到这一点。
How would we track this?
我们如何追踪这个？
I’d propose measuring three things from new AI models across a variety of problem times.
我建议在不同的问题时间内衡量新人工智能模型的三件事。
Across each task type, and especially across real world AI research, measure three parameters and track how they change as newer models arrive:
在每种任务类型中，尤其是在现实世界的人工智能研究中，测量三个参数并跟踪它们随着新模型的出现如何变化：
G - Intelligence scaling: How much verified task performance does each multiplication of agents add?
G - 智能扩展：每个代理的倍增会增加多少经过验证的任务性能？
Measure local score gains, slopes, or exponents as appropriate.
酌情测量局部得分增益、斜率或指数。
This is a task-specific proxy for collective intelligence, not a universal intelligence unit.
这是集体智慧的特定任务代理，而不是通用情报单元。
λ - Speed-cost scaling: How much faster does the team reach the same result, and how much more total compute or agent-time does it use?
λ - 速度成本扩展：团队达到相同结果的速度有多快，以及它使用的总计算或代理时间多了多少？
Track the speed exponent λ over successive team-size ranges.
跟踪连续团队规模范围内的速度指数 λ。
N - Where gains weaken or saturate: At what team size do the score or speed curves bend, or approach a ceiling?
N - 收益减弱或饱和的地方：在团队规模多大时，得分或速度曲线会弯曲或接近上限？
Does the bend or onset of saturation move outward with newer models?
新型号的弯曲或饱和起始点是否向外移动？
Track intelligence gains, the speed-cost tradeoff, and where gains weaken or saturate.
跟踪情报收益、速度与成本的权衡，以及收益减弱或饱和的情况。
G and N answer two different questions.
G和N回答两个不同的问题。
How much extra benchmark score does a defined team expansion buy, and how many agents can we add before those gains flatten?
确定的团队扩展可以购买多少额外的基准分数，以及在这些收益趋于平缓之前我们可以添加多少代理？
In the illustration below, G is the gain from one to ten agents.
在下图中，G 是 1 到 10 个代理的增益。
N is the agent count where the curve reaches its plateau.
N 是曲线达到稳定状态时的人员数量。
A real curve may have a gradual bend rather than a sharp plateau, or no plateau within the tested range.
真实曲线可能会逐渐弯曲，而不是急剧稳定，或者在测试范围内没有稳定。
One hypothetical benchmark curve.
一条假设的基准曲线。
G is the added score from a specified expansion of the team; N is the agent count where gains flatten.
G是指定的队伍扩展所增加的分数； N 是收益趋于平缓的代理数量。
The illustrated sharp plateau is not assumed for every real task.
所示的陡峭平台并不适用于每项实际任务。
Watch these parameters over time, especially G and N.
随着时间的推移观察这些参数，尤其是 G 和 N。
Don’t stop at one to ten agents.
不要止步于一到十个代理。
Measure the gains from 1→10, 10→100, and 100→1,000 agents , separately, on each benchmark.
在每个基准上分别测量 1→10、10→100 和 100→1,000 个代理的收益。
A newer model might scale better over the first interval and worse over the next.
较新的模型可能在第一个时间间隔内扩展得更好，但在下一个时间间隔内扩展得更差。
We need to know whether the gains improve and whether the bend moves outward.
我们需要知道增益是否改善以及弯曲是否向外移动。
Three hypothetical trends for added benchmark score across model generations.
跨模型各代基准分数增加的三个假设趋势。
Measure each expansion separately: 1 to 10, 10 to 100, and 100 to 1,000 agents.
分别测量每个扩展：1 到 10、10 到 100 和 100 到 1,000 个代理。
Different intervals can follow different paths.
不同的间隔可以遵循不同的路径。
We especially need to track these trends for actual internal AI research .
我们特别需要跟踪这些趋势以进行实际的内部人工智能研究。
Are newer models getting bigger gains from larger teams?
较新的模型是否能从较大的团队中获得更大的收益？
Are those teams producing more verified discoveries and better successor models?
这些团队是否产生了更多经过验证的发现和更好的后继模型？
That’s what would tell us whether massive parallel teams are getting closer to the conceptual breakthroughs RSI may need.
这可以告诉我们大规模并行团队是否正在接近 RSI 可能需要的概念突破。
Anthropic has given us a concrete test to watch.
Anthropic 给了我们一个值得关注的具体测试。
Its October employment report shows research agents logging about nine hours for every human work-hour by mid-July , and nineteen by late August .
其 10 月份的就业报告显示，到 7 月中旬，研究人员的每人工作时间约为 9 个小时，到 8 月底则为 19 个小时。
The company-wide ratio reached about six.
全公司的比例达到了六左右。
Research agent runtime rises to about nine hours per human work-hour by mid-July and nineteen by late August.
到 7 月中旬，研究人员的工作时间将增至每人工作小时约 9 小时，到 8 月底则增至 19 小时。
These hours include longer operation as well as parallel agents.
这些时间包括更长的操作以及并行代理。
Source.
来源。
Let’s make a generous assumption: each agent-hour is as valuable as a human research-hour , with the number of human researchers unchanged.
让我们做一个大胆的假设：在人类研究人员的数量不变的情况下，每个代理时间与人类研究时间一样有价值。
Agents can run throughout the 168-hour week.
客服人员可以每周运行 168 小时。
Anthropic’s runtime ratios already count those extra operating hours against a forty-hour human week.
Anthropic 的运行时间比率已经将这些额外的工作时间计入人类每周 40 小时的工作时间中。
On that assumption, July’s nine agent-hours plus one human hour amount to ten times the labor of humans alone.
根据这一假设，7 月份的 9 个代理小时加上 1 个人工小时相当于人类单独劳动的十倍。
August’s total is twenty times.
八月份的总数是二十倍。
But Anthropic already had roughly one agent-hour per human hour in April.
但四月份，Anthropic 已经实现了大约每人每小时 1 个代理小时。
Compared with that observed baseline, total labor grew fivefold by July and tenfold by August .
与观察到的基线相比，总劳动力到 7 月增长了五倍，到 8 月增长了十倍。
What should that do to research?
那应该做什么研究呢？
Treat the added labor as parallel research effort.
将增加的工作视为并行的研究工作。
Davidson and Houlden’s relationships imply about 2–4 times April’s research pace by August , depending on whether experimental compute becomes a bottleneck.
戴维森和霍尔登的关系意味着到 8 月份，4 月份的研究速度大约是 2-4 倍，具体取决于实验计算是否成为瓶颈。
Applying the labor exponent used by Alan Chan and colleagues, including Geoffrey Hinton and Yoshua Bengio , gives about 25 times .
应用 Alan Chan 及其同事（包括 Geoffrey Hinton 和 Yoshua Bengio）使用的劳动指数，得到大约 25 倍。
Log scaling gives about twice the pace.
对数缩放提供了大约两倍的速度。
Our sub-log shapes give only about 1.1–1.8 times.
我们的子对数形状仅给出约 1.1-1.8 倍。
Now translate research pace into the Epoch Capabilities Index (ECI) .
现在将研究速度转化为时代能力指数（ECI）。
Suppose software research supplies half of baseline capability improvement, while the compute-growth contribution stays unchanged.
假设软件研究提供了一半的基线能力改进，而计算增长贡献保持不变。
Then four times the software pace means about 2.5 times the overall ECI slope , not four times.
那么四倍的软件速度意味着大约是整体 ECI 斜率的 2.5 倍，而不是四倍。
We’re already discounting the software boost.
我们已经对软件的提升打了折扣。
Other obstacles could reduce it further.
其他障碍可能会进一步减少它。
Allow six months from research to public release.
从研究到公开发布需要六个月的时间。
Here’s what those assumptions imply for January and February 2027.
以下是这些假设对 2027 年 1 月和 2 月的影响。
The top group uses hypothetical RSI research relationships; the bottom group extends experimental benchmark shapes.
顶层组使用假设的 RSI 研究关系；底部组扩展了实验基准形状。
What different RSI models predict we should see in the pace of AI improvement (using ECI slope), six months after an AI labor increase.
在人工智能劳动力增加六个月后，我们应该在人工智能改进的步伐（使用 ECI 斜率）中看到不同的 RSI 模型的预测。
Source.
来源。
Why do log scaling and Davidson-Houlden’s power 0.3 nearly tie?
为什么对数缩放和 Davidson-Houlden 的幂 0.3 几乎相等？
This is only a tenfold labor increase from April to August.
这只是四月至八月劳动力的十倍增长。
At that scale, both calibrations give about twice the software pace.
在这种规模下，两种校准的速度大约是软件速度的两倍。
By 100 agents and then a billion, the hypothetical power law curve pulls far ahead of the experimentally-observed log and sub-log scaling rules.
通过 100 个智能体，然后是 10 亿个智能体，假设的幂律曲线远远领先于实验观察到的对数和子对数缩放规则。
Around late May, agents were already logging about three hours per human hour: twice April’s total labor on our assumption.
5 月下旬左右，特工们每人每小时的工作时间已经达到了约 3 小时：根据我们的假设，是 4 月份总劳动力的两倍。
With a six-month lag, that earlier signal would arrive around November.
由于有六个月的滞后期，较早的信号将于 11 月左右到达。
These tests have many different assumptions, including equal value per hour and treating added runtime as parallel labor.
这些测试有许多不同的假设，包括每小时价值相等以及将增加的运行时间视为并行劳动。
I’m open to refining them on how others think about this.
我愿意根据其他人对此的看法来完善它们。
In measurement, we should judge the AI pace of improvement as the rate of at which ECI score goes up over successive releases, allowing for development and publication delays.
在测量中，我们应该以 ECI 分数在连续发布中上升的速度来判断 AI 的改进速度，考虑到开发和发布的延迟。
Personally, I don’t expect to see sustained acceleration of Anthropic’s AI progress (as measured by ECI) through the first half of 2027, and certainly not one as large as these simple extrapolations from the hypothetical models of Davidson and Houlden or Chan and colleagues suggest.
就我个人而言，我预计到 2027 年上半年，Anthropic 的人工智能进展（以 ECI 衡量）不会持续加速，而且肯定不会像戴维森和霍尔登或陈及其同事的假设模型所暗示的那样大。
If anything, I expect acceleration at the low end of the sub-log scaling bar at the bottom of the chart.
如果有的话，我预计图表底部的子对数缩放条的低端会出现加速。
I could be wrong for many reasons.
由于多种原因，我可能是错的。
Parallel search of different AI training parameters, or use of AI to code new RL environments, or gains from AIs combining existing ideas in the literature could all make a bigger difference than I expect.
并行搜索不同的人工智能训练参数，或者使用人工智能来编码新的强化学习环境，或者结合文献中现有的想法从人工智能中获得收益，都可能产生比我预期更大的差异。
Or, at the next increment of AI improvement, AIs could start making much more major advances in machine learning theory that allow us to bend the scaling laws and other diminishing returns that AI faces, or to add new axis of scaling, as reasoning models did.
或者，在人工智能改进的下一个阶段，人工智能可能会开始在机器学习理论方面取得更重大的进展，使我们能够改变人工智能面临的缩放法则和其他收益递减，或者像推理模型那样添加新的缩放轴。
On the other hand, while more parallel search and recombination of existing ideas seems highly likely, they may not bring huge advantages, and the AIs we’re making today might not find conceptual breakthroughs.
另一方面，虽然对现有想法进行更多并行搜索和重组似乎很有可能，但它们可能不会带来巨大的优势，而且我们今天制造的人工智能可能无法找到概念上的突破。
Atop that, ideas probably really are getting harder to find in AI, and that may chew up some or all of the increased capacity of AI to work on improving itself.
最重要的是，在人工智能中寻找想法可能确实变得越来越困难，这可能会削弱人工智能自我改进的部分或全部能力。
We shall see.
我们将会看到。
If we see the rate of AI improvement increase significantly, I’ll update my views in proportion to what the data show.
如果我们看到人工智能改进率显着提高，我将根据数据显示的比例更新我的观点。
In summary, for now, multi-agent scaling shows incredibly steep diminishing returns to model intelligence, or the ability to reach conceptual breakthroughs.
总之，就目前而言，多智能体扩展显示出模型智能或实现概念突破的能力的回报急剧递减。
I expect those sorts of breakthroughs to be necessary for major AI advances and for self-sustaining RSI, let alone a runaway intelligence explosion.
我预计这些突破对于人工智能的重大进步和自我维持的 RSI 是必要的，更不用说失控的智能爆炸了。
And at the same time, we’re all operating on far too little real world data.
与此同时，我们所使用的现实世界数据太少了。
The most important thing we can do now and for the months and years ahead is to rigorously collect that data, and use it to update our forecasts of the future.
我们现在以及未来几个月和几年可以做的最重要的事情就是严格收集这些数据，并用它来更新我们对未来的预测。