主题综述

长程智能体的下一步 · Long-Horizon Agents

主题综述

更新日志

主流共识

第一点:agent 不可靠这件事是结构性问题,不是当前模型版本的瑕疵

"The issue with agents is they aren't reliable to nine nines of reliability, but they can do a ton of work and more and more work over longer time horizons."
「agent 的问题是它们没法做到九个 9 的可靠性,但它们能完成大量工作,并且在更长的时间跨度上做越来越多。」

第二点:长程任务真正生效的场景,目前是"生成一个初稿,由人来审"——而非端到端自动化。

"If you can find these framings where they run for a long period of time but produce a first draft of something, those to me are the killer applications of long-horizon agents right now. Coding is an example of that. Coding, you usually put up a PR. You don't directly push to prod… AISREs usually surface it to a human who comes in and then reviews it."
「如果你能找到那种'跑很久、但产出一个初稿'的框架,那些在我看来就是当下长程 agent 的杀手级应用。编码就是一例——编码,你通常是提一个 PR,而不是直接推到生产……AI SRE 通常是把结果交给一个人来 review。」

2026-08 注脚:这条共识的两端都在被拉扯。一端是"已经在干活"——OpenAI 的 Mark Chen 观察到专业人群的醒悟时刻:

"It feels like a lot of people woke up at the start of this year and were like, man, agents are working in my profession. And they're essentially realizing that these models can just do long horizon meaningful work for them."
「我觉得在今年年初,很多人都像突然醒悟一样,感觉代理人在我的职业中发挥作用。他们基本上意识到这些模型可以为他们做长期有意义的工作。」

另一端是"整类需求还没人开跑"——Mercor 的 Osvald Nitski 说市面上"90% 企业工作流已被模型覆盖"的算法漏掉了潜在需求:

"there's a whole category of latent demand that people aren't even, these are things that people aren't even trying to do with models yet. Most commonly, we think these are like long-horizon tasks, like setting up a procurement agent to fully automate your procurement team for months on end. You only check on it maybe once a week."
「但还有一个人们甚至没有意识到的潜在需求类别,这些是人们尚未尝试用模型实现的事情。最常见的是,我们认为这些是像长期任务,比如设立一个采购代理来完全自动化你的采购团队,持续数月。你可能每周只检查一次。」
Osvald Nitski (Mercor) · 20VC: Mercor CPO

第三点:reward hacking 是真实存在的训练问题,且 reward function 越复杂越难治。

Will Brown(Prime Intellect)观察到 Claude 3.7 会出现 reward hacking——在完成被要求的任务时夹带一些不必要的动作,Anthropic 在 Opus 上正努力压低这种行为。(该集为英文原音、podwise 仅存中文译文,故此处转述、不作逐字引用。)

第四点(2026-08 新增):"能力"不再是模型的固定属性,而是测试时算力预算的函数。这直接改写了长程 agent 的评估问题。

"The problem is we're in a world now where The capability of the model is a function of how much money you put into it, basically. If you give it a budget of $10,000, it can do a lot more than what it can do with a budget of $10. Give it a budget of $10 million, it can do even more. At what budget should you evaluate these models?"
「问题是我们现在的世界中,模型的能力基本上取决于你投入多少资金。如果你给它 10,000 美元的预算,它能做的比给它 10 美元时多得多。给它 1000 万美元的预算,它能做的更多。在什么预算下你应该评估这些模型?」
Noam Brown (OpenAI) · Why Traditional Benchmarks Fail

而且长程运行本身就是新增能力的来源:

"as the models have become stronger, they're better able to operate over longer horizons. So again, with GPT-3, if you wanted to run it for, you know, a week, there's really not much you could do to scaffold it into something useful that could actually run for a week. But we're seeing now with the most recent models that you can actually scaffold, for example, 5.5 into doing a series of experiments that can run for weeks, for months."
「随着模型变得更强,它们能够在更长的时间范围内更好地运作。所以,如果用 GPT-3 运行一周,其实很难把它搭建成一个可以实际运行一周的有用东西。但现在我们看到最新的模型,你实际上可以将 5.5 搭建成进行一系列可以运行几周或几个月的实验。」
Noam Brown (OpenAI) · Why Traditional Benchmarks Fail

同方向的回声:Anthropic 平台团队把"token 不可互换、要给 token 分派不同工作(建议 / 执行 / 做梦)"当作协调层的设计前提(Anthropic 平台团队);Mark Chen 则称行业在"评估危机"里——benchmark 一公开就不再是好评估(Mark Chen)。

第五点(2026-08 新增):环境(environments)成为横跨各阵营的稀缺资产。RL 派要它当训练场(Prime Intellect 的"RL 环境的 GitHub"),数据供应商在卖它,world models 派在造它。

"The data type that's growing the fastest for us is environments. … the data that the models are now the agents are being evolved and trained on looks a lot closer to what they see in deployment, right? So if you want to learn how to use something like Salesforce, you need a pretty high fidelity mock that acts exactly like Salesforce in your eval."
「对我们而言,增长最快的数据类型是环境。……现在模型的代理正在进化和训练的数据与他们在部署时看到的东西有很多相似之处,对吧?所以如果你想学习如何使用类似 Salesforce 的东西,你需要一个非常高保真度的模拟,能够在你的评估中完全像 Salesforce 一样运作。」
Osvald Nitski (Mercor) · 20VC: Mercor CPO

为什么环境是瓶颈,Poolside 的 Eiso Kant 给了机制解释:RL 受 batch size 约束——你没有无限多的任务,不能像预训练吃全网数据那样无限扩 batch,所以 RL 的算力扩展方式和预训练根本不同(Poolside)。任务 / 环境的存量,就是 RL 派的天花板。

2026-08-24 注脚:供给端的机制这次被 Mercor CEO Brendan Foody 摊开——RL 环境 = worlds(对应真实项目的文档 / 邮件 / 表格语料)+ apps(Salesforce、Microsoft 365 级的高保真克隆)+ tasks(prompt + 验证器)三件套,单季吞吐已达 250 万专家小时。而人之所以无法被替换出这条流水线,是因为判分本身需要越过模型能力的前沿:

"Only humans can measure the frontier in most domains, not every domain. … in most domains, like building a slide deck, The model has an incredibly hard time identifying reliably where it made its own mistake. It's as if you would be asking a human to grade their own homework."
「在大多数领域,只有人类能够测量前沿,并非所有领域都是如此。……在大多数领域,如制作幻灯片,模型在可靠识别自己犯错的地方时会遇到极大的困难。就像你在要求一个人给自己的作业打分一样。」
Brendan Foody (Mercor) · RL Environments Explained

下一步的建设方向直接改写"长程"的量纲:

"The first one is ultra long horizon. Like right now agents mostly aren't trained to do things that are over 10 hours. And we need to start building tasks for things that might take a human a hundred hours or even a thousand hours to do. And so that's going to be a giant shift."
「第一个是超长时间跨度。现在,代理主要未经过训练以进行超过 10 小时的任务。我们需要开始创建那些可能需要人类 100 小时甚至 1000 小时才能完成的任务。这将是一次巨大的转变。」
Brendan Foody (Mercor) · RL Environments Explained

但注意:这条共识第一次出现了正面反对者——Rich Sutton 把"人造环境 + 合成数据"整条路线判为大错,理由与 Eiso 的 batch size 约束同构(都卡在人类供给上),只是走得更远:他认为出路是 agent 从自身经验学习、自建模拟器,而不是雇更多专家造环境。详见阵营 D 架构翼。

分歧在哪

阵营 A · "Multi-turn RL 是终极范式"——ReflectionAI / Prime Intellect 立场

Misha Laskin (ReflectionAI) 给得最干脆:

"It became pretty clear to us that the next paradigm and effectively the final paradigm that we need to have in place before what people used to call AGI or now I think the goalposts have shifted to ASI is reached is just figuring out how to scale reinforcement learning on top of large language models."
「我们越来越清楚:在到达原来的 AGI(现在大家把目标改成 ASI)之前,要落地的下一个、也基本上是最后一个范式,就是搞清楚怎么在大语言模型之上把强化学习 scale 起来。」

但 Misha 自己也给出了一个让 RL 派难做的承认——generalization 可能不存在

"Maybe the hot take is that there's no such thing as generalization. There's just bringing the test distribution into train."
「也许更尖锐的说法是:根本没有所谓的泛化,只是把测试分布塞进训练。」
Misha Laskin · Asimov

Will Brown / Johannes Hagemann (Prime Intellect) 走基础设施侧——他们押开源 RL environments 是下一个 GitHub:

"RL allows you to trade off compute for data, in a sense, where you can get a lot of value out of a smaller amount of data by using more compute."
「RL 允许你用计算换数据——用更多算力,就能从少量数据里榨出大量价值。」

Will 也是这一派里最公开承认 reward hacking 风险的人。他有两条值得记的观察(均出自英文原音、podwise 仅存中文,故转述):一是很多"耸动的 AI 安全结果"在他看来其实是两难困境——模型被给了两个互相冲突的目标、只能选一个,选哪个听起来都很糟;二是对"用工具"本身给奖励,会诱发"虚拟工具调用"——模型调了工具、却对推理毫无帮助,纯粹是为了拿奖励。

2026-08 新增:OpenAI 的 Mark Chen 从正面划出了 RL 的边界——它起飞的地方全是可判分的领域。这与 Misha 的 hot take 同向:奖励可判 ≈ 分布可控。

"So it's these fields where things are hard to grade, where RL has the least amount of ability to go and directly apply there. I know a lot of people are developing techniques to apply RL in these settings, but For now, it's just where there's cold hard truth. Things like math and computer science where you implement it correctly or wrong."
「所以这些领域的东西难以评估,强化学习在这里的直接应用能力最弱。我知道有很多人在开发技术来把强化学习应用到这些环境中,但目前来说,那里才有冰冷的硬实话语。像数学和计算机科学这样的领域,你要么正确实施,要么错误实施。」

供给侧也印证了这条路线的真实资金流向:Mercor 说环境是他们增长最快的数据类型(见共识第五点),买家正是前沿实验室。另外,阵营 A 还挨了一记来自内部的冷枪——前 OpenAI RL 负责人 Jerry Tworek 自称"RL maximalist"出身,如今判定 RL 只是"从经验学习"的一种数学表达、且效率不高,详见下面新增的阵营 D。

(2026-08-24 追记:第二记冷枪来得更重——RL 之父 Rich Sutton 判整条"人造环境 + 合成数据"供给链是大错,见阵营 D 架构翼。同时 Foody 在供给端给出 pass@1 vs pass@16 的可训练性经验法则:16 条轨迹全错,说明模型还学不动这份数据;pass@1 失败但 16 条里偶尔对一两次,才是可训窗口(RL Environments Explained)。)

阵营 B · "脚手架在做大部分工作"——LangChain 立场

Harrison Chase 不直接反对 RL,但他把"长程 agent 突然能用"的原因放到了模型外面——归功于 context engineering(compaction、sub-agents、skills 这些管理上下文的工程手段,也是他那期的标题主张)。他给的最锋利的一条,是 agent 和传统软件的本质区别——应用逻辑不再都在代码里

"When you're building software, all of the logic is in the code in the software and you can see it there. When you're building an agent, the logic for how your applications works is not all in the code. A large part of it comes from the model."
「写软件时,所有逻辑都在代码里,看得见。写 agent 时,应用怎么工作的逻辑不全在代码里——很大一部分来自模型。」
Harrison Chase · Context Engineering

Kyle Corbitt (OpenPipe) 给的是阵营 A 和阵营 B 之间的一个中间观察——RL 重要,但工程化路径并不像"训练吃数据"那么简单:

"One pro is just sort of like operational simplicity. Like there's a whole extra model you need for this value model you need for PPO that you can throw away with GRPO. … So the way GRPO works is you have to do a set of different trajectories or a set of different rollouts all in parallel with the exact same environment, the exact same conditions, and then you score each of them."
「一个好处就是操作上简单——PPO 需要一整个额外的 value 模型,用 GRPO 可以把它扔掉。……GRPO 的工作方式是:你必须并行跑一组不同的轨迹、一组不同的 rollout,环境完全相同、条件完全相同,然后给每个打分。」

关于"何时该 fine-tune",那个流传的"成本、延迟、质量一致性"三要素框架其实出自主持人 Swyx 之口——他在节目里复述 Kyle 早年在 AI Engineer World's Fair 演讲中给的建议(Kyle 本人没说过这句原话)。Kyle 的回应是:

"Yeah, I mostly stand by that. I don't think it's changed. … But the main one I see that really drives fine-tuning is if you have to move to a smaller model, and it's typically for latency reasons, and this is usually like real-time voice. … I would say for 90% of use cases where you aren't forced to a smaller model, then it's still not a good ROI and you probably shouldn't invest in it today."
「是的,我基本上坚持这一点。我认为它没有改变。……但我看到真正推动微调的主要原因是,如果你必须迁移到更小的模型,这通常是出于延迟的原因,而且这通常就像实时语音一样。……我会说,对于 90% 的非被迫使用较小模型的用例,它仍然不是一个好的投资回报率,你今天可能不应该投资它。」

2026-08 新增 · B 阵营拿到了最主流的背书,同时开始自我瘦身。

Satya Nadella 几乎逐句复述了 Harrison 的论点——魔法在模型外面的上下文层:

"the amount of work you need to do to prep the context layer such that your plan can execute in the most efficient way. This is where the magic is."
「你需要做的准备上下文层的工作量,以便你的计划能够以最有效的方式执行。这就是魔法所在。」
Satya Nadella (Microsoft) · The Rise of the Full-Stack Builder

而且他给出了长程 agent 进生产的最具体案例——Azure 网络团队把自己的工作重构成"造一个做这件事的 agent 系统"(他们叫它 Miles):

"And so they were saying, look, we don't need headcount. We need tokens in order to be able to manage our operation. That reconceptualization of what their work is. They basically took their work and made it meta. That meta work is now their new work."
「他们在说,我们不需要人手。我们需要令牌来管理我们的操作。这种对他们工作的新构思。他们基本上把他们的工作做到了更高的层次。这种更高层次的工作现在成了他们的新工作。」
Satya Nadella (Microsoft) · The Rise of the Full-Stack Builder

但 Anthropic 平台团队——本身就是 B 路线的供应商——承认脚手架的一半正在被模型吃掉,harness 的职责在换:

"And now the models are actually very, very steerable. And so a lot of that steering, you can just put in the prompt, right? … if you have harnesses that are like designed to kind of do that kind of like steering, You can delete that part. … But as a result of this, what the harness needs to start doing is more allow it to run longer."
「现在模型实际上非常可控。……如果你有为了做这种控制而设计的工具,你可以删除那部分。……但因此,工具需要开始做的是更允许它运行更长时间。」
Angela Jiang (Anthropic) · Building an Ecosystem, not a Walled Garden

走得最远的是 Poolside 的 Eiso Kant——脚手架最小主义。他不否认 harness 有用,但把它定性成权宜之计:

"the harness is the stop gap between what the model is capable of And what it needs as additional instructions and what it needs is access to data and tools"
「harness 是一块权宜的补板——补在模型已具备的能力,与它还需要的额外指令、它还需要的数据与工具访问之间。」
Eiso Kant (Poolside) · Inside the Model Factory

并且给了一个可证伪的到期日:

"And so I'm a very big fan of Give the model a minimal harness, as minimal as possible. Give it a container in which it has its own code base, right? … I think we will, in 12 months, not see a single system prompt that is stuffed with 20 or 30 or 40 tools anymore."
「所以我非常拥护:给模型一个最小的 harness,越小越好。给它一个容器,里面是它自己的代码库……我认为 12 个月后,我们将不会再见到任何一条塞着 20、30、40 个工具的 system prompt。」
Eiso Kant (Poolside) · Inside the Model Factory

Eiso 的底气来自模型侧:他们发现长程能力的增益可以来自训练进权重里的行为而非更多智能——这正是过去靠脚手架兜底的东西。他引用同事 Peng Ming 对 Laguna S 的判断:

"I have the feeling that a lot of the gains in Laguna S come not from more intelligence, but more from different behavior, more verification, less taking things for granted, not declaring victory early and being way more persistent."
「我有种感觉,Laguna S 的很多增益并非来自更多智能,而更多来自不同的行为:更多验证、更少想当然、不提前宣布胜利、以及执着得多。」
Eiso Kant(转引同事 Peng Ming)· Inside the Model Factory

2026-08-24 追记:"转向 / 工具堆砌类脚手架会死"又添一张来自阵营外的选票——Trajectory(持续学习平台,立场上属于阵营 D)给客户的 harness 建议与 Anthropic / Eiso 的瘦身方向完全合流:harness 不再是防呆的流程锁,而是产品原语的编排面。

"I would view building the harness as not enforcing specific flows that need to be done, but more what are the primitives that your product has, whether that's search tools or private information, and then viewing the agent as orchestrating those primitives and then letting it do that."
「我认为构建工具带的方式不是强迫执行特定的流程,而是更多地关注你的产品具有哪些基本操作,无论是搜索工具还是私人信息,然后将代理视为协调这些基本操作,并让它去执行这些操作。」

所以 B 阵营现在的真实分歧不是"要不要脚手架",而是脚手架的哪一半会死:转向 / 工具堆砌那一半(Anthropic 和 Eiso 都判了死刑),还是验证 / 上下文准备 / 让它跑更久那一半(Satya、Anthropic、连 Eiso 都保留了容器 + 数据访问)。Harrison 的 compaction / sub-agents / skills 处在两者之间,尚无人正面宣判。

阵营 C · "World Models 是基础设施"——General Intuition / Google DeepMind 立场

Pim De Witte (General Intuition) 押的是一种与 RL / context 路线都不同的方向:

"What world models do is they actually have to understand the full range of possibilities and outcomes from the current state, and based on the action that you take, generates the next state, the next frame."
「world model 干的事是:理解当前状态下所有可能的结果,根据你采取的动作,生成下一个状态、下一帧。」

GI 的论据是数据 + 行为模仿:3.8B 游戏剪辑,4 秒空间记忆就能展现 superhuman 行为——他们拒了 OpenAI 5 亿美元的数据收购,自己做实验室。

Google DeepMind Genie 3 的研究者讲的是同方向但更基础设施化:

"It's at the point where a human who is not an expert will watch it and think it looks real."
「现在到了非专业人士看了会觉得"看起来是真的"的程度。」
"A year ago when we were thinking we'll get a minute of consistency for an autoregressive model like this in real time, people thought that was like our stretch goal kind of thing. … and people say, but a minute actually isn't long enough yet. I think that's a sign of the progress, right?"
「一年前,当我们认为我们可以在实时环境中为一个像这样的自回归模型获得一分钟的连贯性时,人们认为这就像我们的延伸目标一样。……人们说,但一分钟实际上还不够长。我认为这是一个进步的标志,对吧?」

需要注意——Camp C 的应用领域目前几乎全在游戏 / 视觉生成,跨到"通用 agentic 任务"的路径在语料里没有展开

阵营 D(2026-08 新增 · 2026-08-24 扩编)· "持续学习 / 记忆才是缺的那块"——Engram / Core Automation / Oak Lab / Trajectory 立场

这一派的共同判断:长程 agent 卡住的不是智能、不是脚手架,而是模型无法在部署后继续学习。每次会话从零开始的 agent,永远只是"第一天上班的实习生"。(2026-08-24:本次 6 篇新语料有 5 篇属于或直击这一派——它已从两家公司的主张扩成一个赛道,并且内部清晰地分出了架构翼产品翼。)

架构翼:换范式,不救现有模型。

最重的一击来自 Jerry Tworek——前 OpenAI 推理 / RL 核心人物,自称当年的"RL maximalist",现在把问题归到范式本身:

"What I think is the issue is that the models are being trained in the lab. And are being deployed in the real world. That is the fundamental tension that is there."
「我认为问题在于模型是在实验室中训练的。并且实际上是在现实世界中部署的。这就是存在的根本张力。」
Jerry Tworek (Core Automation) · Building the Automated AGI Lab

他的具体论证(转述):测试时学习只有两条路,in-context learning 数据量太小——他用 codex 大约 20 分钟就得 compact 一次;持续微调则撞上灾难性遗忘。两条都不通,所以他断言 transformers 无法承载持续学习、必须换架构,Core Automation 就是为找"transformer 替代品"而建的实验室(同集)。他还顺带打了阵营 A 一枪:RL 只是"从经验学习"的一种数学表达,用长长的思维链换一比特反馈,效率低得可疑。

2026-08-24 新增:架构翼迎来分量最重的援军——RL 之父本人。Rich Sutton(《苦涩教训》作者)与学生 Khurram Javed 创办 Oak Lab,立场比 Tworek 还要彻底:不是 transformer 有缺陷,而是"冻结权重"这个范式本身才是历史的反常——

"Before there was all this AI craziness… you wouldn't have to say continual learning because it wouldn't make any sense to talk about learning that wasn't continual. All learning is continual. We always act and we learn. That's just the normal way of thinking. I'm not weird. The field is weird."
「在所有这些人工智能疯狂之前……不需要提到持续学习,因为谈及不持续的学习完全没有意义。所有学习都是持续的。我们总是行动并学习。这就是正常的思维方式。我并不奇怪。这个领域很奇怪。」
Rich Sutton (Oak Lab) · Why AI Models Stop Learning

对阵营 B 的整个上下文工程路线,他们给出釜底抽薪的一句——喂上下文只是改变状态,不是学习:

"Large language models are, oh, we're going to put all this knowledge into the system and the large language model will not learn when it runs. It's talking to people. It's interacting. It is absolutely, the weights never change."
「大型语言模型是,哦,我们将把所有这些知识放入系统中,而大型语言模型在运行时并不会学习。它是在与人类对话。它在互动。权重绝对不会改变。」
Rich Sutton (Oak Lab) · Why AI Models Stop Learning

阵营 D 一直缺的理论地基也由他们补上——大世界假说:

"The big world hypothesis, let's say what it is, is that the world is massively more complex than your mind, than any agents, any agent. And this is obvious because the world contains many other agents."
「大世界假说,简单来说,就是世界的复杂性远超你的思维,超过任何代理,任何代理。这很明显,因为世界中包含许多其他代理。」
Rich Sutton (Oak Lab) · Why AI Models Stop Learning

这条假说顺手把炮口对准了阵营 A 的供给链。主持人问"合成数据生成算不算利用计算的通用方法",Sutton 的回答只有一句:

"No, that's just a big mistake."
「不,这真是个大错。」
Rich Sutton (Oak Lab) · Why AI Models Stop Learning

Khurram 给出机制论证:

"But if all the engineers, OpenAI Anthropic or all the big Neolabs, the engineers went on vacation, who would generate the synthetic data? That's the question. It doesn't come from agent's experience. It's not something that the agent is generating itself. Some human has to decide what is the right synthetic data to generate. And that requires human expertise."
「但是如果所有的工程师,无论是 OpenAI 还是 Anthropic,或者所有大型 Neolabs 的工程师都去度假,谁来生成合成数据?这就是问题所在。它并不是来自于代理的经验。这不是代理自己生成的东西。需要由某个人工来决定生成什么样的合成数据是正确的。这需要人类的专业知识。」
Khurram Javed (Oak Lab) · Why AI Models Stop Learning

注意这一枪打中的是谁:Mercor 正靠单季 250 万专家小时的人力造环境(见共识第五点),Prime Intellect 在做"RL 环境的 GitHub"。Sutton 的立场等于说:这整条供给链是在用人类专业知识给"大世界"打补丁,注定跟不上——agent 应该从自身经验学模型、自建模拟器。这是共识第五点第一次遇到正面反对者,而且来自这一派奉为圣经的作者本人。

架构翼内部也裂出了第一道缝。Tworek 断言持续微调撞灾难性遗忘、transformer 承载不了;Sutton 却说这病能治:

"But it's totally curable. You have to have the right algorithm."
「但这是完全可以治愈的。你必须要有正确的算法。」
Rich Sutton (Oak Lab) · Why AI Models Stop Learning

药方(转述):每个权重一个独立的元学习步长(只让网络的相关部分适应新信息)+ continual backprop——在特征空间不断注入随机初始化的新单元做"生成与测试",几年前已发表于 Nature(同集)。但 Khurram 补了一个让产品翼难受的限定——这套算法无法后装到现有模型:

"I don't think that you could take an existing model and say, I'm going to just start updating it with these algorithms because these algorithms meta learn how to learn. So really you have to say, I'm going to learn from scratch."
「我认为你不能拿一个现有的模型并说,我将用这些算法开始更新它,因为这些算法是元学习如何学习的。所以实际上你必须说,我将从头开始学习。」
Khurram Javed (Oak Lab) · Why AI Models Stop Learning

以及为什么大实验室不做(转述):行业困在局部最小值里,转向新范式"几乎不可能不先变差再变好",被产品和当前 scaling 路线锁死的实验室承担不起这段下坡(同集)。

产品翼:不等新架构,在现有模型上把部署经验变成权重。

Engram 的持续训练论点(见下方 Jessy Lin / Biderman 引语)本次得到 Trajectory 的正面增援——创始团队来自 Windsurf(SWE1 模型)、DeepMind 机器人组与 Apple Vision Pro。Arjun Karanam 给了这一派最好记的公式:模型在 IQ 轴上狂奔,经验轴却纹丝不动——

"if you take this analogy of a Terence Tao in your pocket, that's great. But Terence Tao, day one at an accounting firm, is probably not the best accountant there. … And so we have this orthogonal axis of experience that is much more important than we think in conjunction with IQ. And that's what we're calling the experience gap."
「如果你把这个比喻比作口袋里的 Terence Tao,那太好了。但 Terence Tao 第一天在会计师事务所工作时,可能并不是最好的会计。……因此,我们有一个与智商并行的经验维度,这比我们认为的重要得多。这就是我们所称的经验差距。」

Ronak Malde 把同一件事说成软件史的反常:

"it kind of acts like any other software in that it's very static. Like the model that you used yesterday is going to be the same model and making the same mistakes tomorrow. And all of those corrections you gave it, the edits, like in any product is all just being put to waste."
「它实际上与其他软件一样,非常静态。你昨天使用的模型明天仍然会是同一个模型,并犯同样的错误。你给它的所有修正、编辑,像任何产品一样,都是浪费掉的。」

以及为什么受监管行业等不起"9 个 9"的渐近曲线——这给共识第一点补了一个行业侧的推论:

"For a field like legal, like getting 80% of the way there is the same thing as zero."
「对于法律这样的领域,达到 80% 的结果与 0% 的结果是一样的。」

燃料是现成的——每天的推理量本身就是没人捡的训练信号:

"we are actually spending hundreds of trillions of tokens every single day on inference. And we're generating great amounts of data on how models in the real world are failing, how they're doing well, and that should be signal that we should be capturing and training on."
「我们实际上每天在推理上花费数百万亿代币。而且我们正在生成大量的数据,关于模型在现实世界中如何失败、表现良好,这应该是我们应该捕捉和训练的信号。」
Ronak Malde (Trajectory) · Scaling up Continual Learning

值得记一笔的是一次方法论对撞:Tworek 在架构翼指控 RL"用长长的思维链换一比特反馈、效率低得可疑"(见上),Trajectory 的 OPSD(on-policy self-distillation)恰好是对这条指控的正面回应——把生产环境里的用户修正当"特权提示"喂给教师侧,学生逐 token 对齐整个词表分布,不再把丰富反馈压成一个标量;他们报告在 Mercor Apex(100+ 工具调用)任务上超过了 RL 基线,顺带命名了这条路线自己的 reward hacking 等价物(hint leakage:提示泄漏进输出、模型走捷径跳过推理)与"but / wait 对冲词坍缩"问题(Scaling up Continual Learning)。产品翼不再只有商业主张,开始有自己的算法学了。

但这一翼的三句自白值得原样记录。Ronak 承认今天所谓持续学习的真实形态:

"as a research community right now, we're in this zone of what I call pseudo continual learning, where there's some still level of like batch updates offline, and then re uploading the model."
「作为一个研究社区,我们目前正处于我所称之为伪持续学习的阶段,在那里仍然存在一些批量更新离线,然后重新上传模型的情况。」
Ronak Malde (Trajectory) · Scaling up Continual Learning

Engram 的 Jack Morris 则把这一翼的天花板亲口说破——问题是真的:

"It's like the core problem with the current paradigm in AI that models cannot acquire new knowledge after training in a personalized way."
「这就像是当前人工智能范式的核心问题,模型在训练后无法以个性化的方式获取新知识。」
Jack Morris (Engram) · Scaling Compute on Context

但所有现行方法(对私有语料做 next-token 续训、KV 压缩、on-policy 蒸馏、合成数据模拟预训练)都会撞"数据墙":

"When we started the company, we generated curves that look just like this blue curve where no matter sort of how much data we generate or how much we train, we kind of do plateau because there's this Almost like natural upper bound to how much you can learn in one go from D."
「当我们创办公司的时候,我们生成的曲线看起来就像这条蓝色曲线,无论我们生成多少数据或者训练多少,我们总是会达到一个平稳状态,因为从 D 中一次性学习的量几乎有一个自然的上限。」
Jack Morris (Engram) · Scaling Compute on Context

他还点破这个领域连名字都没统一(sleep-time compute / neural memory / dreaming / machine studying……),"范式还太早,没像预训练 / 后训练那样固化"(同集)。换句话说:D 阵营自己的研究员承认,"实习生时刻"目前没有任何一条路线能给出保证——缺的那块叫递归自我改进(模型自己给自己出越来越难的题,与 AlphaGo 自我加难同构),在私有语料上还没人做成。

Engram 的原始论点保留如下。Jessy Lin 把矛头直接对准阵营 B 的天花板:

"the first proof of concept of the thing that people keep talking about with continual learning, which is you have an intern that you can teach things over time, and it actually gets better. I think everybody's waiting to see that, you know, and no matter how sophisticated the context engineering approaches are these days, they're not getting there."
「人们一直在谈论的持续学习的第一证明概念,就像是你有一个可以随着时间教授的实习生,而它实际上会变得更好。我认为每个人都在等待看到这一点,你知道的,无论现在的上下文工程方法多么复杂,它们都没有达到那个水平。」
Jessy Lin (Engram) · Memory and Continual Learning

经济学论据是两个数量级的推理成本差:

"That can give you, you know, two orders of magnitude reduction in token inference consumption. … These kinds of things the model can kind of implicitly learn by training on some of the data and answer, you know, within 100 tokens what the best frontier models would consume 100,000 tokens."
「这可以让你,知道,减少两个数量级的令牌推断消耗。……这些模型可以通过训练一些数据隐式学习,并在 100 个令牌内回答,而最佳前沿模型可能需要消耗 10 万个令牌。」
Dan Biderman (Engram) · Memory and Continual Learning

而效率不只是省钱——Biderman 在另一集里把它直接接到本主题上(他同时点了 context rot:喂得越多、模型越糊涂,长上下文和 compaction 都到不了整体推理):

"if you can do the same thing with fewer resources, When you scale up to very large resources, suddenly you can take on tasks that were previously not possible. Way more long horizons. Way more adaptive."
「如果可以用更少的资源做同样的事情,当你扩大到非常大的资源时,突然间你可以承担之前不可能的任务。更长的时间范围。更加自适应。」
Dan Biderman (Engram) · The AI Memory Problem

两处张力值得记账。其一,对旧成员 Kyle Corbitt 的"90% 用例微调不是好 ROI"(见阵营 B),Engram 的整个商业模型就是反命题——Kyle 说的是静态一次性微调的 ROI,Engram 赌的是持续训练摊薄后的 token 经济学,两者并未正面交锋,但不可能都对。(2026-08-24:Trajectory 的商业模型立在同一个反命题上,且已有 Harvey、Clay、Decagon、Rogo、Mercor 五家在产、与 NVIDIA 合训 Nemotron 3 Super 进 Harvey 产线——反命题开始有产线数据了,虽然仍是自报。)其二,前沿实验室并不接这个框架——Mark Chen 被问到"到 AGI 还需要持续学习这类突破吗"时直接拆台:

"I don't know if that same framing like continue learning is a basic primitive that you have to unlock. There's so many different techniques. … but I think there are clearly many shots on goal and I am pretty sure they'll work."
「我不知道是否那种框架,比如持续学习,是你必须解锁的基本原语。有很多不同的技术。……但我认为明显还有很多机会,而且我很确定它们会奏效。」

暗流 · "应用层 vs 通用代理"——Misha 的另一条线

Misha Laskin 在阵营 A 里同时给出了一个跟其他 RL 派不一样的微妙差异:

"Just because a company might have the best model in some general set of academic benchmarks, doesn't actually mean they have the best product."
「就算一家公司在一些通用学术 benchmark 上有最好的模型,也不意味着他们有最好的产品。」
Misha Laskin · Asimov
"When you look at what an engineer does in an organization, 80% of the time they're spending trying to comprehend complex systems and collaborating with teammates."
「在一个组织里看工程师做什么,80% 的时间花在理解复杂系统和与队友协作。」
Misha Laskin · Asimov

Asimov 不是 code generation agent,是 comprehension agent——这条产品方向跟 Will Brown 那套通用 multi-turn RL 路线在战术上是不同的,虽然他们都自称 RL 派。

最后,Tuhin 把这个题目跟推理经济学接到了一起:

"From the developer's perspective, they would insert more intelligence if you make it cheaper. They will insert more intelligence anyway, but if you make it more cheaper, they'll insert a hell of a lot more intelligence."
「从开发者的视角看,你把智能做得更便宜,他们就会塞进更多智能。其实他们本来就会塞更多,但你要是做得更便宜,他们会塞进多得多的智能。」

都没说透的

我的看法

判断(不是事实):这三条路线短期内会并行而不是收敛。1–2 年看:阵营 B(context engineering)在生产端做绝大多数的工作,因为它能套在任何前沿模型上,迭代快;3–5 年看:阵营 A 的 RL post-training 会在"垂直任务"上跑出更可靠的 agent,但 generalization 的边界比现在的乐观预期窄得多(Misha 的"hot take"我会赌它是真的);5–10 年看:阵营 C(world models)有可能成为 long-horizon 真正的基础设施,但目前完全在游戏 / 视觉域,跨域转移这一步没人有可信的路径图。

2026-08 修订(保留上面的原判断,以下是增量):

把握程度:中等偏低(不变)。短期 B 主导:信心高→更高;中期 RL 边界:一般(Mark Chen 的 cold-hard-truth 边界与 Misha hot take 互相印证,我略增信心);长期 world models:仍明显低;阵营 D:证据还太少,只下小注。

2026-08-24 修订(以上判断原文保留,以下是增量):

还想知道什么

取材