跨国深谈

harness

44分钟阅读

跨国深谈← 返回←harness
HA

harness

概念
本站收录 65 集 · 35 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

35 条

就像如果我只是让它生成，那将是垃圾。如果我把我自己的上下文给它，像我的策展、我的品味、我的 harness，它就会说，嗯，不要供应商之类的东西，或者它必须有四个声音。
Like if I would just have it generate, it will be rubbish. If I had like my own context to it, like my curation, my taste, my harness, it would be say, well, not vendors or kind of like things or like it has to have four voices.
—— Patrick Debois · [12:35]

指向原始笔记的链接

所以这里出现了一种类比:数据之于模型,正如模型之于 harness。
So there’s a sort of analog that emerges where it’s what data is to a model, models are to a harness.
—— Matan Grinberg · [20:06]

指向原始笔记的链接

我认为另一个有趣的问题是,如果你是一家前沿实验室,真正的护城河在哪里?是在智能中,还是在工具链中?
I think the other interesting question is, where is the real moat if you’re a frontier lab? Is it in the intelligence or is it in the harness?
—— Sriram Krishnan · [07:20]

指向原始笔记的链接

思考这个等式最简单的方法是,如果你不是那个模型,你就是那个框架。
the easiest way to think about this equation is that if you’re not the model, you’re the harness.
—— 嘉宾 · [00:56]

指向原始笔记的链接

Markdown，而不是魔法。胖技能，瘦线束。
Markdown, not magic. Fat skills, thin harness.
—— Garry Tan · [16:56]

指向原始笔记的链接

如果你正在构建一个 UX，并且有一些东西与对话并排很好渲染，但它既不存在于消息列表中，也没有被机具卸载并提供发现它的方式，那么它对智能体来说本质上就是不可见的。
if you’re building a UX and you have something that’s nicely rendered side by side with the conversation, but it doesn’t exist either in the message list or offloaded by the harness with a way to discover it, it’ll be essentially invisible to the agent.
—— 嘉宾 · [08:32]

指向原始笔记的链接

当模型演进而安全带不变时,它会拖累智能体的表现。
When the model moves and the harness doesn’t, it degrades the agent.
—— Isabella Kai He · [08:57]

指向原始笔记的链接

安全带的修复程序不再被需要,并且实际上在 Opus 4.5 的模型性能中起反作用。
The harness fixes were no longer needed and were actually detracting from model performance with Opus 4.5.
—— Isabella Kai He · [08:49]

指向原始笔记的链接

harness 的主要工作是在正确的时机把上下文带给模型。
The main job of the harness is to bring context to the model at the right point in time.
—— Harrison Chase · [02:07]

指向原始笔记的链接

我的回答通常是，你越是处于模型训练内容分布之内，现成的 harness 就越好用。
My answer generally is the more in distribution you are of what the models are trained on, then the better the off-the-shelf harness will be.
—— Harrison Chase · [07:19]

指向原始笔记的链接

我认为我们经常看到，你越偏离分布，就越会想要定制 harness。
I think we often see that the more out of distribution you get, the more you’re going to want to customize the harness.
—— Harrison Chase · [21:51]

指向原始笔记的链接

如果他们正在构建正确的产品，他们将能够利用这种判断力，以便今天的判断力成为明天的智能。
And if they’re building the right product, they will be able to harness that judgment so that the judgment of today is the intelligence of tomorrow.
—— Julien Bek · [61:45]

指向原始笔记的链接

我喜欢用的一个类比是：如果模型是引擎，那 harness 本质上就是引擎周围的那辆车。
The analogy I like to use is that if the model is the engine, the harness is essentially the car around the engine.
—— Santi Garza · [03:50]

指向原始笔记的链接

我认为 harness 长期以来一直被贬低为二流研究，然而它字面上就给我们带来了 18% 的提升——这就是 harness 一和 harness 二之间的差距。
So I think harnesses have long been belittled as subpar research, yet it literally gives us an 18% bump in the difference between harness one and harness two.
—— 嘉宾 · [01:14]

指向原始笔记的链接

而仅仅加上一些 harness——这种不配称为研究的东西，只是一些包装和脚手架——我们就能达到 95%，NVIDIA 的 ABO 更是达到了 100%。
And just with some harness, this thing that doesn’t deserve any research, just some wrapper and some scaffolding, we can get to 95% and ABO from NVIDIA got to 100%.
—— 嘉宾 · [03:56]

指向原始笔记的链接

要么我们在学习系统提示词，要么我们在学习 harness 本身——这非常迷幻。
Either we’re learning the system prompt or we’re learning the harness itself, which is very trippy.
—— 嘉宾 · [14:18]

指向原始笔记的链接

从某种意义上说,harness 总是比模型领先一点。
So the harness, in a sense, is always a little bit ahead of the model.
—— Tibo Sottiaux · [36:27]

指向原始笔记的链接

我不得不和模型对抗，不让它表现得像它被训练时所用的那个 harness。
I have to fight against the model to not act like the harness that it was trained with.
—— Demetrios Brinkmann · [37:22]

指向原始笔记的链接

五年后，今天成功的公司里有七成到那时将完成从 SaaS 到 harness 的转变。
Five years from now, seven out of 10 companies today that are successful then will have made the switch from SaaS to harness.
—— Tim Sanders · [58:22]

指向原始笔记的链接

一个明显的趋势是，随着模型能力提升，我们可以移除编排代码；但如果你的 harness 在模型进步时反而变得更复杂，你很可能是在过度工程化你的 harness。
An obvious trend that with better model capabilities, we can remove orchestration code, but if your harness is getting more complex as the model improves, you are most likely overengineering your harness.
—— Philipp Schmid · [16:05]

指向原始笔记的链接

把一个智能体里的模型部分拿掉,剩下的所有东西,那就是 harness。
If you take an agent, remove the model part from it and everything that you have left, that’s the harness.
—— Mike Chambers · [04:54]

指向原始笔记的链接

Harrison Chase 说，你的 harness、你的记忆，如果你不拥有你的 harness,你就不拥有你的记忆，这是关键。
Harrison Chase said, your harness, your memory, and if you don’t own your harness, you don’t own your memory, which is key.
—— Kay Malcolm · [18:20]

指向原始笔记的链接

所以 harness 工程的目标是一次又一次地创造可靠且可预测的输出，而推理模型是非常非确定性的。
So the goal of harness engineering is to create reliable and predictable outputs over and over, whereas a reasoning model is very non-deterministic.
—— Ignacio Martinez · [11:26]

指向原始笔记的链接

如果有一个实操配方我希望你们带走，那就是：冻结测试框架，定义场景，记录轨迹，给行为打分，并且只暴露一个很小的策略面。
If there’s one practical recipe I want you to take away, freeze the harness, define scenarios, log traces, score behavior, and expose only a small policy surface.
—— Erina Karati · [19:36]

指向原始笔记的链接

它们能通过,是因为我们构建了一个不会让它出错的挽具。
They pass because we built a harness that won’t let it be wrong.
—— Andrew Orobator · [18:01]

指向原始笔记的链接

我们是在自己的计算机使用框架上训练我们的模型，所以使用模型分布内的那一套是有优势的。
we train our models on our computer use harness, so there is an advantage to using the one that’s in distribution for the model.
—— Ari · [16:31]

指向原始笔记的链接

你觉得我们会作为人类盯着这个看吗?绝对不会。但一个自主的框架、一个痴迷于 token 成本的自动化编码智能体,就会做到这一点。
Do you think we look at this as human? Absolutely not. But an autonomous harness, an automated coding agent that is obsessed with token costs is going to do just that.
—— Laurent · [06:46]

指向原始笔记的链接

第一，我们认为 harness 需要是模型中立的，也就是你不应该被耦合到某个特定的模型上。
One, we think the harness needs to be model neutral, so you shouldn’t be coupled to a particular model.
—— 嘉宾 · [04:40]

指向原始笔记的链接

慢慢地加入越来越多的逻辑，因为要在像 Claude Codes 和 Codexes 这样的东西里添加额外的 harness 逻辑非常困难，因为它们的核心循环就是一个 while 循环。
That slowly add more and more logic because it’s very hard to add additional harness logic into like the Claude codes and codexes because their core loop is just a while loop.
—— Diogo Almeida · [82:37]

指向原始笔记的链接

但我们学到的一件事是，实际上我们当时做的那些事情，大部分用一个更好的 harness 就能做得同样好，而且便宜得多。
But one of the things that we’ve learned is actually most of those things that we were doing, we could do just as well, a lot cheaper, with a better harness.
—— Woodson Martin · [28:35]

指向原始笔记的链接

但这里的要点和收获是：搜索帮你找到信息，而 harness 给你可信赖的结果。
But the idea here and the takeaway is that search helps you find information, and the harness gives you the results that you can trust.
—— George He · [22:49]

指向原始笔记的链接

而 Gemini API 里的 anti-gravity 智能体仍然是同一个 harness,意思是如果我们改进 Gemini 使它在那个 harness 下变得更好，我们在 anti-gravity 里受益，但我们在 Gemini API 里也受益。
Then the anti-gravity agent we have in the Gemini API, but it’s still the same harness, meaning if we improve Gemini to become better with that harness, we benefit inside anti-gravity, but we also benefit inside of the Gemini API.
—— Philipp Schmid · [08:51]

指向原始笔记的链接

Harness 是双手，大脑是 LLM。
Harness is the hands, the brain is the LLM.
—— Viren Baraiya · [10:02]

指向原始笔记的链接

智能体只是盒子里的智能，而你围绕它们构建的 harness 才是赋予它们能力、在下游交付这些成果的东西。
Agents are just intelligence in a box and the harness that you surround them with is what provides them with the ability to deliver these outcomes downstream.
—— Abhi Arya · [12:59]

指向原始笔记的链接

第一个是，你应该真正发布你自己版本的 harness。
The first one is you should actually ship your own version of the harness.
—— Dominik Kundel · [16:05]

指向原始笔记的链接

② 出现在这些集

65 集

③ 关联

点进去有真内容 —— 本页主要出口

智能体 · Anthropic · 沙箱 · Claude Code · 评估 · OpenAI · Codex · MCP · 护栏 · 推理

反向链接