集里怎么说它

• 《速度病：当团队 10 倍速写代码却推不出产品，怎么治》(13:25起):本集提出「状态与动作分离」，把智能体当作动作、把文档当作共享且持久的状态，以取代孤立的聊天上下文

① 提到它的金句

31 条

Bash 本质上是相当危险的，并且它能以意想不到的方式改变系统状态。但与此同时如果你必须手动批准每一条 bash 命令，作为一名工程师这超级烦人，而且你无法真正提高生产力，因为你只是在不断地批准每一条命令。
Bash is inherently pretty dangerous and it can change system state in unexpected ways. But at the same time if you have to manually approve every single bash command, it’s super annoying as an engineer, and you can’t really be productive because you’re just constantly approving every command.
—— Boris · [24:21]

指向原始笔记的链接

所以，其中一条是文档是一种失败状态。
And so, one of them was documentation is a failure state.
—— Bob Baxley · [40:23]

指向原始笔记的链接

但现在有了 AI 和我的同伴，就像是，等等，AI 使得让我自己做许多这些工作成为可能。现在，我不会在所谓的博士或最高 1%、10% 的水平上做这些。但如果我处于第 0 或第 10 百分位，它肯定可以让我即使在今天也很快达到就最先进水平而言的第 60、70 百分位。
But now with AI and my companion, it’s like, wait a second, AI makes, allows me to do many of those jobs myself. Now, I’m not going to do them at the what’s called the PhD or the highest 1%, 10% level. But if I was at the zero or 10th percentile, it can certainly get me even today very quickly up to like the 60th, 70th in terms of what the state of the art is.
—— Julie Zhuo · [15:08]

指向原始笔记的链接

但我想鼓励听众将自己视为了领导者，把人们聚集在一起实现自我实现，并且他们必须主动克服默认状态，那就是盲目、苦干和孤独。
But I want to encourage listeners to think of themselves as leaders who bring humans together to self-actualize, and that they have to actively overcome the default state, which is blind, grind, and loneliness.
—— Rachel Lockett · [101:22]

指向原始笔记的链接

你意识到,你对自己心理状态的控制比你以为的要多
You realize that you have more control than you think over your own mental state.
—— Eric Landau · [19:48]

指向原始笔记的链接

所以当我们微调更小、更笨的模型时,它们只是没那么通用,但在我们想要它们做的特定任务上,它们实际上优于那些大型、聪明、最先进的模型。
So when we fine-tune smaller, dumber models, it’s that they’re just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models.
—— Jesse Zhang · [05:34]

指向原始笔记的链接

文件系统是……可能是目前上下文管理的最先进技术,智能体将利用文件系统来存储它们可能需要参考的信息,比如便签本。
The file system is… Probably the state of the art when it comes to context management today, where agents will utilize the file system to store information that they might need to reference, such as a scratch pad.
—— 嘉宾 · [03:29]

指向原始笔记的链接

当我们创办公司时,抗体设计的最先进技术大约是 0.1% 的结合率。
When we started the company, the state of the art for antibody design was about like a 0.1% binding rate.
—— Josh · [16:37]

指向原始笔记的链接

你想要的是将智能体作为操作，将文档作为状态分离开来。
What you want is to separate the agent as the action and the doc as the state.
—— Matt Dailey · [13:44]

指向原始笔记的链接

我们发现,随着模型演变得更有能力,如果你将逐字稿和记忆状态作为一个周期性批处理过程,并使用我们所谓的做梦,它允许我们提取新的洞察和新的组织结构,这基本上会反馈并编辑记忆,从而让第二天的智能体会话自动地变得更加智能。
What we found is that as models have evolved and become more capable, if you feed the transcripts and the memory state as a periodic batch process with what we call dreaming, it allows us to extract new insights and new organized structures that essentially feedback and edit the memory as needed to make the next day’s agent sessions automatically much more intelligent.
—— Gagan Bhat · [27:45]

指向原始笔记的链接

我认为在很多情况下，一个人认为，我们很好，我们就让它保持在这种稳定状态，实际上事情正在以一种方式内爆。
I think in a lot of situations where one thinks, we’re pretty good, we’re just going to let it be in this steady state, that’s actually things are imploding in a way.
—— Roblox CEO · [07:01]

指向原始笔记的链接

仅仅是格式上的变化,就曾被观察到能让基准成绩波动 76 个百分点;而重排少样本示例的顺序,能把模型在某些基准上从接近随机(约 50%)翻转到接近最先进水平。
Formatting changes alone have been seen to swing benchmark by 76 percentage points, and reordering few-shot examples flips a model from near-random, so near 50%, to near-state-of-the-art on some benchmarks.
—— Jared Joselowitz · [11:37]

指向原始笔记的链接

你知道，我本科时使用的那台巨大的最先进的超级计算机、并行连接机器计算机，实际上比我口袋里的 Pixel 手机还要慢。
You know, the giant state-of-the-art supercomputer, parallel connection machine computer that I used in my undergraduate years is actually slower than the Pixel phone that I have in my pocket.
—— Zubin Gharemani · [41:43]

指向原始笔记的链接

我们为此确定的最强大的架构，是为每个账户都存在一个智能体，并维护该账户的持久状态。
So the architecture that we’ve landed on for this that I think is most powerful is an agent that exists for each account and maintains a persistent state of that account.
—— Everett Berry · [11:19]

指向原始笔记的链接

如果我没能力准确表达出我追求的审美，又不会用 Photoshop 或 Ableton,我就处于一种品味超出能力的状态。
And so if I don’t have the ability to articulate the exact aesthetic that I’m looking for, and if I don’t know how to use Photoshop or Ableton, I’m left in this state where my taste exceeds my capability.
—— Charlie Guo · [05:34]

指向原始笔记的链接

但我认为，通过观察真实系统如何失败——包括我们自己的和全行业的系统——并真正试图提取出「是什么把系统踢进了这种状态」，我们学到的东西多得多。
But I think we’ve learned a huge amount more by looking at how real systems fail, both you know, our own and systems across the industry, and really trying to extract like, what was it that kicked the system into this state?
—— Marc Brooker · [12:30]

指向原始笔记的链接

如果你有一个产品组织却本质上无法应对非常尖刺型人才，我认为那是一种失败状态。
I view it as a failure state if you have product orgs that can’t essentially handle really spiky talent.
—— Peter Sellis · [13:12]

指向原始笔记的链接

所以未来将走向一种状态:你甚至可能不必用英语定义任务,而是更多定义目标和评分标准,然后智能体会用它所有可用的上下文真正解决任务。
So the future will move towards a state where you actually might not even have to define the task in English, but actually more the goal and a scoring rubric, and then the agent will use all available contexts available to it to actually solve the task.
—— Jerry Liu · [04:57]

指向原始笔记的链接

所以在稳态下，一个更大的语言模型大概总是会胜过更小的模型，但小模型用 LoRA 微调足够便宜，以至于在足够窄的语料库上，你也可以非常快地微调一个小模型，让它在个性化语料库上变得更好，并为特定任务对模型进行个性化。
And so a larger language model will probably always outperform a smaller one at steady state, but the smaller ones are cheap enough to fine tune using LoRa that on narrow enough corpuses, you can very quickly also fine tune a small model on a narrow corpus and make it better at individualized corpuses and personalize the model for specific tasks.
—— Hamad Bashir · [32:22]

指向原始笔记的链接

它找到目标后,就能访问让无人机跟踪跟随的工具——这不需要对规则做任何专门编码,而是用更智能体的方式:给智能体它需要的所有工具,让它根据无人机状态和可用的信息与上下文自己做决策。
It finds something, then it has access to the tools that allow the drone to track and follow, and that’s without any specific coding of that rules, but instead having a more sort of agentic, giving the agents all the tools that it needs to be able to understand the drone state and make decisions given the information and the context that’s available there.
—— Suchet Bargoti · [17:21]

指向原始笔记的链接

我们意外地创造了一个机器人领域最先进的模型——仅仅通过扩展视频模型。
we accidentally created a state of the art model for robotics. By just scaling video models.
—— Anastasis · [56:05]

指向原始笔记的链接

最优做法可能是把一个新的状态-动作-奖励放回上下文里，因为这非常快，但之后就需要某种类似睡一会儿的东西，几乎所有有智能的东西都会睡觉，告诉我一个不以某种方式睡觉的智能系统，对吧，然后……
The optimal thing to do may be to put a new SAR state action reward back into context because it’s very quick, but then there needs to be some like of go to sleep for a while, like almost everything that is intelligent sleeps, like tell me an intelligent system that doesn’t sleep, right, in some way, and then…
—— Francois · [27:46]

指向原始笔记的链接

所以它不是知道科学的状态，而是知道如何做科学。
So rather than knowing the state of science, it knows how to do science.
—— Doj Chubuk · [03:09]

指向原始笔记的链接

我知道我是我，我知道我早餐吃了什么，我知道自己总有一天会死，所有这些都是意识，但如果你没有那些，就像在那个状态中，无论剩下的是什么，显然都是高度有意识的。
The fact that I know I’m me, I know what I had for breakfast, I know I’m going to die one day, all of that is consciousness, but you can also, if you don’t have that, like in that state, whatever remains is clearly highly conscious.
—— Dr. Christof Koch · [26:24]

指向原始笔记的链接

目前它位于集群之间，我们用的是线性可插拔光模块，但我认为未来五年这些光学元件会越来越贴近 GPU，走向共封装光学的状态。
At the moment it’s between the clusters and we’re using linear pluggable optics, but I think in the next five years we will see these optical components get closer and closer to the actual GPU and we will go towards a state of co-packaged optics.
—— Adam · [09:48]

指向原始笔记的链接

比如我的很多创造力会因为线程处理时的分心而被压制。所以你越能让我保持某种心流状态，就越好。
Like a lot of my creativity gets stamped down by being distracted when my thread is processing. So the more you can keep me… Sort of like in flow state, the better.
—— 嘉宾 · [29:22]

指向原始笔记的链接

虽然说出来很简单，我不知道在当时，也许甚至到现在，还有哪个任何人展示过的通用模型，能在几个不同的任务上达到 99% 以上的成功率。
As simple as it is to state, I don’t know of at the time, and maybe even still, another like general model that anybody had shown could achieve like 99% plus success rates on several different tasks.
—— Pete Florence · [10:36]

指向原始笔记的链接

但提示词工程是你觉得「哦，我必须用这句咒语」的那种东西。我不认为我们现在还处于那种状态。
But it’s prompt engineering where you see, oh, I’ve got to use this magic phrase. I don’t think we’re in that state anymore.
—— Gabriel Chua · [30:31]

指向原始笔记的链接

我把那本质上视为引擎的一种失败状态，如果我到了出现 KV cache 抖动或者东西被刷到磁盘的地步，那说明我把引擎部署得不对。
I think of that as essentially like a failure state of the engine that we, like I’ve deployed my engine improperly if I get to the case where I have KV cache thrashing or like things being flushed to disk.
—— Charles Frye · [45:46]

指向原始笔记的链接

KVCache 有点像到处使用全局变量，但如果你能对自己的状态聪明一些呢？
The KVCache is a little bit similar to having global variables everywhere, but what if you could be smart about your state?
—— Diogo Almeida · [80:48]

指向原始笔记的链接

如果是一个所谓的「愚蠢」任务，有时候用一个更笨或更小的模型反而更好，而不是用那些非常昂贵、非常臃肿的最先进模型。
If it’s a quote-unquote dumb task, it’s sometimes better to use a dumber or a smaller model as opposed to these state-of-the-art models that are really expensive and really bulky.
—— Ben · [07:18]

指向原始笔记的链接

② 出现在这些集

1 集

• 《速度病：当团队 10 倍速写代码却推不出产品，怎么治》 — 作为概念

③ 关联

点进去有真内容 —— 本页主要出口

Matt Dailey · REF · 速度病 · 智能体 · 决策层 · 上下文工程 · 原型重力 · 智能体破产 · IDE · 计划模式