跨国深谈

LLM

43分钟阅读

跨国深谈← 返回←LLM
LL

LLM

概念
本站收录 64 集 · 54 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

54 条

因为 LLM 正在总结许多引用，所以你需要尽可能多地被提及。
Because the LLM is summarizing many citations and so you need to get mentioned as many times as possible.
—— Ethan Smith · [11:12]

指向原始笔记的链接

Webflow，我们看到 LLM 流量和 Google 搜索流量之间的转化率差异是 6 倍。
Webflow, we saw a 6X conversion rate difference between LLM traffic and Google Search traffic.
—— Ethan Smith · [14:47]

指向原始笔记的链接

我认为有很多它表现不太好的情况，但我觉得如果你把 LLM 当裁判的用途从一个基准重新定义为一个异常检测器，那它其实是可以的。
Well, I think there’s a lot of cases where it doesn’t work very well, but I think if you reframe the utility of an LLM as a judge from being a benchmark to being an anomaly detector, then I think it’s actually okay.
—— Ankur Goyal · [25:10]

指向原始笔记的链接

一个模型或者一个非常重的 LLM 产品，然后展示一堆关于性能的柱状图，这说白了，全是营销。
A model or a really LLM heavy product and showing a bunch of bar charts about performance, that is like, it’s all marketing.
—— Ankur Goyal · [55:28]

指向原始笔记的链接

当我们试图请求一个 LLM 做这个错误分析时，我们通常发现它只是说追踪看起来很好，因为它没有理解某样东西是否可能是坏的产品味道所需的上下文。
What we usually find when we try to ask an LLM to do this error analysis is it just says the trace looks good because it doesn’t have the context needed to understand whether something might be bad product smell or not.
—— Shreya Shankar · [24:09]

指向原始笔记的链接

如果你因为市场还没准备好而失败,我实际上认为那样更好。至少你尝试过某种深刻的、新颖的、困难的事情,而不是转型成为另一家 LLM 包装公司。
If you fail because the market isn’t ready yet, I actually think that’s way better. At least you took a swing at something deep, and novel, and hard instead of pivoting into another LLM wrapper company.
—— Edwin Chen · [30:13]

指向原始笔记的链接

一旦你接受了评测，你就会意识到 LLM 所做的不过是你在产出的评测的一个函数。
once you embrace evals, then you realize that what the LLM does is merely a function of the evals that you’re producing.
—— Ankur Goyal · [05:34]

指向原始笔记的链接

良好的变更管理并承认每次你例如替换智能体中的 LLM 时，它的行为都会不同。如果你不在你的治理中考虑到这一点，你的最终用户将承担其后果。
Good change management and acknowledging that every time you, for example, replace the LLM in an agent, it will behave differently. And if you don’t take that into account in your governance, your end users will bear the burden of that.
—— Emil Lassen · [20:09]

指向原始笔记的链接

LLM 不会神奇地从你的大脑中提取你的知识，除非你把那些知识变成规则，否则你不可避免地会以沮丧告终。
The LLM is not going to magically extract your knowledge from your brain and you will inevitably end up frustrated unless you turn that knowledge into rules
—— Kitsa · [04:37]

指向原始笔记的链接

LLM 就像一个从来没出过门、从未离开过自己出生的那个房间的人，但几个世纪以来每天读遍所有的书。
LLM is like someone who has never gone outside, never left the room they were born in. But read all the books every day for centuries and centuries.
—— Alex · [08:02]

指向原始笔记的链接

VLA 是一个非常糟糕的 hack：你有一把锤子，你有一个 LLM，然后你看什么都像钉子。
VLA is a very bad hack where you have a hammer, you have a LLM, and you see everything as a nail.
—— Alex · [10:31]

指向原始笔记的链接

在我看来:LLM 是发动机,而我们是车、是道路、是整个系统。
The way I see it: the LLM is the engine, and we’re the car, the roads, and the whole system.
—— Carlos González de Villaumbrosia · [16:37]

指向原始笔记的链接

人类与大语言模型的混合模式现在真的非常成功。但是没有人类的大语言模型,就不怎么行,根本不行。
The human-LLM hybrid is really, really successful right now. But LLMs without humans, not so much, not at all.
—— Jerry Tworek · [39:00]

指向原始笔记的链接

我们不盲目使用 LLM 代码的原因是我们还不信任它,因为它们底层的系统没有足够的刚性。
The reason that we don’t use LLM code blindly is because we don’t trust it yet, because the systems underneath them don’t have enough rigidity.
—— Vaibhav Gupta · [17:17]

指向原始笔记的链接

从 Demis Hassabis 往下,他基本上都把世界模型视为前进的道路,视为最重要的东西,并且他们愿意在 LLM 开发上退居二线
from Demis Hassabis on down he just basically views world models as the way forward as the most important thing and they’re willing to sort of take a back seat on llm development
—— 嘉宾 · [55:14]

指向原始笔记的链接

问题是 LLM 做不了任何事。
The problem is the LLM can’t do anything.
—— Frank Coyle · [08:53]

指向原始笔记的链接

上下文溢出到主上下文，因为上下文意味着 tokens，tokens 意味着钱，而且你有越多的上下文，LLM 在给你答案时就会越困惑。
Context spill over into the main context because context means tokens, tokens mean money, and the more context you have, the more confused the LLM is gonna be in giving you an answer.
—— Frank Coyle · [13:11]

指向原始笔记的链接

当每个人谈论智能体时，他们真正谈论的只是一个在循环中运行、调用工具的 LLM。
When everyone talks about agents, what they really talk about is just an LLM running in a loop calling tools.
—— Harrison Chase · [02:47]

指向原始笔记的链接

这导致了一些非常令人惊讶的行为，因为实际上你离能够访问 10,000 个上下文桶只有不到两次 LLM 调用的距离，其中每一个都包含 200,000 个 token。
And that leads to some really surprising behaviors because you’re literally no more than two LLM calls away from being able to access 10,000 context buckets each one of which contains 200,000 tokens.
—— Flo Crivello · [32:49]

指向原始笔记的链接

如果把我的品味排除在外，LLM 讨厌 Grock 并且它喜欢 Claude Opus。它喜欢 Sonnet 并且它不喜欢 GPT-56 Sol。
If you take my taste out of it, The LLM hates Grock and it loves Claude Opus. It loves Sonnet and it does not like GPT-56 Soul.
—— 嘉宾 · [25:08]

指向原始笔记的链接

一种在这里能运作得很好的模式是，如果你在平台中强制执行一个更宽泛的 agent 定义，把 LLM 和人类都囊括在内，那么你就可以做到让任何 LLM 能执行的动作，人类也能执行。
One pattern that can work very well here is if in your platform you enforce a wider definition of agent which encompasses both LLMs and humans, then you can make it such that any action that can be taken by an LLM could also be taken by a human.
—— Christopher Lovejoy · [13:49]

指向原始笔记的链接

其次，你有这种人类与智能体的等价性，这意味着对于任何任务，你可以让智能体——也就是 LLM 智能体——和人类都执行它，而你们的差异就是你的 eval。
Secondly, you have this human agent equivalency, which means that for any task, you could get both the agent, the LLM agent, and the human to perform it, and your difference is your eval.
—— Christopher Lovejoy · [16:16]

指向原始笔记的链接

当你使用 LLM 时你要付费，无论你是否得到了你想要的东西。
You pay when you use an LLM, regardless of whether you get what you want.
—— 嘉宾 · [15:32]

指向原始笔记的链接

如果你只围绕某个特定的 LLM 裁判去爬山，你很可能是在对它过拟合。
And if you’re only hill climbing on a specific LLM judge, you’re probably overfitting on it.
—— Vishu · [07:35]

指向原始笔记的链接

你的请求永远不应该被你的护栏耗时所束缚。决定速率的步骤应该始终是 LLM。
Your request should never be bound by your guardrail timing. It should always be the LLM that is the rate determining step.
—— Kanish Manuja · [11:21]

指向原始笔记的链接

有几项学术研究表明，最开头的 7K 和最末尾的 7K token 是最重要的，而中间的那些东西最终只会在 LLM 试图想出答案时把它搅浑、让它困惑。
There’s a couple academic studies that show that like the first 7K and like the last 7K tokens are the most important and the stuff that’s in between can end up just muddying and confusing the LLM as it’s trying to come up with an answer.
—— 嘉宾 · [56:56]

指向原始笔记的链接

我会担心任何这样的记忆架构：它不依赖成本更低的嵌入器和重排序器，而是依赖 LLM 的多遍处理来帮你分类并缩小数据语料库。
I would worry about any memory architecture that instead of relying on lower costs embedders and re-rankers is relying on multiple passes of the LLM to help you categorize and shrink the corpus of data
—— 嘉宾 · [66:35]

指向原始笔记的链接

大多数 LLM 系统本质上就是一个搜索问题。
Most LLM systems are just a search problem.
—— Jean-Denis Greze · [01:22]

指向原始笔记的链接

我认为在未来六个月内，我们会看到一堆这样的系统：公司把一项政策托付给 LLM，让它自动把越来越多的、原本属于私密的信息披露到公开空间里。
And I think in the next six months we’ll have a bunch of systems where companies have trusted an LLM with a policy to automatically surface more and more information that otherwise would have been private into a public space.
—— Jean-Denis Greze · [11:12]

指向原始笔记的链接

你给一个 LLM 一个目标，它会在护栏之内竭尽所能去解决那个目标。
You give an LLM a goal, it will do everything it can within guardrails to solve that goal.
—— 嘉宾 · [19:26]

指向原始笔记的链接

我不认为只靠一个 LLM 就能做到。它们今天还不够，我也不认为它们有足够的创造力，但它们确实非常有帮助。
I don’t think you can get there with just an LLM. They’re not enough today. I don’t think they’re creative enough either, but they’re definitely super helpful.
—— Alexander Whedon · [47:11]

指向原始笔记的链接

顺便说一句，这比大多数 LLM-as-a-judge 方法——也就是让一个 LLM 来判断那是优质的人类作品还是 AI 生成的 slop——表现都要好。
This performed better, by the way, than most LLM-as-a-judge methods of asking an LLM to judge if that is great human quality versus AI-generated slop.
—— Thais Castello Branco · [08:29]

指向原始笔记的链接

整个「LLM 套壳」或「模型套壳」的概念其实正在跑通
the whole concept of being a sort of LLM wrapper or model wrapper is actually working out
—— Aaron Levie · [02:18]

指向原始笔记的链接

从安全视角看，这是一个非常令人不安的世界，因为突然间你实际上是在一个业务环境里，运行由 LLM 在运行时生成的代码。
It’s a very unnerving world from a security lens because suddenly you are literally running code emitted at runtime from an LLM into a business environment, essentially.
—— Samar Abbas · [12:23]

指向原始笔记的链接

如果人们是在 LLM 内部消化信息，那么 LLM 就是新的网站访客。
if people are consuming information inside an LLM, the LLM is the new website visitor.
—— Liam Dunne · [07:09]

指向原始笔记的链接

我们发现如果你发布即使方向上大致正确的数字，效果会好得离谱，因为存在空白地带，而 LLM 本质上不想产生幻觉。
we found this works like insanely well if you publish numbers that are even like yards directionally correct because there’s blank space and the LLM doesn’t want to hallucinate, essentially.
—— Ben Moore · [35:02]

指向原始笔记的链接

如果某个 LLM 选中这段话并把它放回一个答案里，我们是否被放在有竞争力的位置上？
if an LLM picked this paragraph and put it back inside to an answer, are we being positioned competitively?
—— Liam Dunne · [43:07]

指向原始笔记的链接

大多数人让 VLM 或 LLM 给标注打一个分数，但这些分数完全不管用，尤其是当你的模型和裁判规模不同的时候。
most of the people ask the VLM or LLM to assign some score to it, but those scores absolutely don’t work, especially if your models are of different size with judges.
—— Merve Noyan · [06:28]

指向原始笔记的链接

在合适的地方,我们不是用 LLM 替换检索和排序系统,而是把推理和理解蒸馏进某种生产级排序架构。
Where it seems right, we’re not replacing the retrieval and ranking system with an LLM, but rather distilling the reasoning and understanding into some production ranking architecture.
—— Raghav Saboo · [09:06]

指向原始笔记的链接

在线的 LLM 调用往往不是你所需要的产品架构。
The online LLM call is often not the product architecture you need.
—— Raghav Saboo · [21:26]

指向原始笔记的链接

所以如果你尝试比较内容信息流和 AI 聊天应用,我认为 LLM 推荐器在结构上其实比 AI 聊天更省 token。
And so if you try to compare content feeds and AI chat apps, I think that LLM recommenders are actually structurally more token efficient than the AI chat.
—— Devansh Tandon · [14:46]

指向原始笔记的链接

所以对我而言，就是，好吧，为什么不直接构建一个非常好的基础 LLM，然后用它来控制机器人，而不是训练一个专门面向机器人、更依赖于纯机器人数据的模型。
And then so to me, it’s like, okay, why not just build a really good foundational LLM and then use that to control robots rather than training like a model that’s specifically for robotics and that’s more dependent on just robot data.
—— Ham · [06:52]

指向原始笔记的链接

他们开始以某种方式利用测试数据，用 LLM 作为裁判来训练模型，用各种花招来刷基准，拿季度奖金之类的。
They start to kind of use leverage test data to use LLM as a judge for training of the models, use all the various tricks to basically gain benchmarks, get like quarterly bonuses and so on.
—— Alex Mashrabov · [23:52]

指向原始笔记的链接

网页已经变成给智能体看的了，对人类来说已经不可读了，因为所有东西都是假设 LLM 会去读而写的，这太让人沮丧了。
The web has become for agents, like, it’s unreadable for humans anymore because everything’s being written, assuming an LLM is reading it and it’s so frustrating.
—— Rania Khalaf · [14:53]

指向原始笔记的链接

新的 LLM 迫切希望拥有比现在高出几个数量级的内存带宽，以便运行得更快。
a new LLM desperately wants to have orders of magnitude more bandwidth to memory in order to run faster
—— Walter Goodwin · [16:33]

指向原始笔记的链接

为什么 LLM 删掉了我的数据库？嗯，因为它没有因果性的概念。
Why did the LLM delete my database? Well, because it has no concept of causality.
—— Sam Newman · [95:59]

指向原始笔记的链接

你的 LLM 的表现取决于你喂给它们的上下文的好坏。
Your LLM are as good as the context that you feed them with.
—— Eli Cohen · [14:56]

指向原始笔记的链接

一开始可能没问题，但如果你的 LLM 不理解位置和空间推理，到了某个时候就会崩掉。
It might be fine at the start, but if your LLM doesn’t understand the location and spatial reasoning, it falls apart at some point.
—— George He · [14:58]

指向原始笔记的链接

如果你有一个 oracle 能告诉 LLM 它做的事情是对还是错,那你基本上就赢了。
If you have an oracle that can tell the LLM if what it did is correct or not, then you basically won.
—— Mario Zechner · [13:31]

指向原始笔记的链接

那是编程中唯一通常没有以文本形式编码在某个地方、因而 LLM 无法从中学习的东西,也就是构建某个东西架构的过程。
That is the part, that’s the only part in programming that is actually not usually encoded somewhere in textual form so an LLM can learn from it, that the process of building the architecture of something.
—— Mario Zechner · [16:29]

指向原始笔记的链接

你知道,我最近采用的一个说法是,我不想为别人经历 LLM 精神错乱付费。
And you know, the phrase I’ve sort of recently taken to is, I don’t want to pay for someone else to go through LLM psychosis.
—— Sean Wang · [01:23]

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

但如果你把文档完全交给 LLM 来编写和维护，研究指出，纯由 LLM 维护和编写的文档往往会随时间呈指数级退化。
But if you offload documentation purely to LLMs to author and maintain, research points out that purely LLM-maintained and written documentation tends to degrade exponentially over time.
—— Kamalakannan Nandagopal · [02:26]

指向原始笔记的链接

而我所相信的这条路径，实际上并不以基于 LLM 的产品作为中间站。
And the path that I believe in doesn’t actually go through LLM-based products as an intermediate point.
—— David Silver · [04:22]

指向原始笔记的链接

② 出现在这些集

64 集

③ 关联

点进去有真内容 —— 本页主要出口

智能体 · Anthropic · ChatGPT · OpenAI · MCP · Claude Code · Claude · Google · Lenny · 推理

反向链接