跨国深谈

大型语言模型 (llms)

15分钟阅读

跨国深谈← 返回←大型语言模型 (llms)
大型

大型语言模型 (llms)

概念
本站收录 6 集 · 37 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

37 条

我在整个职业生涯中构建过很多工作流软件，如果你只是给工作流构建器加一个 AI 步骤，那它不是智能体，而且它很可能是在极其做空和看空 LLM，我认为这个房间里没有人是这样想的。
I’ve built a lot of workflow software, you know, throughout my career, and if you’re just adding, you know, an AI step to a workflow builder, it’s not an agent, and it’s also probably incredibly short and bearish on LLMs, which I don’t think anyone in this room is.
—— 嘉宾 · [15:43]

指向原始笔记的链接

他们将成为 100 倍工程师，因为他们能够管理 LLM。
They’re going to be 100X engineers because they’re able to manage LLMs.
—— Maor Shlomo · [33:58]

指向原始笔记的链接

对我来说，与 LLM 合作的一个有争议的观点是不要使用 TypeScript，使用纯 JavaScript，使用 JSX。
Controversial take for me working with LLMs is don’t use TypeScript, use plain JavaScript, use JSX.
—— Maor Shlomo · [71:46]

指向原始笔记的链接

而且 Webflow，他们 8% 的注册来自 LLM。它现在是你最大的渠道之一，所以它很大。
And Webflow, they get 8% of those signups from LLMs. It’s now one of your top channels so it’s large.
—— Ethan Smith · [46:45]

指向原始笔记的链接

有趣的是，LLM 本身在下国际象棋方面相当糟糕。
Interestingly, LLMs themselves are quite bad at playing chess.
—— Albert Cheng · [49:46]

指向原始笔记的链接

所有这些大语言模型在晚上和周末都闲置着，而人类不在那里。
All these LLMs are sitting idle overnight and on weekends, while humans aren’t there.
—— Dhanji Prasanna · [33:38]

指向原始笔记的链接

我很快，但我快不过使用 LLM 的工程团队。
I’m fast, but I’m not faster than engineering with LLMs.
—— Kitsa · [01:04]

指向原始笔记的链接

Sam Altman 说我们会像 Windows 一样,但 LLM 没有网络效应。
Sam Altman said we’re going to be like Windows, but LLMs don’t have network effects.
—— Benedict Evans · [05:27]

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

人们把生物学视为一个定制问题或定制领域,但实际上它就像自动驾驶、LLM 的相同原理,它们都适用。
people kind of treat biology as this like bespoke problem or like bespoke field, but really it’s like the same principles as like self-driving, LLMs, they all apply.
—— Matt · [18:00]

指向原始笔记的链接

LLM 很少告诉你不,而这个技能会绝对告诉你不,别做这个。
LLMs very rarely tell you no, and this skill will tell you absolutely no, don’t do this.
—— Oji Udezue · [17:07]

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

尤其是图像和设计,LLM 有点更可替换一些,但在图像和设计上,我们真的需要掌控自己的命运。
Particularly for image and design, LLMs is a bit more sort of fungible, but for image and design we really needed to be in control of our own destiny there.
—— Cliff Obrecht · [16:46]

指向原始笔记的链接

就像我们确实有这些表现很好、能对抗企业 LLM 的开源 LLM，我认为在生物学领域也会有这些开源基础模型，它们将产生非常、非常广泛的影响。
And much like you could imagine, or we do have these open source LLMs that do very well against corporate LLMs, I think there’ll be these open source foundation models in biology that will have very, very broad impact.
—— Vijay Pande · [17:36]

指向原始笔记的链接

手写你的 LLMs.txt 是好得多的。
It is much better to write your LLMs.txt from hand.
—— Christopher Burns · [06:20]

指向原始笔记的链接

我认为 LLM 会很快比人类少犯很多这类错误。
I think the LLMs will make many fewer of these mistakes than humans pretty quickly.
—— JD · [18:18]

指向原始笔记的链接

因为 LLM 来到你的网站时，它们就像有视觉障碍的用户或人。
Because LLMs, when they come to your website, they’re like users or people with vision disabilities.
—— Liad Yosef · [19:31]

指向原始笔记的链接

你没法在 VLLM 或 SGLang 上运行这些基于扩散的 LLM。我们必须构建自己的服务引擎，而且我们能处理真实生产工作负载的很多复杂性。
You cannot run these diffusion-based LLMs on VLLM or SGLang. We had to build our own serving engine, and we can handle a lot of the complexity of real production workloads.
—— Stefano Ermon · [13:01]

指向原始笔记的链接

我愿意把我们的模型想象成,UDP 之于 LLM,TCP 之于我们的模型。
I would like to think of our model like UDP as LLMs and TCP as our models.
—— Diogo Almeida · [23:56]

指向原始笔记的链接

只要我们还在使用基于 transformer 的 LLM，这就是它们工作原理的数学本质，使用更少的上下文你就会得到更好的质量。
As long as we’re using transformer-based LLMs, that’s just like the math of how they work, and you will get better quality if you use less context.
—— Dex Horthy · [52:08]

指向原始笔记的链接

我听到很多工程师谈论如何把 LLM 变成确定性流程,而我的主张是:有很多问题是确定性算法无法解决的。
I hear a lot of engineers talk about how we turn LLMs into deterministic processes, and my contention is actually there’s a lot of problems that you can’t solve with deterministic algorithms.
—— Cody Menefee · [16:22]

指向原始笔记的链接

是啊，在某种意义上，现在的 LLM 比你能指定的真人还要强大。
Yeah, in some ways, the LLMs are more powerful than the humans you can designate now.
—— Eric Newcomer · [06:17]

指向原始笔记的链接

第一个是,推荐系统的扩展方式和 LLM 一样,而且这个领域在扩展曲线上还处于非常早期的阶段。
The first is that recommendation systems scale just like LLMs do, and that the field is very early in that scaling curve.
—— Devansh Tandon · [00:30]

指向原始笔记的链接

我们观察到的一件事是，对于 Fable 级别的 LLM，它们的延迟大约每月改善 2 倍，这是非常非常快的。
So one thing that we saw is that for Fable class LLMs, their latency is improving by around 2x per month, which is very, very fast.
—— Jay · [17:35]

指向原始笔记的链接

LLM 不再只是角色球员了。
LLMs aren’t just role players anymore.
—— Kevin Hou · [00:57]

指向原始笔记的链接

Scaling loss 起作用不只是科学概念。视频只是比 LLM 和编程晚了两三年，但很显然它最终在视频领域也会起作用，我就决定赌一把。
It’s not just a concept from science that scaling loss work. Video just takes maybe two, three years longer than LLMs and coding, but it was clear that finally scaling loss should work in video as well, and I just decided just to take a bet.
—— Alex Mashrabov · [10:50]

指向原始笔记的链接

我们恰好认为它可能比 LLM 还要大，但，你知道，大概早三到五年。
We happen to think that it could be even larger than LLMs, but, you know, probably three to five years earlier.
—— David George · [49:45]

指向原始笔记的链接

模型在编写和调用代码方面比调用工具更好，因为工具调用基本上是一种人工构造，我们必须教会 LLM 去做这件事。
Models are better at writing and calling code than calling tools because tool calling is an artificial construct, basically, that we have to teach the LLMs to do.
—— Jan Čurn · [06:28]

指向原始笔记的链接

因为说到底，这一切真正有趣的地方在于，如果说我们对 LLM 有任何了解的话，那就是它们太喜欢搜索了。
Because at the end of the day, right, what’s really funny about all this is that if we know anything about LLMs, it’s that they just love to search.
—— Dan Adler · [07:23]

指向原始笔记的链接

基本上所有强大的 LLM 都是为字符串优化的，而字符串是供人类或其他 LLM 消费的。
Basically all the string LLMs have been optimized for strings and strings are meant to be consumed by humans or other LLMs.
—— Diogo Almeida · [05:18]

指向原始笔记的链接

我关于科幻的论点，尤其是和 LLM 相关的，我觉得这往往就是如今人们说 AI 时的意思，因为它似乎是智能的最大压缩，同时却有着最少的实用性。
My argument with sci-fi is especially related to LLMs, which I think tends to be what people mean by AI these days, because it seems to be like the greatest compression of intelligence while having like the least utility.
—— Diogo Almeida · [16:08]

指向原始笔记的链接

这就是为什么围绕 LLM 的那些护栏真的不是长期来看正确的解决方案。
This is why all the guardrails around LLMs are really not long-term going to be the right solution.
—— Sam Newman · [96:53]

指向原始笔记的链接

所以 LLM 生成的输出中有 62% 是不安全或损坏的。
So 62% of the output generated by LLMs is not secure or it’s broken.
—— Eli Cohen · [02:25]

指向原始笔记的链接

不被锁定在单一实验室，意味着你解锁了 LLM 的整条帕累托前沿。
Not being locked into a single lab means you unlock the entire Pareto frontier of LLMs.
—— Davis Palmie · [08:51]

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

换句话说，如果你已经决定要把一个军用级别的、故障频发的实验品——也就是 LLM——作为能产生营收的点子从私人公司发布给大众，那你已经做出了那个决定。
In other words, if you’ve decided that you were going to release a military-grade malfunctioning experiment, which is what LLMs are, into the general public as a revenue-producing idea in private companies, you’ve already decided to do that.
—— 嘉宾 · [62:03]

指向原始笔记的链接

而且很有可能，LLM 中使用的所有这些人类数据，正在拖累一切。
And it might just be that all of this human data that’s been used in LLMs is kind of weighing things down.
—— David Silver · [13:42]

指向原始笔记的链接

② 出现在这些集

6 集

③ 关联

点进去有真内容 —— 本页主要出口

Facebook · a16z · AI · 强化学习 · Lenny · Willowryan · David George · Sophia Du · Liam Fedus · Annie Lamont

反向链接