集里怎么说它

• 《Benedict Evans:AI 价值会落在哪一层?》(27:53起):本集说算力是当下唯一的护城河:只要 scaling 持续,就应该预期最前沿模型公司的数量像半导体那样收窄;若曲线慢下来则一切都会改变。
• 《Tokens 进、互动出：推荐系统正在像 LLM 一样扩展》(00:30起):本集论证推荐系统遵循与 LLM 一样的幂律扩展曲线:加大模型、数据、算力,推荐质量可预测地提升,且行业在这条曲线上仍非常早期。

① 提到它的金句

19 条

创办一家公司从未如此容易，将其扩展为一个持久的高影响力组织从未如此艰难。
Starting a company has never been easier, scaling one into a durable high impact organization has never been harder.
—— Lenny · [26:04]

指向原始笔记的链接

任何把 GPT-5 和缩放定律联系起来的说法都是疯狂的。
Any reference to GPT-5 and scaling laws is crazy.
—— David George · [22:48]

指向原始笔记的链接

我想达到一个收入是毫无悬念、不用想的地步。因为最终，一旦你在创业公司里踏上了收入的列车，你就得开始规模化。
I wanted to get to a point where the revenue was a slam dunk and no brainer. Because ultimately, once you get on the revenue train in a startup, you got to start scaling.
—— Damien Lewke · [38:16]

指向原始笔记的链接

如果我们扩展规模，仅靠扩展规模并没有帮助。你需要增加模型的多样性。
If we scale, scaling alone doesn’t help. You need to increase the diversity of the models.
—— Damian Borth · [24:25]

指向原始笔记的链接

如果 Transformer 是有利可图的,如果你可以花费更多的努力和资源来扩大 Transformer 以在下一个季度获胜,那么很难把大量的注意力和精力投入到可能在一两年内更好或重新定义该领域的事情上。
If Transformer is profitable and if you can spend more efforts and more resources scaling Transformer to win in the next quarter, it’s very hard to put at least a lot of attention and a lot of energy to work on something that will maybe better or maybe will redefine the field in a year or two.
—— Jerry Tworek · [13:28]

指向原始笔记的链接

如果你考虑某人的技能曲线，它可能开始得很高，但如果他们没有一个良好的上升轨迹，他们将会在一家指数级扩张的公司里非常艰难地工作。
If you think about somebody’s skill curve, it might start out really high, but if they don’t have a good trajectory upwards, they are going to have a really hard time working in a company that is scaling exponentially.
—— Max Junestrand · [18:08]

指向原始笔记的链接

而且，你知道，如果你相信这些缩放定律，智能在变得更好，我会说，每年一个数量级，我是说，10 倍 10 倍 10 倍，三年就是一千倍。
And, you know, and the thing, if you believe these scaling laws and intelligence is getting better, I’d say, at one order of magnitude a year, I mean, 10 times 10 times 10, that’s a thousand times in three years.
—— Tony Kim · [26:32]

指向原始笔记的链接

所以多智能体是一种以并行而非纯串行的方式扩展测试时算力的方法。
And so multi-agent is a way of scaling test time compute in parallel instead of purely serial.
—— Noam Brown · [01:49]

指向原始笔记的链接

在 Inception，我们押注 diffusion 模型，因为我们相信最终重要的是推理时扩展。
At inception, we’re betting on diffusion models because we believe that what matters eventually will be inference time scaling.
—— Stefano Ermon · [08:01]

指向原始笔记的链接

在我脑子里,规模定律从来都不是缺陷,恰恰是 OpenRouter 会非常有价值的一个特性。
The scaling laws were never like, in my mind, were always a feature, not a bug for why OpenRouter would be very valuable.
—— Anjney Midha · [16:33]

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

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

目前看来，我们所能感知到的缩放定律似乎没有尽头。
It appears there is no end to the scaling law that we can perceive so far.
—— Parag Agrawal · [14:45]

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

它实际上变得更差了。大型实验室下的赌注是，扩展通用智能就会扩展教学能力。不幸的是，情况并非如此。
It’s actually getting worse. The bet that the big labs have made is that scaling general intelligence is going to scale pedagogical capacity. Unfortunately, that has not been the case.
—— 嘉宾 · [45:55]

指向原始笔记的链接

所以这是那种——底层现实中是非常复杂的基础设施在扩展你的系统——但只要你为它提供合适的接口，比如简单的配置文件，这些能力对任何想要扩展的公司来说都变得容易接触得多。
So this is one of those things, very complex infrastructure in reality under the hood that is scaling your systems, but you provide the right interfaces for it, like simple configuration files, and those things become much more accessible to any company trying to scale.
—— Ben · [13:55]

指向原始笔记的链接

缩放定律依然成立,甚至可能正在朝着有利的方向弯曲,因为数据质量在变好。
Scaling laws are holding or maybe even bending because the data quality is getting better.
—— 嘉宾 · [06:04]

指向原始笔记的链接

我认为我们拥有的最接近理论的东西是缩放定律，而这些是经验性定律。
I think the closest things we have are scaling laws, and these are empirical laws.
—— Daval Ajada · [09:12]

指向原始笔记的链接

那就是，说到底，扩展是行得通的。但在机器人学中，扩展以及如何扩展要难得多，因为没有互联网。
which is like, at the end of the day, scaling will work. But in robotics, to scale and how to scale just much harder because there is no internet.
—— Jason Ma · [42:38]

指向原始笔记的链接

② 出现在这些集

2 集

• 《Benedict Evans:AI 价值会落在哪一层?》 — 作为概念
• 《Tokens 进、互动出：推荐系统正在像 LLM 一样扩展》 — 作为概念

③ 关联

点进去有真内容 —— 本页主要出口

LLM · Jacob Efron · Devansh Tandon · Benedict Evans · Meta · OpenAI · Instagram · Anthropic · YouTube · TSMC