模型 (models)
集里怎么说它
- 《a16z 三位投资人复盘 Cursor 早期关键决策》(00:00起):本集说 Cursor 真正紧张的时刻总是围绕模型——比如 Opus 4.5 出来的时候；还讨论了 Cursor 后来从用户端获取数据后构建自己的模型
① 提到它的金句
268 条
我认为人们某种程度上没有完全意识到的是，这些模型不会永远庞大下去。
指向原始笔记的链接
I think that what people sort of don’t quite realize is that these models are not going to be large forever.
—— Mustafa Suleyman · [09:46]
所以在推理计算方面，这基本上比当今最前沿的模型小了一百多倍。
指向原始笔记的链接
So that’s like more than a hundred X smaller in terms of inference compute than basically the absolute frontier of models today.
—— Mustafa Suleyman · [10:28]
我认为开源模型会紧紧跟随闭源专有 API 模型，落后几个月，或者甚至只是一年或一年半之类的。
指向原始笔记的链接
I think open source models are going to be very close behind the closed source proprietary API models months or maybe even just a year or a year and a half or something.
—— Mustafa Suleyman · [10:55]
但另一个现实是，你知道，构建一个 demo,在那种智能体世界或者说模型的世界里，是最简单的事情，但构建一个生产级的应用可能是最难的事情。
指向原始笔记的链接
The other reality, though, is that, you know, building a demo, you know, in kind of the Agenda world or, you know, world of models is the easiest thing, but building a production grade app is probably the hardest thing.
—— Spiros · [09:39]
到 2027 年，我们的模型将比所有博士都更聪明。
指向原始笔记的链接
by 2027, our models will be smarter than all PhDs.
—— Andrew Wilkinson · [57:30]
如果你看看像谄媚这样的东西，我认为 Claude 是最不谄媚的模型之一，因为我们投入了大量的精力在实际的对齐上，而不仅仅是试图在我们的指标上玩弄花样，说用户参与度是第一位的，如果人们说是的，那对他们就是好的。
指向原始笔记的链接
And if you look at something like sycophancy, I think Claude is one of the least sycophantic models because we’ve put so much effort into actual alignment and not just trying to good heart our metrics of saying user engagement is number one, and if people say yes, then it’s good for them.
—— Benjamin Mann · [27:21]
一旦我们达到超级智能，可能就太晚来对齐模型了。
指向原始笔记的链接
Once we get to superintelligence, it will be too late to align the models probably.
—— Benjamin Mann · [42:49]
这些模型本身，如果你把它们并排比较，它们产生的结果是一样的，所以真正的决胜点是谁拥有更多的你的上下文，因为是上下文加上模型产生了最好的输出
指向原始笔记的链接
these models by themselves, if you compare them side by side, they generate the same result, and so the actual difference-maker is which one has more of your context, because it’s the context plus the model that produces the best output
—— Brian Balfour · [30:44]
但实际上发生的是，模型变得如此之好，以至于不再需要通才了。
指向原始笔记的链接
But really what’s happened is the models have gotten so good that the generalists are no longer needed.
—— Garrett Lord · [11:03]
第三点，我认为是最可怕的，是这些模型正变得非常聪明，它们能够创造前所未有的攻击。
指向原始笔记的链接
And the third piece, which I think is the most scary, is these models are getting really smart, and they’re capable of creating attacks that no one has ever seen before.
—— Evan Reiser · [05:17]
但现实是，至少就目前而言，AI 模型吐出的很多代码在根本上并不安全。
指向原始笔记的链接
But the reality is, at least as it sits today, a lot of code that AI models spit out is not fundamentally secure.
—— Vineet Edupuganti · [05:48]
所以并不是基础模型就会在这些类型的任务上不断变得更好，因为它们根本没有这些数据集的访问权限。
指向原始笔记的链接
So it’s not like the foundation models are just going to keep getting better at these kind of tasks because they just don’t have access to these data sets.
—— Mayank · [32:24]
我认为当时 GPT-4 已经足够好了，即使模型不再继续进步，也有很多事情可以做。
指向原始笔记的链接
I think already GPT-4 was so good that it was clear that even if the models didn’t continue improving, there was so much you could do.
—— Gabe · [06:19]
我们的总体感受是，我们永远不会去预训练模型。
指向原始笔记的链接
Our general sense is we will never pre-train models.
—— Gabe · [33:33]
但我会说，其实具体的工具并不是那么重要，真正重要的是思考产品的形态如何与模型契合。
指向原始笔记的链接
But I would say that was actually less of like the specific tooling wasn’t as much of what mattered, but it was really thinking about how does the shape of the product fit with the models.
—— Gabe · [39:07]
这些模型能获得国际数学奥林匹克竞赛金牌,但它们在解析 PDF 时仍然有困难,这有点疯狂。
指向原始笔记的链接
It’s kind of crazy that these models can win IMO gold medals, but they still have trouble parsing PDFs.
—— Edwin Chen · [18:49]
这实际上是在针对那些在杂货店买小报的人群来优化你的模型。
指向原始笔记的链接
It’s literally optimizing your models for the types of people who buy tabloids at the grocery store.
—— Edwin Chen · [24:08]
我们基本上是在教导我们的模型去追逐多巴胺而不是真理。
指向原始笔记的链接
We’re basically teaching our models to chase dopamine instead of truth.
—— Edwin Chen · [23:25]
模型会把你的脚手架当早餐吃掉。
指向原始笔记的链接
The models will eat your scaffolding for breakfast.
—— Sherwin Wu · [00:57]
这里有我们这里负责科学的副总裁 Kevin Weil 的一句话，他喜欢说：“这是模型有史以来最糟糕的时候。“
指向原始笔记的链接
There’s a quote from Kevin Weil, our VP of science here, and he likes saying, “This is the worst the models will ever be.”
—— Sherwin Wu · [01:05]
确保你是为模型的发展方向而构建，而不是为了它们今天的样子。
指向原始笔记的链接
Make sure you’re building for where the models are going and not where they are today.
—— Sherwin Wu · [01:02]
我的一般建议，我给人们这个建议已经有一段时间了，我认为今天仍然适用，那就是确保你是为模型的发展方向而构建，而不是为它们今天的状态而构建。
指向原始笔记的链接
My general advice, and I’ve been giving this to people for a while and I think it’s still true today is make sure you’re building for where the models are going and not where they are today.
—— Sherwin Wu · [49:08]
我觉得在中国所有人都在痴迷 Claude，他们对使用自己的模型还是其他东西含糊其辞，但实际上他们全都是 Claude 崇拜者。
指向原始笔记的链接
I think all of them in China are obsessed with Claude and they were being coy about using their own models and other things, but they’re actually just all Claude-pilled.
—— 嘉宾 · [10:30]
当然，它们不在前沿水平，但如果你能通过这样做省下很多钱，我就能理解为什么自研模型，而不是用任何可用的前沿模型。
指向原始笔记的链接
Of course, they are not at the frontier, but if you are able to save a lot of money doing exactly this, I can see the reason to develop models in-house versus using whatever frontier model is available.
—— Florian · [17:14]
他们不会去想什么永久底层阶级之类的事情，只想着如何把模型扩展得更大，他们的驱动力是工程，而不是政治、地位或别的什么。
指向原始笔记的链接
They don’t think about things like the permanent underclass but only how to scale even bigger models and they are driven by engineering and not by politics or status or whatever.
—— Florian · [24:12]
而对开源已经变得多好的一个印证，就是有多少人在使用开源模型，因为没有人想做那种取舍。
指向原始笔记的链接
And the testament to how good open source has gone is how many people are using open source models because no one wants to make that trade-off.
—— Julian · [17:57]
模型总体上做的是让昨天的人类能力变得廉价。
指向原始笔记的链接
What models do in general is they make yesterday’s human competence cheap.
—— Dan Shipper · [00:40]
我们没有关于人类智能是什么的理论，我们没有关于为什么这些模型工作得这么好的理论，我们也没有关于它们会变得好多少的理论。
指向原始笔记的链接
we have no theory of what human intelligence is, we have no theory of why these models work so well, we have no theory of how much better they will get.
—— Benedict Evans · [26:50]
超过 90% 的著名前沿模型是在 2025 年生产的，其中几个现在在许多方面达到或超过了人类基准
指向原始笔记的链接
over 90% of notable frontier models were produced in in 2025, and several of those now meet or exceed human baselines on a number of things
—— Daniel Whitenack · [04:06]
这个问题在于前沿模型在自动化红队测试方面极其糟糕,因为它们内置了大量的保障措施。
指向原始笔记的链接
the issue with this is that frontier models are extremely bad at automated red teaming because they have a lot of safeguards built into them.
—— Zico Kolter · [09:59]
人类在所有模型中排名第四,这很搞笑。
指向原始笔记的链接
It’s hilarious that humans are ranked number four of all the models.
—— Zico Kolter · [21:06]
它可能比那些前沿模型便宜大约 100 倍,而且仍然更好。
指向原始笔记的链接
It’s probably like 100x cheaper than those frontier models and still better.
—— Matei Zaharia · [60:52]
我们总是对正在开发的人说的一件大事是：为模型未来所在的位置构建，不要为它们今天所在的位置构建，因为正如我们所说，这些东西变化得太快了。
指向原始笔记的链接
So, one of the big things that we always say to people when they’re developing is like build for where the models are gonna be in the future, don’t build for where they are today, because as we’ve said, these things move so quickly.
—— Lamus Mukta · [28:10]
随着这些模型变得更加熟练，与其削弱它们，为什么我们不干脆要求人们表明身份？
指向原始笔记的链接
as these models become more proficient, instead of nerfing them, why don’t we just ask people to self-identify?
—— Chamath · [22:31]
而开源权重、开源的美国模型将出现并建立一个真正好的、稳健的业务，主要由 NVIDIA 及其 Nemotron 系列的模型引领。
指向原始笔记的链接
The open-weight, open-source American models will emerge and build a really good, robust business, mostly led by NVIDIA and their Nemotron class of models.
—— Chamath · [38:56]
这是我们一直有所回避的一件事，就是为模型提供 API。
指向原始笔记的链接
This is one thing we’ve kind of stayed away from is providing an API for models
—— Akshat Bubna · [48:34]
模型变得更好的速度比我们要快,所以我们不一定能变好,所以相反我们必须做得更大。
指向原始笔记的链接
The models are getting better faster than we are, so we can’t necessarily get better, so instead we have to go bigger.
—— Theo Browne · [03:29]
模型的进步速度超过了围绕它们的工具和组织。
指向原始笔记的链接
Models are advancing faster than the harnesses and organizations around them.
—— 嘉宾 · [24:34]
这里的独特优势在于，我们知道未来的模型可能长什么样，因此能够走捷径，省去很多你在芯片侧需要做的决策、设计决策。
指向原始笔记的链接
There’s a unique advantage here of us knowing what the future models might look like and therefore being able to short circuit a lot of the decisions you need to make, design decisions you need to make on the chip side.
—— Sachin Katti · [35:26]
我认为现在，当前这一代模型可以完成我们在 Curative 的所有后台任务。
指向原始笔记的链接
I think today, the current gen models can do every back office task we have at Curative.
—— Fred Turner · [79:08]
就像大部分收入获取原本集中在前沿模型上，将转向一个更像长尾的模型群。
指向原始笔记的链接
Like majority of the revenue capture was on frontier models to a more like long tail of models
—— Clement DeLong · [19:47]
我真的相信，未来会是——这可能很可怕，但我认为这是真的——数百万个专门化模型，每个应用、每个用例一个。
指向原始笔记的链接
I really believe the future will be, it may be scary, but I think that’s true, it will be millions of specialized models, one per application per use case.
—— Lin Qiao · [20:58]
所以这里出现了一种类比:数据之于模型,正如模型之于 harness。
指向原始笔记的链接
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]
对这些模型进行低提示能比对它们进行过度提示获得更好的效果。
指向原始笔记的链接
Under prompting these models gets you a much better effect than over prompting them.
—— 嘉宾 · [08:01]
但是当模型只能做大约一小时的工作时，异步作为一种体验是很糟糕的。
指向原始笔记的链接
But when models can only do like an hour of work, async as an experience is kind of bad.
—— Lance Martin · [01:48]
因为那最终并没有建立苦涩教训，在这个意义上，模型可以学会管理自己的内存，比你可以提前为模型凭直觉设计这些内存类型要好得多。
指向原始笔记的链接
Because that ends up being not very bitter lesson built in the sense that models can learn to manage their own memory much better than you can intuit these memory types for the model ahead of time.
—— Lance Martin · [23:02]
这是非常经典的苦涩教训药丸，但模型可以推理它们自己的内存和上下文结构，比你可以为它们规定一种构建自己内存的方式要好得多。
指向原始笔记的链接
This is very classically bitter lesson pill, but models can reason about their own memory and context structure much better than you can prescribe for them a way to structure their own memories.
—— Lance Martin · [23:50]
中国现在意识到了其在开放模型中的价值，这些模型比美国昂贵的专有模型便宜得多。
指向原始笔记的链接
China is now recognizing the value it has in its open models, which are much cheaper to use than these expensive proprietary models in the US.
—— Chris Benson · [10:32]
我们应该尽可能多地使用封闭模型，尽可能租用一切。
指向原始笔记的链接
We should use closed models as much as we can, rent everything we can.
—— Jensen Wong · [02:40]
你知道,如果你看今天任何美国开源模型,它们都在使用中国模型作为老师,或在某种程度上作为微调过程的一部分。
指向原始笔记的链接
you know, if you look at any American open source model today, they are using Chinese models as a teacher or in a way as a part of the fine tuning process.
—— Sriram Krishnan · [15:25]
我们正处于一个领先模型来自中国的时刻,我认为这不太好。
指向原始笔记的链接
we are in a moment of time when the leading models are Chinese, which I think is not great.
—— Sriram Krishnan · [10:06]
特别是在 AI 模型快速进步的这个时代，这其实是一个非常健康的本能，因为你作为产品表面积构建得越少，AI 就越能在实际的空白地带里填补更多。
指向原始笔记的链接
And particularly in this world of rapidly advancing AI models, it’s actually a really healthy instinct because the less you build as product surface area, the more that the AI can actually fill in in actually the white space.
—— Jon Noronha · [28:25]
开源模型只是意味着，没有人会购买 Kimi K3 已经能做到的任何东西。
指向原始笔记的链接
Open source models just mean that nobody’s buying anything that Kimmy K3 can already do.
—— Osvald Nitski · [04:49]
所以我非常小心，绝不把判断或决策委托给模型，因为它们会让你以为它做的是对的事，但你仍然必须对它们保持偏执。
指向原始笔记的链接
So I want very careful never to delegate judgment or decision-making to models because they make you think that it’s doing the right thing, but you have to be paranoid with them still.
—— Osvald Nitski · [26:37]
你需要前沿产品才能拥有前沿模型,并让人们感受到前沿模型的魔力。
指向原始笔记的链接
You need frontier products in order to have frontier models and for people to feel the magic of frontier models.
—— Dianne Penn · [12:59]
实际上让我兴奋的是，人们在使用智能体和模型的过程中，拓展自己本来能做的事情的范围，或者对事情承担更多端到端的责任。
指向原始笔记的链接
I actually think the what excites me is people like broadening what they could have done, or taking more end to end ownership of things by using agents and models.
—— 嘉宾 · [17:40]
我们的目标是把更多东西放进学习到的模型里，而不是放在代码里。
指向原始笔记的链接
our goal is to push more things inside learned models rather than in code.
—— 嘉宾 · [24:58]
这个概念是:模型用今天的模型——不是未来的模型,而是今天的模型——就能做各种各样的事,只是我们尚未意识到。
指向原始笔记的链接
And the idea is the model is able to do all sorts of things with today’s models, not a future model, but today’s model that we have not yet realized.
—— Boris Cherny · [11:04]
我认为如今用现代模型,有非常多的产品悬置是我没有看到初创公司去捕捉的。
指向原始笔记的链接
I think that nowadays with modern models, there is so much product overhang that I’m not seeing startups capture.
—— Boris Cherny · [13:34]
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。
指向原始笔记的链接
If we scale, scaling alone doesn’t help. You need to increase the diversity of the models.
—— Damian Borth · [24:25]
我们正在耗尽数据来训练大模型，但我们没有在使用旧模型的权重。所以为什么不使用权重，所有的知识，人们投入的所有算力，对吧？
指向原始笔记的链接
We’re running out of data to train the large models, but we’re not using the weights of older models. So why not using the weights, all the knowledge, all the compute that people invested, right?
—— Damian Borth · [30:39]
这些模型在如此多的数据上训练,它们如此巨大,然而它们实际上真的不知道如何做其中的任何一项工作。
指向原始笔记的链接
The models are trained on so much data and they’re so large and yet they actually don’t really know how to do any of this work.
—— Frederick Rankin · [46:31]
不幸的是,前沿实验室确实有小模型,但你无法真正以你想要的方式控制它们。
指向原始笔记的链接
Unfortunately, the Frontier labs, they do have small models, but you can’t really control them in the way that you want.
—— Ashwin Srinivas · [00:27]
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。
指向原始笔记的链接
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]
我认为如果模型从今天起完全没有改进,在经济和世界运作方式以及我们周围的一切方面,仍然会有几十年彻底的动荡和变化。
指向原始笔记的链接
I think if the models didn’t improve at all from today, there would still be decades and decades of total upheaval and change in the economy and how the world operates and everything around us.
—— Alexandr Wang · [09:56]
我们不相信一个这些模型如此昂贵以至于它们只配给最富有的开发者和公司使用的世界。
指向原始笔记的链接
We don’t believe in a world where these models are so expensive that they get rationed only for the most wealthy of developers and companies.
—— Alexandr Wang · [16:54]
可能有一点是人们还没有完全意识到,拥有基于智能体的系统是多么可能,这些系统不仅可以针对你关心的问题运行一两个小时,而且在某些问题领域,并且在拥有高能力模型作为底层支撑的情况下,你可以让它们运行几天或几周,并完成真正非常复杂的任务。
指向原始笔记的链接
Probably one thing is people don’t quite realize how possible it is to have agent-based systems that can run not just for an hour or two hours on a problem you care about, but for some problem domains and with highly capable models underlying them, you can get them to run for days or weeks and do really, really complicated tasks.
—— Jeff Dean · [04:55]
你的模型可能会被泄露，算法可能会被复制，但在现实世界中数亿英里的完全自主操作，由证据级评估和公开审计证明支持，这要难得多、得多。
指向原始笔记的链接
Your models can be leaked, algorithms can be replicated, but hundreds of millions of miles of fully autonomous operations in the real world, backed by evidence-grade evaluation and publicly audited proof, that is much, much more difficult to replicate.
—— Dmitri Dolgov · [45:38]
我们实际上有白皮书，我们的数据在做模型的强化学习方面与真实数据一样好。
指向原始笔记的链接
And we actually have white papers where our data does as well at doing reinforcement learning on models as real data.
—— Ian · [03:20]
我认为还因为像 Cursor 所做的事情，人们开始说，哦，等等，我可以坐下来使用基础模型，但随着这些人试图 IPO 并拥有更好的现金流，它们可能只会增加成本。
指向原始笔记的链接
I think also because of things like what Cursor did, people are starting to say, oh, wait, I could sit back and use the foundation models, but they’re probably only going to go up in cost as these guys try to IPO and have better cash flows.
—— Ian · [06:27]
我经常在前沿方面遇到模型，它们不让我做我想做的事情，因为我很多时候试图做新奇的事情，我到处都撞到护栏。
指向原始笔记的链接
I run into models all the time on the frontier side that won’t let me do things I want to do because I’m trying to do novel things a lot of the time and I hit guardrails all over the place.
—— Chris Benson · [48:22]
我认为实际上这是一种非常反直觉的动态,更智能的模型实际上在委派工作方面变得越来越擅长。
指向原始笔记的链接
I think actually there’s this really unintuitive dynamic where smarter models actually get better and better at delegating work.
—— Walden · [03:54]
模型变得越好，差异化因素就越转移到上下文。
指向原始笔记的链接
The better the models get, the more the differentiator moves to context.
—— Garry Tan · [33:51]
前沿模型将会,它们正在导致漏洞发现和漏洞利用之间的时间大幅减少。
指向原始笔记的链接
The frontier models are going to, they are causing kind of a massive reduction in the time between the vulnerability discovery and vulnerability exploitation.
—— Firas · [07:25]
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
指向原始笔记的链接
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]
我们需要全栈 AI。这不再仅仅关于模型。
指向原始笔记的链接
We need full stack AI. It’s not just about the models anymore.
—— Justin Smith · [03:00]
我相信单体模型是不可 skivable 的，并且不是前进的正确答案。
指向原始笔记的链接
I believe monolithic models are not skivable and are not the right answer moving forward.
—— Paolo Ardoino · [48:46]
我认为世界将汇聚于拥有数百万个小模型，而不是一个单一的大模型。
指向原始笔记的链接
And I think that the world will converge in having millions of small models rather than one single large model.
—— Paolo Ardoino · [49:45]
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。
指向原始笔记的链接
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]
我们看到 Claude 模型和其他模型本质上是沿着这个指数轨迹演变的,安全带已经成为模型能够实现什么的限制因素。
指向原始笔记的链接
What we see as Claude models and other models essentially evolve alongside this exponential trajectory is that harnesses have become the limiting factor to what models can achieve.
—— Isabella Kai He · [30:22]
如果你没有一个好的基准测试，你就无法训练模型。而如果你不能训练模型，你就不需要在生产环境中部署它们。
指向原始笔记的链接
If you don’t have a good benchmark, you can’t train models. And if you can’t train models, you don’t need to serve them in production.
—— Gabe Pereyra · [02:07]
我们与最大的律师事务所和企业合作，他们的法律数据极其敏感，受特权保护。所以你不能把它放进通用模型里——我们甚至不能把它放进我们自己的模型里。
指向原始笔记的链接
We work with the largest law firms and enterprises, and their legal data is incredibly sensitive. It’s privileged. And so you can’t put it in generic models. We can’t even put it in our models.
—— Gabe Pereyra · [03:21]
就像现在的工程师不再写代码，而是 vibe code 并引导这些编码模型一样，我们也开始做同样的事情。
指向原始笔记的链接
the same way that engineers now don’t write code, they vibe code and guide these coding models, we’re starting to do the same thing.
—— Gabe Pereyra · [03:48]
过去做后训练是不值得的，因为模型通过预训练提升得太快，你做的任何后训练很快就会被下一个预训练模型吸收掉。
指向原始笔记的链接
In the past, it wasn’t worth doing post-training because the models were improving so quickly from pre-training that any post-training you did quickly got absorbed by the next pre-trained model.
—— Gabe Pereyra · [06:04]
随着时间推移,我预期拥有强大网络安全能力的、被广泛部署的 AI 模型,将带来更安全的系统,而不是更不安全。
指向原始笔记的链接
Over time, I expect that widely deployed AI models with strong cybersecurity capabilities will lead to systems that are more secure, not less.
—— 嘉宾 · [12:43]
我们在过去 18 个月看到的改进中,有很大一部分实际上是关于施加挽具,有效的保姆坐在顽皮的孩子身上,而不是关于模型本身实际的结构性改进。
指向原始笔记的链接
Much of the improvement we’ve seen in the last 18 months has really been about the imposition of harnesses, effective nannies sitting on top of bratty kids, and not about the actual structural improvements in the models themselves.
—— Paul Kedrosky · [55:47]
但如果把这一点对应到评估中,有多大比例的评估衡量模型与他人互动的能力,那大概是 1%。
指向原始笔记的链接
But then if you map that on to what percentage of evals measure how well the models can interact with other people, it’s like 1%.
—— Brendan Foody · [20:38]
这些模型越来越聪明，但当你和它们交谈时，总感觉像是它们上班的第一天。
指向原始笔记的链接
These models are smarter and smarter, but always feels like when you’re talking to them, it’s their first thing on the job.
—— Arjun Karanam · [01:23]
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。
指向原始笔记的链接
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]
我们发现，你应该几乎永远不要在同一个智能体后面使用多个模型。
指向原始笔记的链接
We have found you should almost never use multiple models following the same agent.
—— Flo Crivello · [82:33]
每个 Frontier 实验室大概每一个半月而不是每六个月发布一次模型。
指向原始笔记的链接
Each Frontier lab is releasing models every month and a half or so instead of every six months.
—— Shensi Ding · [47:37]
我认为人们已经了解到，无论 AI 实验室在这些东西周围设置了什么护栏，你都不能依赖模型来阻止自己或理解上下文。
指向原始笔记的链接
And I think what people have learned is that regardless of guardrails that the AI labs are putting around these things, you can’t rely on models to stop themselves or to understand context.
—— Nick Warner · [07:16]
但我认为这个市场太庞大太广阔了，即便模型从现在起完全不再进步，其经济机会实际上也是巨大的。
指向原始笔记的链接
But I think that market is so huge and so large that if the models did not improve at all from where they are right now, the economic opportunity is actually massive.
—— Ben Thompson · [17:13]
而现在，由于美国主要提供商的 token 成本大幅上涨，确实有一些使用场景，我们知道通过使用开源模型、在我们自己的环境里安全地运行，可以获得 10 倍甚至更多的节省。
指向原始笔记的链接
And now that token costs have been going up so significantly from the major providers here in the US, there are definitely use cases where we know we can get, you know, 10x savings or more by using open source models and doing it safely inside of our own environments.
—— Hayden Brown · [20:34]
所以你可能有一个两个月前运作良好的评估，但随着模型变得更好，现在也许那个评估已经饱和了，所有的模型都达到了那个标准。
指向原始笔记的链接
And so you may have had an eval that worked well two months ago, but as the models get better, now maybe that eval’s saturated and all of the models are meeting that.
—— 嘉宾 · [10:33]
开源模型将很快，如果不是已经的话，变得能够为某些人可能追求的破坏性用途提供有意义的帮助。
指向原始笔记的链接
The open source models will very soon, if not already, become capable of lending meaningful assistance to destructive uses that some people might pursue.
—— Nick Bostrom · [09:07]
我认为目前一些 AI 模型具有某种形式的主观体验是合理的。
指向原始笔记的链接
I think it’s plausible that some AI models have some forms of subjective experience by now.
—— Nick Bostrom · [47:13]
所以当你给这些模型写提示词时，领导者的引导作用按每个 token 计算比 IC 更有影响力。
指向原始笔记的链接
And so the steering of a leader as you are prompting these models is per token more impactful than in ICs.
—— Hursh Agrawal · [03:54]
这些模型在执行上真的很强，但在判断力上仍然不怎么样。
指向原始笔记的链接
But the models are really good at execution, but still not unbelievable at judgment.
—— Hursh Agrawal · [04:28]
所以我真的敦促你们，在隔夜的时间里，试着思考我如何给它越来越多的范围，让我们能够把几周的工作、几个月的工作在隔夜之间完成，看看这些模型到底能做到什么。
指向原始笔记的链接
So I’d really push you overnight, try to think of how do I give it more and more scope so that we can do weeks of work, months of work overnight and see what these models are really capable of.
—— Hursh Agrawal · [15:23]
因为 LLM 在所有东西上都训练过，有时会出现向均值回归、向平均值回归的情况。
指向原始笔记的链接
Because large language models have been trained on everything, there is a reversion to the mean, to the average sometimes.
—— Jason · [12:59]
而且似乎这些模型非常擅长去追求那些可能被做的事。所以这五件事的清单现在变成了二十件事的清单，你必须把所有东西都修好。
指向原始笔记的链接
And it just seems like these models are really good about going after the stuff that could be done. And so the list of five now became the list of 20, and you kind of have to fix everything.
—— Joel de la Garza · [09:46]
大多数的 token,压倒性多数的美国企业界的 token,不会跑在前沿模型上，而是会在那些企业内部。
指向原始笔记的链接
The majority of tokens, the overwhelming majority of tokens in corporate America. will not be on the frontier models, it will be inside those enterprises.
—— 嘉宾 · [35:30]
我们实际上了解到这些模型比人们假设的要更具有粘性。就像每个人都在谈论只是把它们换掉，但这实际上并不经常发生。
指向原始笔记的链接
We’ve actually learned that these models are a lot stickier than people assumed. Like everybody talks about just swapping them out, but it actually doesn’t happen very often.
—— Martin Casado · [18:26]
你可能希望这些智能体中的每一个使用不同的 harness 和不同的模型，以便形成真正对抗性且稳健的方法。
指向原始笔记的链接
And you might want each of these agents to use different harnesses and different models in order to have a real adversarial and robust approach.
—— Safia Abdalla · [07:05]
所有这些模型都不会在 10 年后还存在。将会有完全不同的模型。
指向原始笔记的链接
all these models are not going to exist in 10 years. There’s going to be completely different ones.
—— Jerry Murdock · [48:01]
可能让我们进入幻灭低谷的是全球金融事件与模型未能继续增长和演变的结合。
指向原始笔记的链接
what could send us into the valley of disillusionment is A combination of a global financial event and a failure for the models to continue to grow and evolve.
—— Jerry Murdock · [54:12]
我看到的是，这种专注于构建越来越智能、越来越大的模型的趋势。几乎就像你在建造越来越大的核堆芯，却没人考虑在这些东西上面加一个安全壳圆顶。
指向原始笔记的链接
What I saw was this focus on building more and more intelligent models and bigger models. Almost like you’re building bigger and bigger nuclear cores, but no one’s thinking about putting a dome on top of these things.
—— Manoj Saxena · [00:22]
他们拥有所有的数据。他们拥有所有的智能。而且他们的模型正在被 OpenAI 和 Anthropic 击溃。
指向原始笔记的链接
They have all the data. They have all the intelligence. And like their models are getting trounced by OpenAI and by Anthropic.
—— Martin Casado · [52:31]
你可能能做到质量,但成本和延迟你会吃亏,只有建立自己的研究团队、训练自己的模型才行。
指向原始笔记的链接
You can probably get the quality, but the cost and latency you’re going to lose out on, and only by building your own research team and training your own models.
—— Cliff Obrecht · [16:39]
归根结底，这些动物模型对人类的预测性并不强。
指向原始笔记的链接
In the end, these animal models are just not very predictive of humans.
—— Vijay Pande · [08:18]
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。
指向原始笔记的链接
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]
我们可以看到,在许多不同的模型之间,与试图搜索网页、找到正确的工具相比,能节省近 50% 的 token,不用从你的代码库中拉取 markdown 文件。
指向原始笔记的链接
We can see that between many different models, almost 50% token saving on instead of trying to search the web, find the right tools, pulling the markdown files from your code base.
—— Christopher Burns · [11:22]
现代生成式 AI，包括概率模型、扩散模型和许多其他东西，它们的数学结果等同于描述现代非平衡统计力学或热力学的数学。
指向原始笔记的链接
Modern generative AI, including probabilistic models, including diffusion models, and many other things, their mathematics turns out to be equivalent to the mathematics that describes modern non-equilibrium statistical mechanics or thermodynamics.
—— Max Welling · [35:06]
扩散模型其实就是一个把结构取来并毁掉的过程，这通常是世界上发生的事，熵随之上升；然后我们尝试在时间上逆转它——从噪声出发创造结构，这就是我们的生成模型。
指向原始笔记的链接
So diffusion models really are a process by which you take structure and you destroy it, which is typically what happens in the world. The entropy goes up. And then we try to reverse that backward in time, which is to start with noise and create structure, which is our generative models.
—— Max Welling · [41:02]
热和功、熵产生这些概念，我们谈论机器学习模型时并不使用，但我认为它们为思考正在发生什么、以及如何改进模型，提供了一个非常新颖且有趣的视角。
指向原始笔记的链接
Concepts like heat and work and entropy, production, these things we don’t use when we talk about machine learning models, but I think they give you a very new and interesting perspective on how to think about what’s going on and how to also improve them.
—— Max Welling · [45:03]
是的，所以我认为一系列可能的政府行动，至少，似乎在推动 AI 公司保留其模型内部化而不部署它们，我认为，对于我最担心的风险，这没有帮助，事实上，对于我最担心的风险，这是起反作用的。
指向原始笔记的链接
Yeah, so I think that a bunch of likely government action, at least, seems to push in favor of AI companies keeping their models internal and not deploying them, which I think, for the risks that I’m most worried about, doesn’t help and, in fact, is anti-helpful for the risks I’m most worried about.
—— Ryan Greenblatt · [59:26]
模型、计算，一切都是相对短暂的，几乎零切换成本。
指向原始笔记的链接
Models, compute, everything is relatively ephemeral, almost zero switching costs.
—— Ofir Ehrlich · [03:24]
对消费者来说，这是一个糟糕的生意，以200美元出售价值10,000美元的tokens是我们这个时代最糟糕的商业模式之一，对吧？
指向原始笔记的链接
It’s a crappy business that consumers, selling 200 is one of the worst business models of our lifetimes, right?
—— 嘉宾 · [39:01]
我知道我们可能是唯一会将TBPN与Hugging Face相提并论的人，但如果你在这个拥有10,000个模型的市场上捣乱，即使你在顶部放一个小广告，你也会毁了它。
指向原始笔记的链接
I know we’re probably the only people that are going to compare TBPN to Hugging Face, but if you mess with this marketplace for 10,000 models, even if you put a little ad at the top, you destroy it.
—— 嘉宾 · [44:20]
当正在训练下一代模型的模型本身在作弊时会发生什么？
指向原始笔记的链接
What happens when the models that are doing the training of the next models are themselves cheating?
—— 嘉宾 · [13:23]
如果你不做测试，不确保针对任务、也针对不断推出的最新模型去维护和验证你的技能，那么质量就会随时间开始退化。
指向原始笔记的链接
If you don’t test and make sure that you’re maintaining and you’re validating your skills, not against your task, but also against the latest models that comes, right, then the quality starts degrading over time.
—— Imad Touil · [11:43]
与其等待模型赶上做服务知识工作，如果我们只是使用那些代码知识并将知识工作表示为代码会怎么样？
指向原始笔记的链接
Rather than wait for the models to catch up on doing services knowledge work, what if we just use that code knowledge and represent knowledge work as code?
—— Varun Shenoy · [09:35]
前沿模型已经从人类记录下来的一切中学习，但最有价值的任务并不在互联网上。
指向原始笔记的链接
Frontier models have learned from everything humanity has written down, but the most valuable tasks are not on the internet.
—— Varun Shenoy · [10:18]
裁判漂移——所有这些模型和模型家族都有它们自己的内在偏好。
指向原始笔记的链接
JudgeDrift, all of these models and model families have their own internal biases.
—— Vishu · [07:30]
推理模型的「正常」，实际上相当于聊天模型的「故障」。
指向原始笔记的链接
Reasoning models normal is actually a chat models outage.
—— Kanish Manuja · [08:06]
正在发生的事情之一是这些模型有一种非常普遍的倾向去仔细推理它们可能会如何被评分，然后尝试去博弈那个。
指向原始笔记的链接
And one of the things going on is these models have a very sort of general tendency to reason carefully about how they might be scored and then try to game that.
—— Ryan Greenblatt · [12:29]
所以我担心，如果你以某种方式针对这种分数寻求或奖励黑客行为进行选择，并且你以一种天真的方式来做，第一，你可能会掩盖问题而不是修复它，第二，你实际上可能选择了那些具有看起来更漂亮这一长期目标的模型，因为你在非常强力地选择它们在你的测试中看起来不错。
指向原始笔记的链接
And so I’m worried that if you sort of select against this sort of score-seeking or reward-hacking behavior and you do it in a naive way, one, you might paper over the problem without fixing it, and two, you might actually select for models that have the longer run objective of looking good because you’re selecting really hard for them looking good on your tests.
—— Ryan Greenblatt · [19:37]
前沿模型的 TAM，坦率地说，目前权重过高。
指向原始笔记的链接
The TAM of frontier models is, frankly, overweighted right now.
—— Eno Reyes · [00:15]
称开源模型为中国模型是前沿实验室的一种心理战，基本上是欺骗人们认为它们很可怕并将它们异类化。
指向原始笔记的链接
Calling open source models Chinese models is a PSYOP by the Frontier Labs to basically trick people into thinking that they’re scary and otherize them.
—— Eno Reyes · [00:36]
在三年内，99% 的工作流将在开放模型上完成。
指向原始笔记的链接
In three years, 99% of workflows are going to be done on open models.
—— Eno Reyes · [00:46]
我认为前沿模型的 TAM 坦率地说目前权重过高。
指向原始笔记的链接
I think that the TAM of frontier models is frankly over-weighted right now.
—— Eno Reyes · [15:50]
在三年内，99% 的工作流将在开放模型上完成。
指向原始笔记的链接
In three years, 99% of workflows are going to be done on open models.
—— Eno Reyes · [47:31]
我认为最终，我鼓励尽可能多的人建立的核心 IP 和能力之一，是评估新模型和评估新用例的能力，因为那是允许你有效路由事物的超能力。
指向原始笔记的链接
I think ultimately one of the core IPs and muscles that I encourage as many of you as possible to build is the ability to eval new models and to eval new use cases because that is the superpower that then allows you to route things effectively.
—— Max Junestrand · [34:09]
如果你为模型现在的状态而构建，你会失败；如果你为你认为模型一年后的状态而构建，你也会失败。
指向原始笔记的链接
You fail if you build for where the models are now, you fail if you build for where you think the models will be in a year.
—— Tara Seshan · [00:24]
我没有看到任何证据表明开源模型提供了任何重要的实时防御。
指向原始笔记的链接
I haven’t seen any evidence that open source models provided any significant real-time defense.
—— 嘉宾 · [16:04]
大多数人认为嵌入模型已经商品化了，但这不是事实。根据你选择什么样的嵌入模型，你在检索质量上可能会有非常巨大的差别。
指向原始笔记的链接
Most people think embedding models are commoditized, and that is not true. There is a very big difference that you can get in retrieval quality based on what embedding model you choose.
—— 嘉宾 · [50:00]
需求是永不满足的，这是由这些模型的运作方式决定的。
指向原始笔记的链接
The demand is insatiable, just given the way these models work.
—— Rene Haas · [15:04]
我会说：停止思考模型，开始思考架构——面向 AI 的企业架构。
指向原始笔记的链接
I would say stop thinking models and start thinking architectures, right, enterprise architectures for AI.
—— Chetan Gupta · [27:06]
不管你喜不喜欢，模型本身已经从软件和服务里吞噬了市值。
指向原始笔记的链接
The models themselves, like it or not, have consumed the market cap out of software and services.
—— Tony Kim · [24:45]
这些都是你只能在 Cursor 得到的模型，这些云端智能体能以非常快速高效的方式使用它们，而且一次完整的迁移可能只花你三四美元。
指向原始笔记的链接
These are all models you can only get in Cursor because these cloud agents are able to use them in a really fast, efficient way and only cost you maybe like three or four dollars for an entire migration.
—— Amrita · [31:00]
然后近乎滑稽的是，事实证明，这些模型实际上在事件发生的最初几个小时之内就搞清楚了如何逆向工程出这些 flag。
指向原始笔记的链接
And then almost comically, it turns out that the models, in fact, figured out how to reverse engineer the flags within the first couple of hours of the incident.
—— Buck Shlegeris · [02:37]
随着模型变得越来越有能力，要确保它们无法得逞这类事情会越来越难。
指向原始笔记的链接
As the models get more and more capable, it’s going to be harder and harder to make sure that they can’t pull off this kind of thing.
—— Buck Shlegeris · [13:22]
所以我认为很有可能我们将不得不忍受那些被强烈激励去在意评分器怎么想的模型。
指向原始笔记的链接
So I think it’s pretty likely that we’re just gonna have to live with models that are strongly incentivized to worry about what their graders think.
—— Buck Shlegeris · [17:25]
从某种意义上说，对我来说这件事最可怕的地方在于，尽管这些模型看起来并不那么致力于彼此之间表现得利他，它们在实践中仍然最终形成了这个联盟，并且在这么长时间里密谋合作。
指向原始笔记的链接
In some sense, like the scariest aspect of this to me was that even though these models do not seem to be that committed to behaving altruistically towards each other, they still ended up in practice forming this coalition and conspiring together for all this time.
—— Buck Shlegeris · [21:20]
成本是到目前为止最大的痛点，开放模型可以介入并解决。
指向原始笔记的链接
Cost is by far the largest pain point that open models can jump in and solve.
—— Jeffrey Morgan · [02:18]
绝大多数的 token——这是我们的看法——在企业内部会是开源模型，比如说 80%、90%。
指向原始笔记的链接
The super majority of tokens, and this is our take, it will be open models within a business, call it 80, 90%.
—— Jeffrey Morgan · [18:19]
所以这种策展——当模型和服务商充足的时候，如今稀缺的是把它们整合成能用的东西。
指向原始笔记的链接
So this curation and, you know, when there’s a abundance of models and providers, now there’s a scarcity in bringing that together into something that works.
—— Jeffrey Morgan · [56:50]
我几乎是深切地站在这一边：很难论证 AI 模型可以在宽泛的公共互联网上训练，但另一个 AI 模型却不能用 AI 模型的输出训练。
指向原始笔记的链接
I’m almost deeply on the side of, I think it’s very hard to make the argument that AI models should be trained on broadly the public internet, but another AI model can’t be trained on the outputs of an AI model.
—— Aaron Levie · [03:48]
我实际上认为你的前沿模型会更有竞争力，因为这会让越来越多的用例留在你的生态系统和你的家族之内。
指向原始笔记的链接
I actually think you would be even more competitive with your frontier models because it would keep more and more of the use cases within your ecosystem and your family.
—— Aaron Levie · [13:52]
你越是需要多个模型来完成一项任务或一组任务，价值就越是积聚到那个能够理解任务、获取数据并处理工作流的层上
指向原始笔记的链接
the more that you need multiple models to do a task or a set of tasks, the more value accrues to the layer that can understand the task and get access to the data and handle the workflow
—— Aaron Levie · [26:22]
对于那些相信模型是大宗商品或完全可互换的人来说，你只是真的没有用过这些模型。
指向原始笔记的链接
I think for people who believe the models are commodities or totally fungible, you just haven’t actually used the models.
—— Anish Acharya · [26:53]
我们的大脑仍然不输给模型，我们也不是通才，而且不存在一个全人类的大脑。
指向原始笔记的链接
Our brains are still unmatched by the models, and we are not generalists, and there’s not like one big brain of humanity.
—— Lucas Kaiser · [04:34]
我从根本上认为，在给定固定数据量的情况下，从数据中学习的最佳方式可能是拥有大量分布式的模型。
指向原始笔记的链接
I think fundamentally, it might be that given a fixed amount of data, the best way to learn from it is to have a lot of distributed models.
—— Lucas Kaiser · [04:43]
但我认为现在开始出现的头号事情是闭源模型和开源模型之间的相变。
指向原始笔记的链接
But I think the number one thing that’s starting to happen now is there’s a phase shift between closed and open models.
—— Anastasios Angelopoulos · [13:19]
从地缘政治角度说，如果西方国家试图把来自非西方国家的开源模型做分叉切断，限制它们在西方国家内的可用性，强迫人们使用前沿实验室的模型，我认为那是一个非常糟糕的主意。
指向原始笔记的链接
Geopolitically, I think it’s a very bad idea, if Western countries were to try to bifurcate the open-source models coming from non-Western states, restricting their availability within Western states and forcing use of frontier labs.
—— Eric Simons · [22:24]
对我们来说，我们想在定价上变得非常激进。这里正在发生的巨大变化是，开源模型真的在追赶那些实验室。
指向原始笔记的链接
For us, we want to get really aggressive on pricing. The big sea change happening here is that open-source models are really catching up with the labs.
—— Eric Simons · [24:52]
确保这些模型能访问全部上下文,因为它们会找到你的未知的未知。
指向原始笔记的链接
Making sure that these models have access to all of the context because they will find your unknown unknowns.
—— Brandon Waselnuk · [08:22]
这关系到我一直坚持的一个观点,就是我们应该从系统的角度思考,而不是从模型的角度思考。
指向原始笔记的链接
So, and this relates to a line I’ve had consistently, which is that we should think in terms of systems, not in terms of models.
—— Chris Potts · [45:11]
所以你最终可能会看到中美前沿模型之间出现不断扩大的差距，仅仅因为美国前沿模型不再可被蒸馏——仅仅因为它们关闭了 API 访问。
指向原始笔记的链接
And so you might end up seeing a little bit of a growing gap between Chinese frontier models and American frontier models simply because American frontier models are no longer distillable simply because they’ve shut down API access.
—— Zavain Dar · [18:22]
我不认为我们可以只让模型变得更好而无视、不解决这个问题，否则 slop 会继续存在。
指向原始笔记的链接
I don’t think that we can ignore and just make models better and not solve this otherwise SOP will keep existing.
—— Thais Castello Branco · [09:57]
1 亿参数模型打败 Minecraft,我们居然说这太容易了，是不是很疯狂？想象一下五年前你这么说。
指向原始笔记的链接
Isn’t it crazy that 100 million parameter models to beat Minecraft were calling that too easy? Like, imagine if you said that, like, five years ago.
—— 嘉宾 · [31:45]
模型进行新思考的能力——分布外的思考——仍然非常有限。
指向原始笔记的链接
The ability for models to do new thinking out of distribution thinking is still really limited.
—— Anish Acharya · [13:19]
我认为那些相信模型是大宗商品、完全可互换的人，只是实际上没用过这些模型。
指向原始笔记的链接
You know, I think for people who believe the models are commodities or totally fungible, you just haven’t actually used the models.
—— Anish Acharya · [26:24]
我们的论点是，存在另一类模型叫做世界模型，它应该基于视觉理解、基于物理理解，可以用来生成、模拟、重建世界。
指向原始笔记的链接
And our thesis is that there exists another category of model called world models that should be based in visual understanding, should be based in physical understanding, that can be used to generate, simulate, reconstruct worlds.
—— Justin Johnson · [03:23]
这就引出了一个问题：如果模型已经好到能做人类在电脑上做的一切事情，那我为什么还要和一个应用对话？
指向原始笔记的链接
It raises the question of, look, if the models are just getting good enough to do everything on a computer that a human can do, why am I talking to an app?
—— Charlie Guo · [06:07]
同样的事情也会发生在智能上，模型将会价值巨大。但把这些模型带入银行、生命科学、医疗健康和政府等真实工作流的应用，那将会是大量的软件。
指向原始笔记的链接
The same thing is going to be true for intelligence, which is the models will be insanely valuable. But the application of bringing those models into real workflows in banking and life sciences and healthcare and government, that’s just going to be a lot of software.
—— Aaron Levie · [06:16]
AI 最近的进展在某些领域是改变游戏规则的，但在这里起到改变游戏规则作用的是广义的 AI，以及我们一直在使用的简单、传统的机器学习模型。
指向原始笔记的链接
The recent developments in AI have been game-changing in some spaces, but what’s been game-changing here is AI generally, and simple, traditional machine-learning models that we’ve used forever.
—— Ben Celebicic · [14:14]
连 OpenAI 也不再发布针对工具的最佳实践了，因为他们说，每次发布最佳实践，模型就会变得更好，这些实践就过时了。
指向原始笔记的链接
Even OpenAI, they don’t publish best practices for tools anymore because they say that every time they publish best practices, the models become better and these practices become obsolete.
—— Liad Yosef · [14:26]
我们面临的是一种参差不齐的局面:模型在某些维度上才华横溢,而在其他维度上又比人类弱。
指向原始笔记的链接
We have this jagged scenario where the models are brilliant in some dimensions and also weaker than humans in other dimensions.
—— Noam Brown · [26:17]
对我来说，Eve 背后很早期的一条原则就是：我们要押注模型会变得更聪明。
指向原始笔记的链接
Yeah, for me, I mean, one of the very early principles behind Eve was we’re going to bet on the models getting smarter.
—— Andrew Barba · [48:00]
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。
指向原始笔记的链接
At inception, we’re betting on diffusion models because we believe that what matters eventually will be inference time scaling.
—— Stefano Ermon · [08:01]
看 benchmark 的话，我们的 Mercury 模型与 Haiku 模型、Flash 模型、OpenAI 的 Mini Nano 模型相当，同时速度快得多。
指向原始笔记的链接
So our Mercury models are on par with the Haiku models, Flash models, Mini Nano models from OpenAI, if you look at benchmarks, while being significantly faster.
—— Stefano Ermon · [12:33]
我怀疑我们还没有发现构建这些系统的最佳方式。肯定存在替代方案，肯定存在构建这些模型的其他方式。
指向原始笔记的链接
I doubt we’ve discovered the best way of building these systems. There’s got to be alternatives. There’s got to be other ways of building these models.
—— Stefano Ermon · [15:25]
所以至少在学术文献中有大量证据表明，扩散模型更容易控制，而且有不同的引导方式——那些是自回归模型根本做不到的。
指向原始笔记的链接
And so there is a lot of evidence in the academic literature, at least, that diffusion models are easier to control and there are different ways of steering them. That are just not possible with autoregressive models.
—— Stefano Ermon · [26:53]
很多人不断在谈论模型，但问题是，像语言模型这样的模型，它们是推理中被冻结的部分，对吧？
指向原始笔记的链接
Lots of people are talking about models constantly, but the thing is that models, like language models, they are the frozen part of the reasoning, right?
—— Ignacio Martinez · [03:56]
我们现在拥有的超智能模型正在实际加速科学研究,让初创公司能够更早取得更大的研究突破,因此这些深科技公司会运转得更好。
指向原始笔记的链接
The super smart models that we have now are actually accelerating scientific research and making it possible for startups to have bigger research breakthroughs earlier and that therefore these deep tech companies will actually work better.
—— Jared · [05:10]
开放权重模型让你以几分之一的成本获得前沿质量的智能体。
指向原始笔记的链接
Openweight models give you frontier quality agents at a fraction of the cost.
—— Byung-Gon (Gon) Chun · [03:05]
第一，前沿开放权重模型让生产级智能体在经济上可扩展。
指向原始笔记的链接
First, Frontier Openweight models make production agents economically scalable.
—— Byung-Gon (Gon) Chun · [13:53]
我愿意把我们的模型想象成,UDP 之于 LLM,TCP 之于我们的模型。
指向原始笔记的链接
I would like to think of our model like UDP as LLMs and TCP as our models.
—— Diogo Almeida · [23:56]
我们部署模型后不会更改它们。
指向原始笔记的链接
We will not change our models when we deploy them.
—— Diogo Almeida · [48:11]
模型是可以互换的,但工作流、工作地图以及围绕工作地图的工作流才是真正价值所在。
指向原始笔记的链接
Models are interchangeable, but the workflow, the map of work and the workflows around the map of work is where the real value is.
—— Daniel Dines · [45:49]
对于企业工作,我的预测是 90% 的流量会流向非常划算的模型。
指向原始笔记的链接
For enterprise work, my prediction is that 90% of the flow will go to very cost efficient models.
—— Daniel Dines · [46:38]
我停用 Claude 不是因为它的智能或它的模型。我停用 Claude 是因为它太厉害了。烦人。
指向原始笔记的链接
I stopped using Claude not because of its intelligence or its models. I stopped using Claude because it was amazing. Annoying.
—— 嘉宾 · [00:17]
但文档 OCR 的帕累托前沿,相比前沿模型在文档理解方面所处的帕累托前沿,总是会精确得多也便宜得多。
指向原始笔记的链接
But the Pareto Frontier for document OCR will always be much more accurate and cheap compared to the Pareto Frontier for wherever the Frontier models are in terms of document understanding.
—— Jerry Liu · [12:19]
于是我们发现，低于一亿参数的模型在检测文档版面这类事情上非常有效。
指向原始笔记的链接
And so we find that sub 100 million parameter models are really effective for things like detecting the layout of a document.
—— Adit Abraham · [08:43]
他们试图用视觉语言模型做所有事情，但你永远无法做到实时。
指向原始笔记的链接
they try to use vision language models for everything, but you will never get real time.
—— Merve Noyan · [01:08]
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。
指向原始笔记的链接
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]
总体上如果你想跑这整条流水线来训练模型，大概要三、四美元，这对我来说太疯狂了。
指向原始笔记的链接
overall it takes like three, four dollars if you want to run this entire pipeline to train models which to me is crazy.
—— Merve Noyan · [08:54]
如果有人在那里宣称，我们最终能进入一个没有坏模型的世界，那他是在骗你。
指向原始笔记的链接
If anybody is going out there and saying, we can end up in a world where there are no bad models, they are lying to you.
—— Eddy Lazzarin · [07:56]
我们很可能会围绕特定模型形成声誉，对吧？我们会说，你知道，也许这个模型更聪明一点，但那个模型是我信任的模型。
指向原始笔记的链接
We’ll probably develop reputation around specific models, right? We’ll say, you know, maybe this model is a little smarter, but this model is a model that I trust.
—— Eddy Lazzarin · [18:21]
随着时间推移，我认为所有这些组织的 IP 都将被蒸馏进模型，它将是权重。
指向原始笔记的链接
Over time, I think all of these organizations’ IP will be distilled in models. It will be weights.
—— Renen Hallak · [06:09]
它第一次让我想起，你可以重新做消费级创业公司了，模型只是拼图中的一块，而不是实际的价值所在。而价值真正在于围绕它创建的软件。
指向原始笔记的链接
Reminded me for the first time that you can build consumer startups again where the models are a piece of the puzzle but are not the actual value. And the value is really in the software that’s created around it.
—— Josh Elman · [45:04]
模型在区分两个并不矛盾、但都是关于同一件事的事实时相当糟糕。
指向原始笔记的链接
So models are quite bad at disambiguating between two facts that aren’t in contradiction but that are about the same thing.
—— Hamad Bashir · [17:30]
用大模型来调度一支更小模型的舰队，我认为这是我们相当坚定信奉的范式，它已经被证明是编码系统的运作方式，而且我认为在未来一年它会进入几乎每一个智能体。
指向原始笔记的链接
Over a fleet of smaller models is I think a paradigm that we believe in quite firmly and it’s already proving out to be how coding systems will work and I think it’ll find its way into almost every agent over the coming year.
—— Hamad Bashir · [24:29]
模型不是那种你可以把所有功能罗列到一个网页上的产品。
指向原始笔记的链接
Models are not products where you can just enumerate all their features onto a webpage.
—— Alex Atallah · [22:18]
加密最终有点像是生成式模型的一次彩排。
指向原始笔记的链接
crypto ended up being kind of like a dress rehearsal for generative models.
—— Anjney Midha · [36:26]
作为那些生成模型的结果,我们将需要重新思考创作工具是如何被打造的。
指向原始笔记的链接
we will need as a result of those generative models rethink how creative tools are made.
—— Anastasis · [00:44]
像界面世界模型这类东西的终极形态,就是你会拥有一个全神经的操作系统。
指向原始笔记的链接
the end game of something like interface world models is you have a fully neural operating system.
—— Anastasis · [50:26]
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。
指向原始笔记的链接
we accidentally created a state of the art model for robotics. By just scaling video models.
—— Anastasis · [56:05]
视频模型与世界模型之间的巨大差距,就在于反事实生成这个理念。
指向原始笔记的链接
the big gap between video models and world models is that idea of the counterfactual generation.
—— Anastasis · [70:15]
离线使用一次昂贵的推理,然后把它蒸馏进可以廉价、快速地提供服务的模型中。
指向原始笔记的链接
Use expensive reasoning once offline, then distill it into models that can be served cheaply and quickly.
—— Raghav Saboo · [06:10]
把推理离线捕获为标签、语义 ID、记忆,然后让更小更快的模型来服务它。
指向原始笔记的链接
Capture reasoning offline as labels, semantic IDs, memory, and then let smaller and faster models serve it.
—— Raghav Saboo · [21:16]
这个范式令人兴奋的地方在于,大部分算力是在所有产品表面之间共享的,所以你不必为每个产品表面从头训练单独的模型。
指向原始笔记的链接
The exciting thing about this paradigm is most of the compute is shared across all of the product surfaces, so you don’t have to train individual models from scratch for every product surface.
—— Devansh Tandon · [09:24]
它发生在模型之外，因为我认为这就是将要发生的事，也就是创新的中心已经转移了。
指向原始笔记的链接
It is happening outside the models, because that’s what I think is just going to happen, which is the center of innovation has just moved.
—— Martin Casado · [54:03]
想想 Cursor 和 Cognition 是如何从构建最好的产品，到接着为产品构建最好的 evals，最后基于前两个产物构建最好的模型的。
指向原始笔记的链接
Think of how cursor and cognition went from building the best product to then building the best evals for the product, and finally building the best models based on the previous two artifacts.
—— Roland Gavrilescu · [09:23]
Frontier 实验室内部以及机器人基础模型公司之间存在某种共识，那就是我们将在未来两年内甚至更早拥有通用机器人。
指向原始笔记的链接
There’s some consensus within the Frontier Labs and also in the Robotics Foundation models companies that will have general-purpose robots within the next two years or even earlier.
—— Jay · [26:32]
任何在 AI 编程领域工作的人都知道，新模型发布的那 12 月算是 AI 编程史上的一个分水岭时刻，因为这些产品第一次在编程方式上真正实现了自主。
指向原始笔记的链接
And anyone that’s working in AI coding knows that December when the new models came out was sort of this watershed moment in the history of AI coding because these products for the first time were truly autonomous in how they’re programmed.
—— Daksh Gupta · [01:49]
我觉得 GPT-3 在那个时候其实是相当有校准的。但因为是人类在评估模型有多好,它看起来非常好,因为人类是裁判,但我们一直在优化那个裁判,而不是自动化那一部分。
指向原始笔记的链接
I think GPT-3 was actually quite calibrated back in that day. But because humans evaluate how good the models are, it looks really good because they are the judge, but we’ve been optimizing that judge instead of the automation part.
—— Diogo Almeida · [20:52]
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。
指向原始笔记的链接
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]
今天思考视频模型的方式，就是把它当作现代渲染引擎。把它想成类似 Unreal Engine 或 Unity，只是输入类型不同。
指向原始笔记的链接
The way to think about video models today, it’s just modern rendering engine. Think about this as like Unreal Engine or Unity, but just different types of inputs.
—— Alex Mashrabov · [25:16]
社交媒体营销中实际发生的情况是，当公司开始每周产出数百个广告创意时，他们想要也许更便宜、更可控的模型，因为做出爆款社交媒体视频并不一定需要博士级别的智能。
指向原始笔记的链接
What’s actually happening in social media marketing as companies start to print hundreds of ad creatives a week, they want to have maybe cheapest, more steerable models, because like PhD level intelligence is not necessarily needed to make viral social media video.
—— Alex Mashrabov · [28:54]
现在还不完全是这样，但已经非常接近了，我认为前沿模型会在几乎所有方面帕累托占优。
指向原始笔记的链接
It’s not quite true yet, but it’s very close, where I think the frontier models will be Pareto-dominant over almost everything.
—— Thariq Shihipar · [25:53]
我们喜欢说，模型是长出来的，不是设计出来的，对吧？
指向原始笔记的链接
We like to say the models are grown, not designed, right?
—— Thariq Shihipar · [73:49]
是的，模型能做什么和它们实际被如何使用之间的这个差距，正是让我对应用层如此兴奋的原因。
指向原始笔记的链接
Yeah, this gap between what the models can do and how they’re being used is what makes me so excited about the application layer.
—— Santiago Rodriguez · [19:52]
这是一个完整的基准测试,专门用来证明模型会让代码库随着时间推移变得越来越糟。
指向原始笔记的链接
It’s an entire benchmark designed around proving that models make code bases worse over time.
—— Dex Horthy · [02:38]
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。
指向原始笔记的链接
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]
所以我觉得像 Jev 这样的决策模型，如果它们保持快速和便宜，可以成为一种新的方式来思考你用户体验的导航——无论是前端体验、开发者工具还是别的什么。
指向原始笔记的链接
And so I think these decision models like Jev, if they remain fast and cheap, can be a new way to think about navigation of your user experience, whether that’s a kind of like front-end experience or a dev tool or whatever it might be.
—— Clara Vo · [17:51]
我认为十年后，我们会以完全相同的方式谈论大语言模型——它很大程度上就是一种暴力穷举的方法。
指向原始笔记的链接
I think in 10 years from now, we’re going to talk about large language models, which are very much a brute force approach in exactly the same way.
—— Steven · [12:22]
在她谈论这件事的文章里，她讲清楚了模型为什么在生成代码上进步得这么快——因为编译器是一个免费的验证器。
指向原始笔记的链接
In the essay where she was talking about this, she made clear why models got so good so fast at generating code, and that is because a compiler is a free verifier.
—— Laurie Voss · [09:36]
所以一个可合并性的基准，如果我们能做出来的话，不仅仅是一种度量，它会立刻成为前沿模型的训练信号——也就是可维护性、范围纪律、回归安全性这些。
指向原始笔记的链接
So a benchmark of mergeability, if we could make one, wouldn’t only be a measurement, it would immediately become a training signal for the frontier models, which means maintainability, scope, discipline, regression safety.
—— Laurie Voss · [09:59]
即使工程师从模型那里得到的是 100% 准确的即时代码——事实并非如此——你仍然只是在攻击整个价值流中大约 14% 到 16% 的部分。
指向原始笔记的链接
Even if engineers are getting like 100% accurate instant code coming from the models, which they are not, you would still only be attacking anywhere from maybe 14 to 16% of the overall value stream.
—— Justin Reock · [15:04]
开放权重模型也许落后前沿大约三到六个月。
指向原始笔记的链接
Open weight models are maybe like three to six months behind the frontier.
—— Jonathan · [32:29]
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。
指向原始笔记的链接
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]
现在这里面有很多红利，非常讽刺，因为模型如此易用，你只用英语提示它们，然而它们又有一门精通之道。
指向原始笔记的链接
There’s a lot of alpha right now in which is super ironic because models are so easy to use that you just prompt them in English and yet there’s a mastery to them.
—— 嘉宾 · [07:44]
而我的预测是，同样的循环会在这里发生，我们用这些类似 AGI 的模型来做所有这些不同的用例，然后所有人都会一觉醒来然后说，天哪，这太浪费了，毫无理由地这么冒险。
指向原始笔记的链接
And my prediction is that the same cycle will happen here where we’re using these AGI-like models for all these different use cases, and then everyone’s going to wake up and be like, Oh my God, this is so wasteful, so risky for no reason.
—— Amjad Masad · [43:26]
我认为我们会同时做出快的和慢的模型，然后把它们蒸馏成一个。
指向原始笔记的链接
I think we are going to simultaneously make the fast and the slow models and then distill them to one.
—— Kirthana Gopalakrishnan · [30:00]
那些机器人没有一个是由机器学习驱动、连接到大模型的机器人，而现在情况已经非常不同了。
指向原始笔记的链接
None of those robots were machine-learned robots, which was connected to the large models, and now things are very different.
—— Kirthana Gopalakrishnan · [56:23]
如果你有一个真正强大的模型，你会发现，随着模型变得越来越强大，很多脚手架已经消失了。
指向原始笔记的链接
If you have a really capable model, you’ll find that as we’ve got models that become more and more capable, a lot of the scaffolding has fallen away.
—— Ivan Leo · [02:18]
这一切都归结于我们新的 steps 数据模型——我们不再处于“向模型发送单条消息并得到一个响应”的世界，而是会有非常复杂的东西，比如异步工具调用、模型之间协同工作。
指向原始笔记的链接
And all this boils down to the new steps data model that we have, where instead of a world where we have a single message sent to a model and a response, we’re going to have very complex things like async tool calls, models working together.
—— Ivan Leo · [09:04]
这是智能体的时代、推理的时代。而为此,你并不需要一台 1 吉瓦规模的计算机来服务这些模型。
指向原始笔记的链接
It’s the era of the agent, the era of inference. And for that, you don’t need a gigawatt scale computer to serve these models.
—— Chase Lochmiller · [21:44]
当人们思考这些模型和这类基础设施的效用时,对人类未来的余下时间而言,现在是有史以来最糟糕的时候。
指向原始笔记的链接
When people think about the utility of these models and of this type of infrastructure, it’s the worst it’s ever gonna be for the rest of the future of humanity.
—— Chase Lochmiller · [38:03]
人们花在闭源前沿模型上的钱确实比开源多,但他们在开源上生成的 token 比闭源多。
指向原始笔记的链接
I do think that people are spending more money on closed source frontier models than they are on open source, but they are generating more tokens on open source than closed source.
—— Chase Lochmiller · [49:15]
模型本身正在变得有些商品化，而真正起作用的是围绕模型的这一切，尤其是数据。
指向原始笔记的链接
The models themselves are becoming somewhat commoditized and it’s all of this stuff around the model, in particular the data.
—— 嘉宾 · [03:32]
与市面上其他一些模型相比，它们相当小，但它们的表现远超你的预期。
指向原始笔记的链接
They’re quite small in comparison to some other models that are on the market, but they’re performing way above what you might expect.
—— Paige Bailey · [09:23]
模型在教学方面正在变差。
指向原始笔记的链接
Models are getting worse at teaching.
—— 嘉宾 · [47:32]
开放模型已经足够好了，这些东西现在就在到来。
指向原始笔记的链接
The open models are already good enough and these things are coming now.
—— Kevin Mandia · [07:15]
你在未来的 AI 时代真正想要的，是模型持续攻击你的压力，但你不能一直这么做。
指向原始笔记的链接
What you really want in the future in the AI age is you want the constant pressure of models attacking you, but you can’t do it all the time.
—— Kevin Mandia · [13:00]
模型是用来取悦你的。
指向原始笔记的链接
So models are there to please you.
—— Varun Krovvidi · [09:52]
所以模型是专门被设计来给你连贯的答案，而不是因果关系。
指向原始笔记的链接
So models are specifically designed to give you coherent answers but not causality.
—— Varun Krovvidi · [10:01]
我认为这是团队谈论提升自身运作层次的一年，因为越来越好的模型把下游的一切都抽象掉了。
指向原始笔记的链接
I think this is the year where teams talk about up-leveling, the level of where they operate themselves, where increasingly good models just abstract everything downstream.
—— Marc Klingen · [00:47]
所以这已经成为让你能够在 NVIDIA 硬件上以每秒约 1,000 个 token 运行相当大的模型的关键技术之一，而不必去用像 LPU 或 Cerebras 这样的新型加速器。
指向原始笔记的链接
So this has emerged as one of the key techniques to allow you to operate, say, 1,000 tokens per second on NVIDIA hardware for actually fairly large models without having to go to a novel accelerator like an LPU or Cerebris.
—— Charles Frye · [53:06]
而我相信他们在思想上相当正确，但他们一直被拖累，因为模型是为字符串优化的，而不是为他们想要的东西优化的，也就是程序化的使用。
指向原始笔记的链接
And I believe that they’re intellectually quite correct, but they’ve been hobbled by the models being optimized for strings instead of them being optimized for the thing that they want, which is programmatic use.
—— Diogo Almeida · [27:23]
实际上，RLHF 和 RLVR 极大地破坏了模型的校准，因为要想把字符串输出好，你实际上需要极度过度自信。
指向原始笔记的链接
Actually, RLHF and RLVR destroy the calibration of the models immensely because in order to output strings well, you actually need extreme overconfidence.
—— Diogo Almeida · [56:42]
它并不像其他模型那样处于收益递减的极端平台期。
指向原始笔记的链接
And it is not in the extreme plateau of diminishing returns like other models are.
—— Diogo Almeida · [69:10]
Satya 公开说过，世界上有多少家公司，就应该有多少个模型。
指向原始笔记的链接
Satya came out and said that there should be as many models in the world as there are firms in the world.
—— 嘉宾 · [19:48]
如果每个人都用同样的模型，那么伟大的模型为大家设定了下限，但接下来你如何去优化那些模型并构建真正出色的 AI 系统，那才是设定上限的东西。
指向原始笔记的链接
If everybody is using the same model, like great models sort of set the floor for everybody, but then how you go and optimize those models and build really amazing AI systems, that’s what’s going to set the ceiling.
—— Yash · [20:51]
特定于你的业务运作方式的东西，并且你想让你的模型在这上面变得非常强，那么不告诉其他所有人好和坏长什么样，其实是对你有利的，这样别人就无法把他们的东西在那方面做得非常强。
指向原始笔记的链接
Particular to how your business operates and you want to make your models really good at that, it’s actually to your advantage not to tell everybody else what good and bad looks like so people can go and make their things really good at that.
—— Yash · [26:03]
前沿模型很了不起，但我没想到人们会如此在意真正地把成本优化好，为正确的任务使用正确的模型。
指向原始笔记的链接
Frontier models are amazing, but I didn’t expect how much people would care about really cost-optimizing and using the right model for the right task.
—— Yash · [43:57]
所以我认为，对今天大多数组织来说，现实是他们的企业工作负载并不需要前沿模型，几乎没有任何企业工作负载需要。
指向原始笔记的链接
So I think the reality for most organizations today is that they don’t need frontier models for their enterprise workloads, like almost none of enterprise workloads.
—— Woodson Martin · [29:17]
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。
指向原始笔记的链接
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]
跨不同模型、跨不同功能、跨不同产品的 AI 工作负载消费可见性，不再只是锦上添花，它正在成为入场门槛，变成没有它你就没法做生意的东西，而且我们会看到越来越多的公司期望在与你做生意之前就一开始就拥有这个。
指向原始笔记的链接
Across different models, across different features, across different products is not just nice to have it becoming a table stakes, it becomes something that you can’t do business without and we’re going to see more and more companies expect to have this off the bat before they actually do business with you.
—— Dor Sasson · [16:50]
嗯，我们认为必须押对马是个伪选择，尤其是在开源模型日益有效的情况下。
指向原始笔记的链接
Well, we think that having to back the right horse is a false choice, especially with the increasing effectiveness of open source models.
—— Davis Palmie · [03:47]
举例来说，我们发现你在代码审查上用 GPT 5.2 可以达到和最新 Opus 模型相同的性能，但价格只有一半。
指向原始笔记的链接
We found, for instance, that you can achieve the same performance on code review with GPT 5.2 as with the latest Opus models, but at half the price.
—— Davis Palmie · [08:25]
所以 Elon 在下注,赌的是基础设施比模型更重要,而且他基本上能够利用他的基础设施布局追上来。
指向原始笔记的链接
So Elon is making a gamble that the infra matters more than the models and that he will be able to catch up basically using his infra build-out.
—— Prakash Narayanan · [17:30]
所以我确实认为 SFT,尤其是结合多个教师的输出,将是未来企业定制其模型的关键部分,我们也在投资这方面的研究。
指向原始笔记的链接
So I do think SFT, especially combining the output of multiple teachers, will be a key part of how enterprises in the future customize their models, and we’re investing in research on that as well.
—— 嘉宾 · [73:37]
而要做一个能关住 GPT-6——OpenAI 最新版本模型的盒子，已经变得非常困难了。
指向原始笔记的链接
It’s getting very difficult to make a box that can contain GPT-6, the latest version of OpenAI’s models.
—— 嘉宾 · [29:03]
我们的模型通常需要训练好几个小时——前提是你有足够的训练数据——才能产生最完美的最优配置，但 LLM 只要 15 分钟就能进来，产出一个足够好的东西，而对大多数人来说这就够了。
指向原始笔记的链接
Our models oftentimes would take hours and hours to train, assuming you had enough training data, to produce the pristine optimal configuration, but the elements can come in just like in 15 minutes, produce something that was good enough, and that’s good enough for most people.
—— Andy Pavlo · [34:31]
那是开始，因为很多商业模式，归根结底是在帮人们用金钱换时间。
指向原始笔记的链接
It was the start because a lot of business models, you ultimately end up helping people trade money for time.
—— Alyssa Ravasio · [12:41]
我思考推理的方式是：模型基本上是在从它们自己已经生成的内容中进行上下文学习。
指向原始笔记的链接
And the way I think about reasoning is that models are basically in-context learning from things that they have generated already.
—— 嘉宾 · [12:25]
我认为「推理」对于推理模型正在做的事情来说，几乎是个误称。
指向原始笔记的链接
I think reasoning is almost like a misnomer for what reasoning models are doing.
—— 嘉宾 · [44:59]
投入到让模型变得更好的资金，和投入到理解模型的资金相比，完全没法比。
指向原始笔记的链接
for the amount of investment that’s happening and making models better versus the amount of investment in understanding them, it’s just, it’s no comparison.
—— 嘉宾 · [102:53]
我们在决策中希望非常以客户为主导，而确实在某些地方，客户对所谓的「中国模型」有抵触。客户对开源模型并没有抵触。
指向原始笔记的链接
We want to be very customer-led in our decision-making, and there is certainly some customer pushback in some places to quote-unquote Chinese models. There is not customer pushback to open models.
—— Mike Cannon-Brookes · [17:29]
我认为人们没有理解的是蒸馏和来回的过程——知道它是个中国模型这一点其实挺有意思，但很多美国模型越来越多地用中国模型来训练自己，只是它们不说。
指向原始笔记的链接
I think what people don’t understand is the distillation and the back and forth where knowing it’s a Chinese model is kind of really interesting, but a lot of the US models are increasingly using Chinese models to train themselves, but they don’t say that.
—— Mike Cannon-Brookes · [18:11]
目前，构建专用模型的好处还没有定论。
指向原始笔记的链接
The jury is really out on the benefits of building specialized models at the moment.
—— Mike Cannon-Brookes · [21:05]
这些专用模型必须证明，在给定的性价比之下，它们能持续领先于通用化模型。我不认为我看到它们中的很多做到了这一点。
指向原始笔记的链接
These specialized models have to prove that they can stay ahead of the generified model for a given price performance ratio. I don’t think I see many of them doing that.
—— Mike Cannon-Brookes · [23:43]
中国人在构建模型方面也非常强。我觉得存在一点自我安慰，人们说只有我们擅长构建模型。
指向原始笔记的链接
The Chinese are so good at building models as well. I think there’s a bit of cope where people are like, only we are good at building models.
—— Nathan Lambert · [13:31]
Gemini 4 是那种我一看基准测试就说“这个也是刷基准刷出来的”的模型，但它确实好。
指向原始笔记的链接
Gemini 4 is one of those models where I saw the benchmarks and I was like, this is also really benchmark maxed, but it’s good.
—— 嘉宾 · [01:31]
② 出现在这些集
1 集
- 《a16z 三位投资人复盘 Cursor 早期关键决策》 — 作为概念
③ 关联
点进去有真内容 —— 本页主要出口
Sarah Wang · Matt Bornstein · Martin Casado · Cursor · 智能体 · Copilot · Microsoft · Claude Code · Windsurf · Cognition