BuildSpeak每日 builder 文摘
今日归档生词本关于
🎙 播客Unsupervised Learning· 2026 年 6 月 12 日· 13,818 词 · 约 69 分钟

AI Vibe Check: Lab Wars, Why APIs Might Vanish & Future Predictions

SPACE 播放 / 暂停·←→ 上一句 / 下一句
Speaker 100:00 - 00:22
I'm Jacob Efron, and this is unsupervised learning. We've had a bunch of new subscribers over our last few months, and so wanted to welcome you to the show. We basically probe the sharpest minds in AI on everything that's happening today, what's real, and what's coming up, where the space is headed. And today's episode is one of my favorite formats we do. It's an AI vibe check, that we do with Ari from Datalogy.
Speaker 100:00 - 00:22
我是 Jacob Efron,这里是 unsupervised learning。过去几个月我们新增了不少订阅者,所以想先欢迎大家来到这档节目。我们的内容基本上就是和 AI 领域最敏锐的一些头脑深入聊聊当下正在发生的一切:什么是真的,接下来会发生什么,以及这个领域将走向哪里。今天这一期也是我最喜欢的节目形式之一——和来自 Datalogy 的 Ari 一起做一次 AI vibe check(氛围检查/现状盘点)。
Speaker 100:22 - 00:46
Ari was a former researcher at DeepMind and Meta, now runs a really exciting AI startup, and Rob at Radical, one of the great AI venture firms. The three of us talk about everything happening in the AI world today. We talked about Fable, of course, and the reaction around the release as well as model capabilities. We talked about how close we are to RSI. We hit on some pretty spicy predictions, including that the labs may actually get rid of their API business as the compute crunch continues.
Speaker 100:22 - 00:46
Ari 曾是 DeepMind 和 Meta 的研究员,现在在运营一家非常令人兴奋的 AI startup;还有 Radical 的 Rob,Radical 是非常优秀的 AI venture firm(风投机构)之一。我们三个人聊了当今 AI 世界里正在发生的各种事。我们当然谈到了 Fable,也谈了它发布后的市场反应以及 model(模型)能力。我们还聊到了我们距离 RSI 到底还有多近。我们也抛出了一些相当火辣的预测,包括随着 compute(算力)紧张持续加剧,各家 lab(实验室)可能真的会砍掉自己的 API 业务。
Speaker 100:47 - 01:01
And we just got to hit on all the main topics of today. Just really fun to sit down with two friends and and great minds in AI. I think folks will really enjoy this. Without further ado, here's our conversation. It's time for another roundup episode.
Speaker 100:47 - 01:01
总之,今天最核心的话题我们都聊到了。能和两位朋友、也是 AI 领域的优秀头脑坐下来一起聊,真的非常有趣。我觉得大家会很喜欢这一期。闲话少说,下面就是我们的对话。又到了新一轮 roundup episode(盘点节目)的时间。
Speaker 101:01 - 01:16
I always love doing this with you guys, Ari and Rob. I feel like we had a ton of fun in the last one, but, like, god, it's AI world. Things have changed. I feel like we we last sat down after NeurIPS, and I think since then, we've had IPO filings. We've had, you know, models not launched and then launched.
Speaker 101:01 - 01:16
我一直都很喜欢和你们一起做这个,Ari 和 Rob。我感觉上一期我们聊得特别开心,但,天啊,这可是 AI 世界,事情变化太快了。我觉得我们上次坐下来聊,好像还是在 NeurIPS 之后,而从那以后,我们已经看到了 IPO filing(IPO 申报)。还有,你知道的,一些 model(模型)先是不发布,后来又发布了。
Speaker 101:16 - 01:24
We've had, you know, SpaceX becoming an AI infra company. No shortage of headlines to to discuss here. So excited to have you both back on the show.
Speaker 101:16 - 01:24
还有,你知道的,SpaceX 都快变成一家 AI infra(基础设施)公司了。可以讨论的 headlines(头条)真是一点都不缺。所以非常高兴你们两位能再次回到节目里。
Speaker 201:24 - 01:25
Excited to be here.
Speaker 201:24 - 01:25
很高兴来到这里。
Speaker 301:25 - 01:26
Yeah. Thanks for having us.
Speaker 301:25 - 01:26
对,谢谢你邀请我们。
Speaker 101:26 - 01:40
So I I I think to kick it off, you know, six months is an eternity in AI world, but I figured I'd at the highest level. What's, like, the single biggest thing that has changed in how you're thinking about the landscape since we last talked? And maybe, Ari, I'll start with you.
Speaker 101:26 - 01:40
那我想就从这里开始吧。你知道,在 AI 世界里,六个月都算得上是沧海桑田了。所以我想先从最高层面问一个问题:自从我们上次聊完之后,在你对整个 landscape(格局)的看法里,最大的那个变化到底是什么?Ari,也许先从你开始。
Speaker 301:40 - 01:53
Yeah. I mean, I think the the most obvious thing that has changed, over the last six months is starting to see the coding agents really start to work at longer time horizons. Right? I think that was just starting when we recorded our last episode at the end of, '25.
Speaker 301:40 - 01:53
对。我的意思是,我觉得过去六个月里最明显的变化,就是开始看到 coding agents 真正开始能在更长的时间跨度上工作了,对吧?我觉得在我们上一次录节目——也就是在 '25 年末——的时候,这才刚刚开始出现。
Speaker 101:53 - 01:58
Feel like everyone went away over Christmas break and was like, holy crap, like these things really work.
Speaker 101:53 - 01:58
感觉大家圣诞假期一回来,都在想:天啊,这些东西居然真的能用。
Speaker 301:58 - 02:33
Yeah. And I think it starts to show how there are these thresholds where if you go beyond the threshold, it can become a lot more valuable. And obviously that's driven the massive rise in token spending and the whole token maxing idea and all this stuff. But I think we're really starting to now see the shift of engineers at least kind of almost all moving from ICs to managers of agents. That's been something that's been very noticeable within Datalogy, for example, over the last number of months is seeing more and more people starting to now context switch between managing different agents rather than just kind of working on the one thing.
Speaker 301:58 - 02:33
对。而且我觉得,这开始说明确实存在这样一些门槛:一旦跨过门槛,它的价值就会大很多。显然,这也推动了 token 支出的暴涨,以及整个 token maxing 这种想法,还有诸如此类的事情。不过我觉得,我们现在真的开始看到一种转变:至少工程师们几乎都在从 ICs(individual contributors,独立贡献者)转向 agent 的管理者。比如在 Datalogy,过去几个月里一个非常明显的现象就是,越来越多的人开始在管理不同的 agents 之间进行 context switch,而不再只是埋头做某一件事。
Speaker 302:33 - 02:39
That is enabled by having these agents be able to run long enough and actually be useful in various ways.
Speaker 302:33 - 02:39
而这之所以成为可能,是因为这些 agents 现在能够运行足够长的时间,并且确实能在各种方面发挥作用。
Speaker 102:39 - 02:44
Mean, everyone likes to ask top AI researchers like yourself, like how much more productive has it made in your work?
Speaker 102:39 - 02:44
我是说,大家都喜欢问像你这样的顶尖 AI 研究者:它到底让你的工作效率提高了多少?
Speaker 302:45 - 03:02
I think that it's interesting. It makes you a lot more productive in some ways, right? But it also produces a lot of challenge as well. Like one of the things that we're struggling with is now it's a lot easier to produce a massive amount of code that can do something. But now you have this pretty massive understanding gap and it's a lot easier to put swap into your code base.
Speaker 302:45 - 03:02
我觉得这很有意思。它在某些方面确实让你高效了很多,对吧?但同时也带来了很多挑战。比如我们现在在挣扎的一件事是,如今要产出大量“能完成某件事”的代码变得容易多了。但与此同时,你会面临一个相当巨大的理解鸿沟,而且也更容易把 swap 塞进你的 code base 里。
Speaker 303:02 - 03:23
So it's definitely made us more productive. I think a lot of times though, the kind of top line numbers tend to be overestimated because it doesn't take into account some of these like later costs of like, we now have big bottlenecks on reviews and we don't wanna go fully to like, oh, you just like, well, my agent will review your agent's output, you know? The bottlenecks just seem to shift.
Speaker 303:02 - 03:23
所以它肯定让我们更高效了。不过我觉得很多时候,那种 headline numbers 往往会被高估,因为它没有把一些后续成本算进去,比如我们现在在 review 上出现了很大的瓶颈,而且我们也不想完全走到那种“哦,你就让我的 agent 去 review 你的 agent 的输出吧”,你懂吧?瓶颈看起来只是转移了位置。
Speaker 103:23 - 03:29
You know, whatever it's hard to improve on an entire process because of because of that. What about you, Rob?
Speaker 103:23 - 03:29
你知道的,也正因为这样,要改进整个流程并不容易。你那边呢,Rob?
Speaker 203:29 - 03:43
There are early signs that seem to suggest over the past six months that that make me question whether open weight AI is going to continue to be a really meaningful force in the ecosystem going for at least like a near frontier
Speaker 203:29 - 03:43
过去六个月里出现了一些早期迹象,让我开始怀疑 open weight AI 是否还能继续在这个生态中作为一种真正有意义的力量存在下去,至少在接近 frontier(前沿)的层面上是这样。
Speaker 103:43 - 03:48
open weight jugular for Ari, like, right off the right off the top here. I like it.
Speaker 103:43 - 03:48
这算是直接对 Ari 的 open weight 要害下手了,一上来就这样。我喜欢。
Speaker 203:48 - 04:15
This yeah. Yeah. We can we can dig it in more detail, but I I think, like months ago or for the past few years, my working assumption had been that the closed source proprietary models would advance the frontier and there are lot of structural reasons for that. But the open frontier would only be a few months behind and that gap might widen a little bit. I didn't think it would shrink altogether, but, like, I thought it would persist as as being relatively small.
Speaker 203:48 - 04:15
对,是这样。对。我们可以再更详细地展开讲,但我觉得,几个月前或者说过去这几年里,我的一个基本判断一直是:closed source(闭源)的 proprietary models(专有模型)会推动 frontier 向前发展,而这背后有很多结构性原因。但 open frontier(开放前沿)只会落后几个月,这个差距也许会稍微拉大一点。我不认为它会完全缩小,但我原本觉得这个差距会持续存在,而且总体上会相对较小。
Speaker 204:15 - 05:02
And I think there are signs now that, like, there it seems like there's a real risk of of near Frontier open weight AI falling off altogether. I think Meta, which historically has been the the the open weight champion in the West, is is pulling back, and it it seems likely that they're not gonna continue with their open source strategy. And then more recently, obviously, the Chinese labs have been the ones driving state of the art open research. It seems like there are strong indications that they may also be pulling back from that. And whether it's Quen or DeepSeq or others, their most high performing models, they're now keeping proprietary behind an API just open sourcing, open weighting, smaller, less performant versions.
Speaker 204:15 - 05:02
而我觉得现在已经出现了一些迹象:near-frontier(接近前沿)的 open weight AI 似乎确实存在一种彻底掉队的风险。我认为 Meta 过去一直是西方 open weight 阵营的代表者,但现在它正在收缩,而且看起来很可能不会继续坚持他们的 open source(开源)策略。再往近了看,显然一直是中国的实验室在推动最先进的开放研究,但现在也有很强的迹象表明,他们可能也在从这条路上后撤。不管是 Quen、DeepSeq 还是其他团队,他们性能最强的模型现在都开始作为 proprietary 模型放在 API 后面,只把更小、性能更弱的版本 open source / open weight 出来。
Speaker 205:02 - 05:29
I think there's real compute incentives behind that. It's just very expensive to service these open way models with no revenue coming in. And I think there's also geopolitical competitive considerations. But it's interesting to contemplate what a world might look like where, like, if you want real frontier artificial intelligence, like, you have to pay a company for a proprietary model as opposed to being able to, you know, have access to the weights yourself.
Speaker 205:02 - 05:29
我认为这背后确实有很现实的 compute(算力)激励因素。因为为这些 open weight 模型提供服务而又没有收入进来,成本实在太高了。我也认为其中有地缘政治和竞争层面的考量。但设想这样一个世界还是很有意思:如果你想要真正 frontier 级别的 artificial intelligence(人工智能),你就必须向某家公司付费购买 proprietary model,而不是能够自己拿到 model weights(模型权重)。
Speaker 305:30 - 06:00
Or But build your I'd actually agree with that. I'll ask you the second one, the other one I was thinking about is, you know, what changed. I don't think that we've seen a major change in the capabilities or the trend line of the capabilities of the open weight models. I actually think if anything, we started to converge and we continue to see that. I do think though the economic decision making around building and releasing open models has definitely changed over the last six months, which I think is what Rob was really getting at.
Speaker 305:30 - 06:00
或者说——但 build your……我其实同意这一点。我再问你第二个问题,我刚才想到的另一个点是,到底是什么变了。我不认为我们看到 open weight 模型的能力,或者它们能力提升的趋势线,发生了什么重大变化。实际上,如果有变化的话,我觉得双方反而开始趋同了,而且这种情况还在继续。不过我确实认为,围绕构建和发布 open models(开放模型)的经济决策,在过去六个月里肯定已经变了,而我觉得这其实正是 Rob 真正想表达的意思。
Speaker 306:01 - 06:27
And I am a lot more bearish on how many open models there will be going forward. It seemed like, you know, there was this kind of cornucopia of open models that was only ever growing in 2025. And I think we're now definitely starting to see that we probably hit the peak number of open models and it's now going to kind of get less and less. And because the financial incentives just don't make sense. Once you've already kind of achieved credibility, it makes sense to invest a lot of money to do that.
Speaker 306:01 - 06:27
而且我对未来 open models 的数量明显更悲观了。之前看起来像是,到了 2025 年,open models 会形成一种只增不减的“丰饶之角”。但我觉得现在我们已经很明确地开始看到,我们可能已经到达了 open models 数量的峰值,接下来只会越来越少。因为这里面的 financial incentives(财务激励)根本说不通。你在还没有建立 credibility(信誉、市场认可)的时候,投入大量资金去做这件事是有道理的。
Speaker 306:27 - 06:43
But after that point, you want to start selling hosted inference of your model and opening it up just fully undermines your business, which, so I I so I think we're all gonna see, like, other Chinese labs start with, big open models, but then probably close-up after that once they've kind of gotten enough press and PR. Is there a
Speaker 306:27 - 06:43
但过了那个阶段之后,你就会想开始卖自己模型的 hosted inference(托管推理)服务,而把它完全开放出来其实就是在彻底削弱自己的生意。所以我觉得,接下来我们会看到其他中国实验室也先推出大型 open models,但在获得足够多的报道和 PR(公关曝光)之后,大概率也会再把它们收回去。有没有一种……
Speaker 106:43 - 06:55
business model for an open source model company or is it literally just it's marketing, to, you know, to as you're on your way to the frontier. And then once you're once you're kind of close, it makes it inevitable to to wanna go close source.
Speaker 106:43 - 06:55
开源模型公司的 business model(商业模式)是什么?还是说它实际上就只是 marketing(营销)——你知道的,在你通往 frontier(前沿)的路上用来做宣传;然后一旦你真的已经接近前沿,就几乎不可避免地会想转向 close source(闭源)。
Speaker 206:55 - 07:29
I don't think there's a business model, honestly. There you know, different things have been tried, the kinda, like, freemium, enterprise, like, you know, Red Hat model. But I just I don't think that it it works in AI given the just the massive upfront investment required to to get to the frontier or close to the frontier in the first place. We'll see what I like. You know, we can come back and revisit this soon, but I I very curious to see if for you know, first of all, for Reflection when Reflection releases a model and, what business model they aim to pursue with it.
Speaker 206:55 - 07:29
说实话,我不认为有这种商业模式。你知道,人们已经尝试过不同的做法,比如有点像 freemium(免费增值)、enterprise(企业版),还有像 Red Hat 那样的模式。但我就是不觉得这在 AI 里行得通,因为首先要达到 frontier 或接近 frontier,就需要极其巨大的前期投入。看看之后会怎样吧。我们很快可以再回来重新讨论这个问题,不过我非常好奇,比如首先,等 Reflection 发布模型时,他们会打算为它采用什么样的商业模式。
Speaker 207:30 - 07:36
But I yeah. I'm I remain skeptical that, like, a great business model exists for OpenAI open source AI.
Speaker 207:30 - 07:36
不过是的,我仍然怀疑,对于 OpenAI 开源 AI 来说,是否真的存在一种很好的商业模式。
Speaker 107:37 - 08:17
It's funny because I was gonna say one of the one of the trends we've actually seen happen, you know, in in the past, like, month or two is that it finally you know, forever, I think people have been like, everyone will use smaller, cheaper, open source models for tasks that those models can do. And, you know, it it felt like the vast majority of usage and tokens was still just, like, at the frontier pushing capabilities, like, figure out what, you know, models can do in different industries. And then finally, I feel like we have a a movement now toward, like, jeez, these bills are pretty expensive, or or usage is pretty high. Like, wouldn't it be nice to to have something that was that was, cheaper and and faster and smaller to use? But it's funny that it's happening at the same time that we're kind of seeing this real you know, the the closed source models, you know, run ahead of of of open source models.
Speaker 107:37 - 08:17
挺有意思的,因为我本来正想说,我们实际上在过去一两个月里看到的一个趋势是:终于,长期以来,我觉得大家一直都在说,凡是较小、更便宜的开源模型能够胜任的任务,所有人都会去用这些模型。可是你知道,之前绝大多数的 usage(使用量)和 tokens(token)似乎仍然都集中在 frontier 上,继续把能力往前推,去弄清楚模型在不同行业里到底能做什么。然后现在我终于感觉到出现了一股转向:天哪,这些账单真挺贵的,或者说 usage 已经很高了。要是能有更便宜、更快、更小的东西可用,不是很好吗?但有意思的是,这种变化发生的同时,我们也确实正在看到——那些闭源模型正在进一步拉开与开源模型的差距。
Speaker 108:17 - 08:23
Ari, I know you spend a lot of time thinking about this stuff. Like, how do we think about those, like, countervailing forces that seem to be happening at the same time?
Speaker 108:17 - 08:23
Ari,我知道你花了很多时间思考这些事情。像这种似乎同时发生的、彼此对冲的力量,我们该怎么理解?
Speaker 308:23 - 08:59
Yeah. I I mean, I think that first off, I I've definitely seen a lot of the the former that you were talking about. Like, it's been very interesting. The combination of the, you know, extreme compute constrained environments that we're operating in, the rise of capabilities of, the coding models in particular, and then even just seeing, you know, from release to release models changing their token efficiency, their output token efficiencies, has resulted in, you know, a lot of companies that were happy using frontier models. All of a sudden, even just going from like Opus 4.6 to 4.7, there was a big difference in token efficiency.
Speaker 308:23 - 08:59
对,我的意思是,首先,我确实已经看到了很多你刚才说的前一种情况。挺有意思的。我们所处环境里的极端 compute(算力)受限、尤其是 coding models(代码模型)能力的上升,甚至还有你知道的,不同版本发布之间模型在 token efficiency(token 效率)、output token efficiency(输出 token 效率)上的变化,这些因素叠加起来,导致很多原本很满意使用 frontier 模型的公司,突然之间——哪怕只是从 Opus 4.6 到 4.7——token 效率就出现了很大的差别。
Speaker 309:00 - 09:33
And a lot of people's bills just doubled overnight. And you've now I'm now starting to see talking to a lot of enterprises in particular, really strong desires to start cutting the cost of of using the models. And I think that wasn't there nearly to the same extent a year ago because the models weren't being used at a scale where those costs were meaningful enough. But now they've really reached that point where they are meaningful enough and you can just consume budgets so quickly because the models think for a long time. And that's now driving a lot of demand to say, okay, how can we make this much cheaper?
Speaker 309:00 - 09:33
结果很多人的账单几乎一夜之间就翻倍了。而现在,尤其是在和很多 enterprise(企业)交流时,我开始看到他们非常强烈地希望削减使用这些模型的成本。我觉得一年前这种情况远没有现在这么明显,因为那时模型的使用规模还没大到让这些成本变得足够有意义。但现在它们确实已经到了那个程度,成本已经足够重要了,而且因为模型会思考很久,你的预算消耗速度会快得惊人。这正在推动大量需求:好,那我们怎么才能把这件事做得便宜得多?
Speaker 309:33 - 09:59
I think that you can do a lot of that with open models as well. Like I think one very notable thing, right, is that like a lot of people were able to reproduce the same or similar level of kind of vulnerability finding that Mythos used with open models by just putting scaffoldings around. I think that's another one of the big changes, right, is that a model is not just a model anymore. It's the model combined with the harness and the scaffolding. And a lot of innovation is happening on the harness and scaffolding layer.
Speaker 309:33 - 09:59
我认为,用 open models(开源模型)也能在很大程度上做到这一点。比如我觉得一个非常值得注意的事情是,很多人只是通过在开源模型外面加上 scaffoldings(脚手架式封装),就能复现 Mythos 所用的、相同或相近水平的 vulnerability finding(漏洞发现)能力。我觉得这也是另一个巨大的变化:现在模型不再只是模型本身了,而是模型与 harness(运行/测试框架)以及 scaffolding(脚手架封装)的结合。很多创新正发生在 harness 和 scaffolding 这一层。
Speaker 309:59 - 10:21
I think that's also possibly how open source models can have a business model. You open source the model, you don't open source the scaffolding in the harness, and you then have an API where people can access the full system. I think that could potentially work. Kimi is kind of doing that, right? Like Moonshot I think is at a couple 100 milli ARR through the API and kind of the chat interface and that's to some extent what they're doing.
Speaker 309:59 - 10:21
我认为,这也可能是 open source model(开源模型)形成 business model(商业模式)的一种方式。你把模型 open source,但不把 harness(框架)里的 scaffolding(支撑层)开源,然后再提供一个 API,让人们可以访问完整系统。我觉得这可能是行得通的。Kimi 某种程度上就在这么做,对吧?比如我记得 Moonshot 通过 API 和聊天界面,ARR(年度经常性收入)大概已经做到几亿美元中的几亿美元区间,而这在某种程度上就是他们正在做的事。
Speaker 310:21 - 10:55
So I think that we will continue to see strong open models, but probably fewer and fewer. And I think that this is going to motivate a lot of folks where they have to now think, how do we survive in an era where there aren't going to be reliable open models? And I think that pushes people more and more towards actually starting to think about, okay, how do I make sure I have the capability to build a model in some way, to maintain a model, to do that repeatably? I think that's a huge part of the reason that NVIDIA has been pushing so hard through Reflection and others, right, to have open model providers in the West.
Speaker 310:21 - 10:55
所以我认为,我们会继续看到强大的 open models(开放模型),但数量可能会越来越少。我觉得这会促使很多人开始思考:在一个不再有可靠 open models 的时代,我们该如何生存?而我认为,这会越来越推动人们真正开始考虑:好,我怎样才能确保自己具备某种 build a model(构建模型)的能力,具备维护模型的能力,并且能够可重复地做到这一点?我觉得这也是 NVIDIA 一直通过 Reflection 等项目大力推动、希望西方拥有 open model providers(开放模型提供方)的一个重要原因,对吧。
Speaker 110:55 - 11:19
Yeah. I mean, certainly, like, the more model providers, the merrier for for them. And it's interesting. Mean, it kind of almost feels like, you know, you saw an early manifestation of this with Meta, right, where it's like early on, you know, they you know, idea of having this controlled by a bunch of companies that weren't them was it was kind of like it, you know, would never work for them as a business. And so it's like we have to have something here to be competitive even if we're not gonna be, you know, the leading model provider just for our own business.
Speaker 110:55 - 11:19
对。我的意思是,当然,对他们来说,model providers(模型提供方)越多越好。这很有意思。我的意思是,这几乎让人感觉,你在 Meta 身上其实已经看到了这种情况的早期体现,对吧?就是在早期,他们会觉得,如果这件事由一批并不是他们自己的公司来控制,那从商业角度看基本上是行不通的。所以他们会想:哪怕我们不会成为领先的 model provider,仅仅为了我们自己的业务,我们也必须在这里有点东西,才能保持竞争力。
Speaker 111:19 - 11:22
Like, we wanna have something that that that that we can use ourselves.
Speaker 111:19 - 11:22
比如,我们得有一个我们自己也能拿来用的东西。
Speaker 311:22 - 11:44
If you are one of these companies that's potentially threatened by this, you kind of have two options. Like option one is that you go and work with a frontier model provider. You share your data and your domain expertise with them. They use that to improve the model with you. Maybe you're able to sign some agreement that gives you some long term certainty around that, but eventually they just outcompete you because you've given up any of your proprietary advantages.
Speaker 311:22 - 11:44
如果你是一家可能因此受到威胁的公司,那你基本上有两个选择。第一种选择是去和一家 frontier model provider(前沿模型提供方)合作。你把自己的数据和领域专长分享给他们,他们用这些东西和你一起改进模型。也许你还能签一些协议,为这件事争取某种长期确定性,但最终他们还是会把你挤掉,因为你已经交出了自己任何 proprietary advantages(专有优势)。
Speaker 311:44 - 12:08
The only other option is that you try to really compete on the niche that is your focus and where you have unique capabilities. And I think we're starting to see that across the board, right? Look at what's happened with Cursor, look what happened with lots of other folks where they're realizing, okay, we have to kind of move away from this. There's also a margin perspective from that as well, right? Where just like fundamentally, if you're competing against somebody who has better margin than you do, then it's very hard to win.
Speaker 311:44 - 12:08
唯一的另一个选择,就是你真正围绕自己所专注的 niche(细分领域)去竞争,围绕你所拥有独特能力的地方去竞争。我觉得我们现在已经在各个领域看到这种趋势了,对吧?看看 Cursor 发生了什么,再看看很多其他公司发生了什么,他们正在意识到:好吧,我们必须某种程度上摆脱这条路。这里面还有一个 margin(利润率)视角,对吧?因为从根本上说,如果你是在和一个 margin 比你更好的人竞争,那就很难赢。
Speaker 112:08 - 12:18
Yes. They will always give themselves better rates than they give you as an app company on top. Rob, Rob, what was your whole reaction to the apps are cooked and SaaSpocalypse narrative of the past months?
Speaker 112:08 - 12:18
是的。他们给自己的价格条件,永远都会比给你这种叠加在上层的 app company(应用公司)更好。Rob,Rob,过去几个月里那种“apps are cooked”和“SaaSpocalypse”的说法,你整体上的反应是什么?
Speaker 212:19 - 12:45
Yeah. That's interesting. I think, like, market narratives have a way of, like, swinging so far in one direction or the other to the to the extent that a lot of the nuance is lost. So I do think, like, a couple things are true simultaneously. Like, I I do think a lot of traditional software companies and even categories potentially are at real existential risk on account of the Frontier Labs and their and their product road maps.
Speaker 212:19 - 12:45
对,这很有意思。我觉得 market narratives(市场叙事)常常会朝某一个方向摆动得太厉害,以至于很多细微差别都丢失了。所以我确实认为,有几件事是同时成立的。比如,我确实认为,很多传统 software companies(软件公司),甚至一些品类本身,确实因为 Frontier Labs 以及它们的 product road maps(产品路线图)而面临真实的生存性风险。
Speaker 212:45 - 13:19
And so I think, like, a lot of that rerating was rational. I think there are a lot of big companies that have big revenues and big customer bases that are in real trouble based on how OpenAI and Anthropic are executing and performing. I also think it, like, for sure was an overreaction and an oversell and, like, painted with too broad of a brush. And I think, like, you know, in in our field of venture capital, it certainly is, like, at the the kind of prevailing narrative is apps are so challenging to invest in. Stay away from them.
Speaker 212:45 - 13:19
所以我认为,很多那种 rerating(重估)其实是理性的。我觉得有很多拥有庞大营收和庞大客户基础的大公司,基于 OpenAI 和 Anthropic 的执行与表现,确实遇到了真正的麻烦。但我也认为,这肯定也有过度反应和过度抛售的成分,而且是一刀切地把很多东西都算进去了。我觉得,在我们 venture capital(风险投资)这个领域里,当前一种相当主流的叙事确实是:apps(应用)太难投了,离它们远一点。
Speaker 213:19 - 13:58
Like, deep tech and hardware and so forth have all become more appealing and consensus over the past few months. And, again, think there's a lot there's a lot that's real there, and there are a lot of deep tech categories that I'm super excited about that we're both investing in and I'm sure we'll talk about. But I also think there are still I I think it just fundamentally, there's no way that one or two or three companies will win every single important market and and important category in the world. It just no no company can have that span of influence and execute excellently so broadly. So for sure, there are some categories that I think I would bet on OpenAI and Anthropic being very well positioned and coding obviously is the first.
Speaker 213:19 - 13:58
像 deep tech(深科技)、hardware(硬件)之类的方向,在过去几个月里都变得更有吸引力,也更接近市场共识了。再说一次,我认为这里面有很多东西确实是真实成立的,也有很多 deep tech 的细分类别让我非常兴奋,我们也确实都在投资,我相信后面也会谈到。但我也仍然认为,从根本上说,不可能只有一两家或三家公司赢下世界上每一个重要市场和每一个重要类别。没有任何一家公司能拥有那么广泛的影响范围,还能在如此宽的战线上都执行得极其出色。所以,当然有一些类别里,我会押注 OpenAI 和 Anthropic 处于非常有利的位置,而 coding(编程)显然是第一个。
Speaker 213:58 - 14:15
There are others which we can talk through. And I think that these horizontal cross cutting areas are the ones that make the most sense for the labs to focus on and go after versus like, you know, vertical software for pet stores or something like this.
Speaker 213:58 - 14:15
还有其他一些领域我们可以展开聊。我认为,那些 horizontal(横向)、cross-cutting(跨领域)的方向,才是这些 labs(模型实验室)最应该聚焦和进攻的地方,而不是去做什么面向 pet stores(宠物店)的 vertical software(垂直软件)之类的东西。
Speaker 114:15 - 14:33
No. I I do wonder if like VCs are just, you know, pushing themselves into a corner of like, damned if you do, damned if you don't. If you're like, if you're only gonna go invest in, you know, pet stores because like that's small enough for the models, like, I don't know if that actually ends up being a super interesting end category. Right? But it turns out, no matter anywhere you look in AI world, there are essential you can't avoid them.
Speaker 114:15 - 14:33
不是。我确实会想,VCs(风险投资人)是不是正在把自己逼进一个“做也不是,不做也不是”的角落。如果你只打算去投 pet stores 这种方向,因为它对模型来说足够小,那我也不确定这最终会不会变成一个特别有意思的终局类别。对吧?但事实是,不管你在 AI 世界里看向哪里,这些核心问题本质上都绕不过去。
Speaker 114:33 - 14:48
Like there's no there's no just like I mean, maybe if you wanna go invest in like a data center builder and you're like, will just build data centers and there will be demands like but sure, then you still have to figure out like, you know, how what makes you a better data center builder than the other 10 people that will compete with aggressive financing for the same site.
Speaker 114:33 - 14:48
就是说,并不存在那种真正的“躲开一切”的地方。我的意思是,也许你可以去投一个 data center(数据中心)builder(建设商),然后想的是:我们就建数据中心,反正总会有需求。但即便如此,你还是得想清楚,和另外那 10 个会为了同一个地块、靠激进融资来竞争的人相比,你凭什么能成为更好的数据中心建设商。
Speaker 214:48 - 15:19
There's no question that we are at a at a period of max uncertainty and volatility. And, yeah, I I mean, I totally agree with you on the on the hard tech point. Like, it there again, I I I do believe there are a lot of really exciting opportunities in in categories in hardware and deep tech where there's a lot of interesting innovation happening right now. But, yeah, to your point, it turns out hard tech is also very, very hard. And, like, the failure rates are much higher and there's a lot of unsolved problems and, like, that There's gonna be a lot of, I think, pain and lost money down that road in the years to come.
Speaker 214:48 - 15:19
毫无疑问,我们正处在一个不确定性和波动性都达到高点的时期。而且,是的,我完全同意你关于 hard tech(硬科技)的观点。再说一次,我的确相信,在 hardware 和 deep tech 的一些类别里,现在确实正在发生很多有意思的创新,也存在很多非常令人兴奋的机会。但是,正如你说的,hard tech 也确实非常非常难。它的失败率高得多,还有很多未解决的问题。所以我认为,在接下来的几年里,这条路上会有很多痛苦,也会烧掉很多钱。
Speaker 215:19 - 15:42
And then, yeah, on on the application side, I I think, as you said, like, I think there's there's still tremendous value in that last mile. And, like, the labs are also starting to try to tool up and stand up these deploy codes and and lean in on the implementation side. But I think there will be so many pockets and opportunities for companies that are delivering applications to also thrive and continue to grow.
Speaker 215:19 - 15:42
然后,是的,关于 application(应用)这一侧,我觉得就像你说的,那最后一公里里依然有巨大的价值。而且这些 labs 也开始尝试把工具链搭起来,建立这些 deploy code,并且更积极地切入 implementation(实施)这一侧。但我认为,仍然会有很多细分空间和机会,让那些交付应用的公司同样能够成功,也能够继续增长。
Speaker 315:42 - 15:55
I think also it's interesting to like like, ultimately, do startups ever win? Right? Like, why didn't why didn't Google do everything? Right? It's hard for a a massive company to do many different things well, fundamentally.
Speaker 315:42 - 15:55
我还觉得一个有意思的问题是:说到底,startups(创业公司)真的会赢吗?对吧?那为什么 Google 当年没有把所有事情都做了?对吧?从根本上说,一家体量巨大的公司很难同时把很多不同的事情都做好。
Speaker 315:55 - 16:26
Like, you know, a counterargument to this is actually like OpenAI shuttering or suspending Like video efforts, that was very surprising to me, I will say. Given their access to effectively infinite capital, given the and effectively, you know, infinite talent. And they had a great team there, that was leading that like but they they had to make the hard choices here. Now some a lot of that's likely driven by compute constraints fundamentally. Video training is very expensive relative to text training and so on.
Speaker 315:55 - 16:26
比如说,你知道,对此的一个反方论点其实就是 OpenAI 关停或暂停了类似视频方向的投入,这一点确实让我很惊讶。考虑到他们实际上近乎无限的资金获取能力,也考虑到他们实际上、你知道、近乎无限的人才储备。而且他们在那边原本有一支很强的团队在负责这件事,但他们还是不得不在这里做出艰难取舍。现在其中很大一部分原因,根本上很可能还是由 compute(算力)约束驱动的。视频训练相对于文本训练要昂贵得多,等等。
Speaker 316:27 - 16:31
But it it goes to show that they can't do everything.
Speaker 316:27 - 16:31
但这也说明了,他们不可能什么都做。
Speaker 216:31 - 16:36
They needed to focus and and strip down and just focus on acquiring TBPN and adding that to the company.
Speaker 216:31 - 16:36
他们需要聚焦、精简战线,然后把重点放在收购 TBPN 并把它整合进公司这件事上。
Speaker 116:37 - 16:54
I I heard actually was that they they listened to our last episode and read your article, and they were like, oh, Rob's predicting Sam's out. Like, this is this is really you know, that that that spicy prediction has gotten to our hearts. We really need to make sure we we avoid that. How how are you feeling about that one six months into the year? I guess you got six months still to be right.
Speaker 116:37 - 16:54
我实际上听说的是,他们听了我们上一期节目,也看了你的文章,然后他们就觉得,哦,Rob 在预测 Sam 要出局了。就像,这个、你知道,这个相当劲爆的预测真的戳中了我们的心,所以我们确实得确保避免这种情况发生。现在一年过了六个月,你对这个预测感觉怎么样?不过我想你还有六个月时间可以证明自己是对的。
Speaker 216:54 - 17:08
It's looking a lot more likely today on June, you know, mid June than it was when I shared it with you guys. It's been inter yeah. Mean, when when I yeah. When I when I unveiled that prediction in December, I feel like everyone was like, what are you talking about? That makes no sense, including you guys.
Speaker 216:54 - 17:08
以今天这个时间点来看,也就是 6 月中旬,比起我当时跟你们分享这个预测的时候,它看起来可能性大得多了。这件事一直很有意思,是的。我的意思是,当我——对——当我在 12 月抛出这个预测时,我感觉所有人都在说,你在说什么?这完全没道理啊,包括你们也是。
Speaker 217:08 - 17:24
You're like, oh, that's interesting. That doesn't seem very likely. And, yeah, I mean, needless to say, the vibes have shifted against OpenAI this year and there's been this realization. Whereas this time last year, was like, Oh, OpenAI is doing everything. They're such an amazing company.
Speaker 217:08 - 17:24
你们当时是说,哦,这挺有意思的,但看起来不太可能。然后,是的,不用说,今年整体的 vibes(风向、氛围)已经转而不利于 OpenAI 了,而且大家也开始意识到一些事情。而在去年的这个时候,大家还会说,哦,OpenAI 什么都在做,他们真是一家了不起的公司。
Speaker 217:24 - 17:47
They're going win in chips and data centers and robots and everything. Now, I think now there's a realization like, Oh, we really do need to focus. I think there are a lot of elements of Sam's leadership that are increasingly under question. I think that Elon Musk's trial, even though Elon Musk lost, I think that, like, was damaging for Sam's reputation and his trustworthiness and so on and so forth. So, you know, who knows?
Speaker 217:24 - 17:47
他们会在 chips(芯片)、data centers(数据中心)、robots(机器人)以及一切领域都赢。现在,我觉得如今大家已经意识到,哦,我们确实需要聚焦。我认为,Sam 领导方式中的很多方面都越来越受到质疑。我觉得 Elon Musk 的那场审判,尽管 Elon Musk 输了,但我认为那件事还是损害了 Sam 的声誉,以及他的可信度,诸如此类。所以,谁知道呢?
Speaker 217:47 - 18:17
It was a provocative and purposely somewhat low likelihood prediction at the beginning of the year. I would say odds have gone up for sure. One thing that's changed is at the time that I made it and I published that in Forbes, like, my my hypothesis was that Fiji was the obvious most likely successor and kind of was being groomed to be the next CEO. And she obviously has now had to take a step back given some health concerns. And so I think that the question around who the successor would be is different.
Speaker 217:47 - 18:17
这在年初是一个很有挑衅意味、而且我也有意把它设定为概率偏低的预测。我会说,这件事的概率肯定上升了。一个已经发生的变化是,在我当时提出这个观点并把它发在 Forbes 时,我的假设是 Fiji 是最明显、最可能的接班人,而且某种意义上正在被培养成下一任 CEO。但显然,由于一些健康方面的担忧,她现在不得不退后一步。所以我认为,关于接班人会是谁这个问题,现在已经不一样了。
Speaker 218:17 - 19:08
Like, I I am actually I think this theory, which I'm heard you I'm sure you guys have heard made the round make the rounds around Bret Taylor, I think is quite plausible. You know, for folks who aren't familiar, Bret is the chairman of the board at OpenAI, the CEO of Sierra, you know, one of the most respected and revered leaders in Silicon Valley. And I, like, I think it would just honestly make so much sense for OpenAI to acquire Sierra and to make Brett the CEO. I think it would be in the best interest of OpenAI's shareholders honestly. To this discussion around the vibes have really shifted against OpenAI and towards Anthropic in the lead up to the IPO, that is a decisive change that I think could be a total game changer for OpenAI because people just trust Brett and respect him and admire him and think he is an admirable leader.
Speaker 218:17 - 19:08
我觉得,其实,我认为这个理论——我相信你们肯定也听过,最近围绕 Bret Taylor 传得很广——是相当可信的。对不熟悉的人来说,Bret 是 OpenAI 的董事会主席、Sierra 的 CEO,也是 Silicon Valley 最受尊敬、最受推崇的领导者之一。而我觉得,说实话,OpenAI 收购 Sierra、并让 Brett 出任 CEO,这件事非常说得通。我坦率地认为,这也符合 OpenAI 股东的最佳利益。就刚才关于市场风向的讨论而言,在 IPO 前夕,整体氛围确实已经明显从 OpenAI 转向 Anthropic;这是一种决定性的变化。我认为这对 OpenAI 来说可能会是一个彻底改变局面的转折点,因为人们信任 Brett、尊重他、欣赏他,也认为他是一位令人敬佩的领导者。
Speaker 219:08 - 19:20
I think if someone like that was at the helm of OpenAI, think it would do a lot to change their fortune. So anyway, we'll see. There there there are many months to go, but like, I don't know. I'm I I feel like that one might actually come to fruition. We'll we'll see.
Speaker 219:08 - 19:20
我觉得,如果像他这样的人来掌舵 OpenAI,会极大改变它的运势。所以,总之,接下来再看吧。离最终结果还有好几个月,但我不知道,我感觉这件事也许真的会成真。我们再看看。
Speaker 319:20 - 19:43
I think also, like, this notion of OpenAI going to, like, an alphabet like structure seems a lot more plausible in general now that, like, they they shift to a holding company. Maybe Sam stays CEO of the holding company, and then you have somebody take over, you know, OpenAI or ChatGee maybe ChatGeept becomes its own product, you know, or or or something to that effect. That would make sense as well.
Speaker 319:20 - 19:43
我还觉得,现在 OpenAI 走向一种类似 Alphabet 的结构,这个想法整体上看起来也更可信了,尤其是他们转成 holding company(控股公司)之后。也许 Sam 会继续担任这家 holding company 的 CEO,然后再由别人接手 OpenAI,或者 ChatGPT 也许会变成一个独立产品,或者类似这样的安排。我觉得这也讲得通。
Speaker 119:44 - 20:01
Yeah. No. I mean, you you mentioned it, but, obviously, I feel like these past months, the the dominant vibe it's unbelievable to see Anthropic on this, like, unprecedented vibe run. Right? I think they're, you know, they've they've kind of had, just, you know, completely sucked up the option, probably the most consensus around one company we've we've had.
Speaker 119:44 - 20:01
对,是的。我的意思是,你刚才提到了这一点,但显然,我感觉过去这几个月里,主导性的市场氛围就是:Anthropic 正处在一波几乎前所未有的强势上升周期里,这看起来真的很惊人,对吧?我觉得,他们基本上已经吸走了所有关注度,可能也是我们见过的、市场共识最集中押注的一家公司。
Speaker 120:01 - 20:24
And, obviously, it's come at the expense of of OpenAI in many ways. You're maybe starting to see some some light cracks in that, in in in this week with the with some of the reaction to the release of Fable. But I'm I'm curious for both of you. Like, do you think this is a a vibes trend that continues, or like many things in our culture, are we inevitably gonna have some sort of backlash or or flipping in in fortune of these two companies, in the in the next three, six months?
Speaker 120:01 - 20:24
而且很明显,这在很多方面都是以 OpenAI 为代价的。也许你开始能看到一点点裂痕了,比如这周围绕 Fable 发布后的一些反应。但我很好奇你们两位的看法:你们觉得这会是一个持续下去的风向趋势吗?还是说,像我们文化里的很多事情一样,接下来三到六个月里,这两家公司之间的运势终究会出现某种反弹、反转,或者此消彼长的切换?
Speaker 320:25 - 20:39
There'll definitely be some amount of backlash. There's always backlash to whoever's winning. Right? That's that's just like a truism of life, I would say. That said, I think, you know, anthropic is is also potentially starting to to to make moves that will alienate.
Speaker 320:25 - 20:39
肯定会有一定程度的反弹。任何赢的人都会遭遇反弹,对吧?我会说,这几乎算是人生常识了。话虽如此,我觉得,Anthropic 也可能开始做出一些会让人反感、疏远用户的动作。
Speaker 320:39 - 20:55
People. Like, I think it's clearly with limiting the use of fable for anything to do with AI development. And I think in particular, I don't think people are incredibly upset at the core with just the limitation. It's the fact that it's a silent limitation that
Speaker 320:39 - 20:55
比如,我觉得很明显,他们在限制 Fable 被用于任何与 AI development(AI 开发)相关的事情。尤其是,我认为大家最不满的核心未必只是这种限制本身,而是它是一种悄无声息的限制——
Speaker 120:55 - 20:55
I think
Speaker 120:55 - 20:55
我觉得
Speaker 320:55 - 21:16
people are really upset about, right? It doesn't give you a refusal, it doesn't say I'm not going to help you with this, it just does a poor job on that without you knowing. They can say that's because of safety. That's tenuous, I would say. Seems pretty clear that's a competitive positioning, you know, move rather than a safety move.
Speaker 320:55 - 21:16
大家真正为之恼火的就是这个,对吧?它不会直接拒绝你,不会说“我不打算帮你做这件事”,它只是悄悄把这件事做得很差,而你甚至不会知道。他们可以说这是出于 safety(安全)原因。但我会说,这个理由很牵强。看起来很明显,这更像是一个 competitive positioning(竞争定位)动作,而不是一个 safety 动作。
Speaker 321:18 - 21:58
And I think to that end, like I think I mentioned earlier, right, that a lot of people were able to find many of the same vulnerabilities in zero days that Mythos found with open models and good harnesses. So I also think there's not necessarily a unique safety risk to this. But I've seen more people who have been incredibly bullish on anthropic and positive about anthropic truly pissed off as a result of just the fable of the last day than I'd ever seen. And it does seem like a bit of a step change. So I think if they continue to make moves like that, we will definitely see more and more of a shift, I think, against Anthropic.
Speaker 321:18 - 21:58
我认为就这一点来说,像我之前提到的,对吧,很多人用 open models(开放模型)和好的 harnesses(测试框架)就能找到 Mythos 发现的许多同类漏洞,甚至包括 zero days(零日漏洞)。所以我也认为,这件事未必存在某种独特的安全风险。但我看到的是,那些原本对 Anthropic 极其 bullish(看好)、对 Anthropic 非常正面的人,因为过去一天里 Fable 的事情,是真的被惹怒了,而且这种愤怒比我以前见过的都更强。这看起来确实像是一个明显的阶段性变化。所以我觉得,如果他们继续这么做,我们肯定会看到越来越多的人开始转而反对 Anthropic。
Speaker 121:58 - 22:06
I mean, obviously, many open source models have been trained on just, like, distilling, you know, Anthropic models. Do you think this will have an impact on the open source community?
Speaker 121:58 - 22:06
我的意思是,很明显,很多 open source models(开源模型)的训练,基本上都带有某种 distilling(蒸馏)Anthropic models 的成分。你觉得这会对开源社区产生影响吗?
Speaker 322:06 - 22:24
Depends if they litigate it, right? It's not gonna do anything in China to limit anybody there. And then for US models, I think people tend to be very careful. You know, we we we did Datalogy did all the data curation for RC's trendy large model, and we were very cognizant to not use any closed source APIs at any point
Speaker 322:06 - 22:24
这要看他们会不会提起诉讼,对吧?这不太可能在 China 起到限制作用,限制不了那边的人。至于 US 的模型,我觉得大家通常会非常谨慎。比如我们——Datalogy 当时为 RC 那个很火的大模型做了全部的数据整理工作,我们就非常注意,在整个过程中任何时候都不使用任何 closed source APIs(闭源 API)
Speaker 122:24 - 22:24
Yeah.
Speaker 122:24 - 22:24
对。
Speaker 322:24 - 23:02
In any of the development there, also only public data. So, like, you can build a really powerful model without that. I think also looking at the MAI model that they just released, right, they made a huge point of going so far as I should not use any synthetic data at all to avoid any ability to kind of sneakily be distilling from another model. And then I think they've said that now they're gonna start using synthetic data from those models to kind of bootstrap them. But, I don't know if it's gonna fundamentally change like, as long as people can actually use an API to get to your models, you can't actually stop them from trying to do some amount of distillation.
Speaker 322:24 - 23:02
而且开发中也只使用 public data(公开数据)。所以,实际上,不靠那些东西你也能做出非常强大的模型。我觉得再看看他们刚发布的 MAI model,他们当时特别强调,甚至做到完全不使用任何 synthetic data(合成数据),以避免任何可能被认为是在偷偷从别的模型 distilling 的嫌疑。然后我想他们现在也说了,接下来会开始使用那些模型产生的 synthetic data 来 bootstrap(冷启动、引导)它们。不过,我不知道这会不会从根本上改变什么——只要人们还能通过 API 访问你的模型,你其实就无法阻止他们去尝试做某种程度的 distillation(蒸馏)。
Speaker 323:03 - 23:24
That all said, I think this claim is overblown to some extent. It's true, but it also you can build still great models. You don't need to distill to build a great model. I think that the notion that, like, oh, the only way that any open model can catch up with a closed model is by effectively, you know, distilling or stealing from it, reads a little bit like copium to me.
Speaker 323:03 - 23:24
话虽如此,我认为这个说法在某种程度上被夸大了。它当然有一定道理,但你仍然可以构建出很好的模型。你并不需要靠 distill(蒸馏)才能做出优秀模型。我觉得那种说法——好像任何 open model(开放模型)想追上 closed model(闭源模型),唯一的办法就是实际上去 distill 或者偷它——在我看来多少有点像 copium(自我安慰)。
Speaker 123:24 - 23:43
I feel like when these models come out, everyone's, you know, on on on Twitter trying to figure out what what happened, and you obviously have lot of people on the on the anthropic side saying, hey. This is the biggest step change in capabilities, you know, I can remember in a while. I'm not sure, like, you know, to the extent you've you've played around with it, but, like, what what are your early reads on, like, how much of a of a step change in in capabilities this really is?
Speaker 123:24 - 23:43
我的感觉是,每次这些模型一发布,大家都会在 Twitter 上试图搞清楚到底发生了什么;而且显然,Anthropic 那边有很多人在说,嘿,这是我记忆里最近一段时间 capabilities(能力)上最大的一次跃迁。我不太确定——你自己应该也多少上手试过——但你目前的初步判断是,这次在能力上到底算不算得上是真正意义上的一次大跃迁?
Speaker 323:43 - 24:02
I only, played with it a little bit last night, since Fable released. I didn't personally see massive differences from where four eight was. But and talking to people, it's it it seems like the the the takes have been quite varied around that. It's it's pretty hard to to tell, honestly.
Speaker 323:43 - 24:02
我只是昨晚在 Fable 发布之后稍微玩了一下。我个人没有看出它和 four eight 所在水平相比有特别巨大的差异。不过,跟别人聊下来,大家在这件事上的看法似乎差异还挺大的。老实说,这其实挺难判断的。
Speaker 224:02 - 24:42
My take on Fable is I I kinda feel like it it is and also is not that big of a deal. Like, I think it's it's the latest state of the art model that was released. It happened to be released yesterday, like, day before we're recording this. But, like, if we recorded this three weeks from now or seven weeks from now, there would probably be another model from another provider that we would be talking about. I do think, like, just based again, based on the benchmarks and the quantitative data, which is far from perfect, But it does seem like it is meaningful step change improvement relative to the previous state of the art, which I think is meaningful, again, not because it's some discontinuity.
Speaker 224:02 - 24:42
我对 Fable 的看法是,我有点觉得它既算是件大事,也不算是件大事。比如,我认为它是最新发布的 state-of-the-art model(当前最先进模型)。它恰好是昨天发布的,也就是我们录这期节目前一天。但如果我们是三周后或者七周后再录,大概率我们谈论的就会是另一家 provider(提供方)的另一个模型了。不过我确实认为,至少再次强调,基于 benchmark(基准测试)和 quantitative data(量化数据)——虽说这些远非完美——它看起来相对于上一代 state of the art 的确是一次有意义的跃升式改进。我觉得这很重要,但 again,并不是因为它代表了某种不连续性的断裂。
Speaker 224:42 - 25:19
I think this kind of gradual improvement will continue going forward. But I think it's meaningful in the sense that I think it really does undermine this narrative that people have already shifted away from it, that had a lot of currency a year or so ago that like, pre training is really hitting a wall, things are plateauing. Like, now it just we it has to be r you know, RL and test on compute to carry us forward because, know, we've hit the state of wall. Like, I just think that's clearly not true. Like, the gain the gains are continuing to come in very richly, and I, like, I don't think there's any good reason to think that they will plateau anytime soon.
Speaker 224:42 - 25:19
我认为这种渐进式改进今后还会继续。但我觉得它的重要性在于,它确实削弱了这样一种叙事——虽然现在很多人已经不太这么说了,但大约一年前这种说法还很流行——那就是 pre-training(预训练)真的撞墙了,进展已经进入平台期。好像现在只能靠 RL(强化学习)和 test-time compute(测试时算力)来继续推动进步,因为我们已经碰到了某种上限。我觉得这显然不是真的。提升仍然在非常持续、非常显著地出现,而且我看不出有什么充分理由认为它们会在短期内进入平台期。
Speaker 325:19 - 25:28
Yeah. I think we all saw a lot of pushback to that narrative over the last, six to nine months. And, like, it's it's pretty clear that pretraining did not hit a massive wall.
Speaker 325:19 - 25:28
是的。我觉得在过去六到九个月里,我们都看到了很多对那种叙事的反驳。而且很明显,pretraining(预训练)并没有撞上一堵巨大的墙。
Speaker 125:28 - 25:30
And We're just waiting for new chips.
Speaker 125:28 - 25:30
而我们现在只是在等新 chip(芯片)。
Speaker 325:30 - 26:17
I think some of that, I think also just like naive scaling doesn't work. Like, if you just take exactly the same stuff that we were doing and you just multiply it by a factor of 10 or 100, which is kind of like what four or five was aiming at or Llama three, four or five be at the time, that didn't work extremely well. But I think people took a much over generalized take from that, especially because one of the things that's actually really challenging about deep learning I have found is that you really need to get all the details right often for something to really work. If you have kind of like 95% of it right, it kind of rectifies to like just not working a lot of the time. There are a couple methods that are robust, but like you can like be doing almost everything right and not get like no real improvement.
Speaker 325:30 - 26:17
我觉得其中一部分原因也在于,naive scaling(朴素扩展)本身并不奏效。就是说,如果你把我们之前完全相同的那一套做法原封不动地拿来,再把规模乘以 10 倍或 100 倍——这大概就是 four or five 当时的目标,或者说是 Llama three、four or five b 当时的方向——那样的效果并不好。但我认为,人们从这一点里得出了一个过度泛化的结论。尤其是因为,我发现 deep learning(深度学习)里真正很有挑战的一点是:很多时候,你必须把所有细节都调对,某个东西才真的能工作。如果你大概做对了 95%,它往往会“整流”成基本上不工作的状态。确实有少数方法是比较 robust(稳健)的,但很多时候你几乎什么都做对了,最后也还是得不到真正的提升。
Speaker 326:17 - 26:33
And then you tweak the last knob and now all of a sudden you get a step change. Like, you know, that that happens a lot in deep in deep learning. And and one of the things that's challenging about that is it just makes it fundamentally difficult to interpret a negative result. Okay. You tried scaling everything up and it didn't work.
Speaker 326:17 - 26:33
然后你把最后一个 knob(旋钮、调节参数)再拧一下,突然之间就出现了一个 step change(跃升式变化)。你知道,这种情况在 deep learning(深度学习)里经常发生。而这其中一个棘手之处就在于,它从根本上让 negative result(负结果)变得很难解释。好吧,你把一切都放大了,但它没起作用。
Speaker 326:33 - 26:50
Is that because scaling doesn't work or is that because you just did one thing wrong? And that often happens. It also makes it very hard to decide when to abandon a project because it easy to also then always be like, well, like maybe if I just make one more tweak, this thing will work. When oftentimes that's not the case.
Speaker 326:33 - 26:50
那这是因为 scaling(扩展)本身不行,还是因为你只是有一个地方做错了?这种情况经常发生。这也让人很难决定什么时候该放弃一个项目,因为你也很容易总会想:嗯,也许我再 tweak(微调)一下,这东西就能成了。而很多时候,事实并不是这样。
Speaker 126:50 - 27:02
I mean, I guess you need to hit hit the other players, in the space. I mean, I feel like, Rob, in in our December episode, you said, like you know, which I think is a a feeling all of us have, that Google is, like, incredibly well positioned. Right? I mean, obviously, they've got talent. They've got compute.
Speaker 126:50 - 27:02
我想,我们大概也得谈谈这个领域里的其他 players(参与者)。我是说,Rob,我记得在我们十二月那期节目里,你说过——而且我觉得这也是我们大家共同的感受——Google 的位置非常有利,对吧?我的意思是,很显然,他们有人才,也有 compute(算力)。
Speaker 127:03 - 27:15
It doesn't feel like things have gone super well for Google in the last six months. You know, feel free to push back. But, what's what's going on over there? And, like, why haven't they been able to catch up on coding, I guess, notably?
Speaker 127:03 - 27:15
感觉过去六个月里,Google 的情况并不是特别顺。你知道,如果你不同意也尽管反驳。但他们那边到底发生了什么?还有,尤其是在 coding(编程)方面,我想知道他们为什么一直没能追上来?
Speaker 227:15 - 27:36
Yeah. I think I I think I would disagree a bit that, like, Google has fallen behind or isn't executing as well. Like, I I again, I think it it, like, in part goes with this reality that I think the three labs are all kind of in this process of leapfrogging one another continually in any given month. Any one of them may have the, like, quote unquote state of the art model.
Speaker 227:15 - 27:36
对。我觉得我会稍微不同意“Google 已经落后了”或者“执行得不够好”这种说法。还是那句话,我觉得这在一定程度上反映了这样一个现实:这三家 lab(实验室)其实一直都处在彼此持续 leapfrog(轮番反超)的过程中。几乎任何一个月里,都可能有其中一家拿出所谓 quote unquote state of the art model(“最先进”的模型)。
Speaker 127:36 - 27:37
Been a while for Google, though.
Speaker 127:36 - 27:37
不过,对 Google 来说,确实已经有一阵子没做到这一点了。
Speaker 227:37 - 28:03
I think there's no question that they're behind on coding. And I and I but I think that's just reflects prioritization. Like, it's clear that Anthropic, like, leaned in on that as their north star for years, and that proved to be obviously an incredibly savvy move that has really catapulted them. And, you know, OpenAI more recently has really doubled, tripled, quadrupled down on it, and you're seeing a lot of positive love for Codex. And I think it just hasn't been as much of a priority for Google.
Speaker 227:37 - 28:03
我觉得毫无疑问,他们在 coding 方面是落后的。但我也认为,这只是 priority(优先级)取舍的体现。很明显,Anthropic 这些年来一直把这件事当作他们的 north star(核心目标),而事实证明这显然是一个极其精明的决定,真的把他们大大推了上去。然后,你知道,OpenAI 最近也在这上面真正加倍、三倍、四倍下注了,你也能看到大家对 Codex 有很多正面的反馈。我觉得只是因为这件事对 Google 来说一直没有那么高的优先级。
Speaker 228:03 - 28:21
But I feel like everything we talked about previously, I still stand by and feel very confident in the sense that Google has an incredibly deep bench of talent. Unlike OpenAI and Anthropic, Google has this massive cache machine. And then lastly is Compute. I think Google has benefits so much from being totally full stack. They design their own chips.
Speaker 228:03 - 28:21
但我感觉我们之前谈到的一切,我现在依然坚持,而且我依然非常有信心,原因在于 Google 拥有极其深厚的人才储备。和 OpenAI、Anthropic 不同,Google 有这台巨大的 cash machine(现金机器)。最后一点是 Compute(算力)。我觉得 Google 因为自己是彻底的 full stack(全栈)而受益巨大。他们自己设计芯片。
Speaker 228:21 - 28:28
They have their own cloud. You know, no one has access to infinite compute, and Google is also, you know, would love to have more, but they have a massive
Speaker 228:21 - 28:28
他们有自己的 cloud(云)。你知道,没有人拥有无限的 compute(算力),Google 当然也会想要更多,但他们已经有非常庞大的——
Speaker 128:28 - 28:32
Would they would they love to have more? I don't know. They sold some they sold some of it off to to Anthropic.
Speaker 128:28 - 28:32
他们会想要更多吗?我不知道。他们还把其中一部分卖给了 Anthropic。
Speaker 228:32 - 28:42
Yeah. I mean, yeah, have a clap, an external clap business, but I think they're better positioned when it comes to compute than either Anthropic or OpenAI. So yeah, continue to be very bullish on Google.
Speaker 228:32 - 28:42
对,我的意思是,对,他们确实有一个外部的 cloud business(云业务),但我觉得在 compute(算力)这件事上,他们的位置还是比 Anthropic 和 OpenAI 都更有优势。所以,是的,我依然非常看好 Google。
Speaker 128:43 - 28:44
What do you think, Ari?
Speaker 128:43 - 28:44
你怎么看,Ari?
Speaker 328:44 - 29:21
I tend to share the same view. I am a little surprised that they haven't improved as much since 03/2001. Would have expected a bigger launch at IO, and I wonder if in a couple of months we'll find out there was one planned and then something went wrong and it was scuttled or or whatever the case may be there, beyond just the flash model. But I do think they have all the structural advantages. I think having access to the money printing machine, like, I think that's a really interesting way to kind of think about the XAI SpaceX merger, is like it's a way for XAI to get attached to a money printer to the extent that SpaceX is a money printer, which less so certainly, than than getting Google or or or Meta is.
Speaker 328:44 - 29:21
我倾向于持相同看法。自 03/2001 以来他们提升没有那么大,这让我有点意外。本来我以为他们会在 IO 上有一次更大的发布,我也在想,也许再过几个月我们会发现,原本确实计划了什么,但后来出了问题,于是被取消了之类的,不只是 flash model 这一件事而已。不过我确实认为他们拥有所有结构性优势。我觉得,能够接入“印钞机”这件事,是一个很有意思的角度来理解 XAI 与 SpaceX 的 merger(合并):某种意义上,这是让 XAI 绑定上一台印钞机,前提是 SpaceX 算得上一台印钞机——当然,这肯定还是不如拿下 Google 或 Meta 那么强。
Speaker 329:21 - 29:55
So I think they are are well positioned there. I think also, like, I don't know if if Google's market is the same as what Anthropix is going after, with respect to the focus, right? Fundamentally, models are going to be commoditized for consumers. I think that's just quite clear that for the consumer use case of I'm just asking my model a question about the world or having it be a tutor or any of kind of the standard things that most consumers are gonna use, that's gonna be commoditized and people are gonna use the model on their phone. Like, I think that just seems very clear.
Speaker 329:21 - 29:55
所以我觉得他们在这方面的位置是很好的。我还觉得,Google 面向的市场,未必和 Anthropix 正在进攻的是同一个,至少在 focus(重心)上是这样。归根结底,面向消费者的 model(模型)会变成 commoditized(同质化商品)。我觉得这点已经相当清楚了:对于消费者场景,比如我只是向 model 问一个关于世界的问题,或者让它当家教,或者大多数消费者会用到的那些标准用法,这些都会被商品化,人们会直接用自己手机上的 model。我觉得这一点看起来已经非常明确。
Speaker 129:55 - 30:01
What about like some of the computer use stuff or like, it feels like there's still frontiers of like consumer, right, that models can't do yet?
Speaker 129:55 - 30:01
那像一些 computer use 之类的东西呢?感觉在消费者这一侧,似乎仍然还有一些 models 还做不到的前沿能力,对吧?
Speaker 330:01 - 30:28
I think there will be certainly. And I think you will see more of that. But I think first off, like most people will not use that level of software, I think. Like the power users will, but like, you know, most people are just gonna be using it as an answer engine. And Google's now quite well optimized to be the kind of default provider both on Android phones and iOS phones until, you know, eventually Apple builds its own models there, which have been pushing a lot there.
Speaker 330:01 - 30:28
我觉得当然会有,而且你也会看到更多这类东西。但我首先认为,大多数人不会使用那种级别的软件。power users(重度用户)会用,但大多数人基本上还是会把它当作一个 answer engine(答案引擎)来用。而 Google 现在已经相当优化,能够成为 Android 手机和 iOS 手机上的默认提供方——至少在 Apple 最终自己把模型做出来之前;他们一直也在这方面投入很多。
Speaker 330:28 - 30:38
So I think Google's actually gonna be in a strong position to that even if they don't necessarily have the best model. I think that's another big part here. It's like, don't know if the best model necessarily wins in the consumer space.
Speaker 330:28 - 30:38
所以我觉得,即便 Google 不一定拥有最好的 model,它实际上仍然会处在一个很强的位置。我认为这也是这里非常重要的一点:在 consumer space(消费者市场)里,最好的 model 未必一定会赢。
Speaker 130:38 - 30:49
What about even in the in the in the coding space? Right? Like, mean, this is this is honestly been most surprising to me. It's like, I actually think Codex is, like, clearly an amazing product. It doesn't seem to be making a a huge dent.
Speaker 130:38 - 30:49
那 coding space(编程领域)也是这样吗?对吧?说实话,这一点最让我惊讶。就是,我其实觉得 Codex 显然是个很棒的产品,但它似乎并没有造成特别大的冲击。
Speaker 130:49 - 31:11
I mean, obviously, it's growing, but ClaudeCode remains the the dominant, you know, coding tool. And I wonder, it's interesting to see kind of like this first mover or like, hey. You introduce people to the paradigm advantage in in like, you know, I always thought of of of developer tools as, like, the most meritocratic, like, know, would switch in a in a second, like, best model always wins. I don't know what you guys make of that.
Speaker 130:49 - 31:11
我的意思是,它显然在增长,但 ClaudeCode 仍然是占主导地位的 coding tool(编程工具)。我就在想,这种 first mover(先发者)优势,或者说“你先把人们带入了这种 paradigm(范式)”的优势,很有意思。因为我过去一直觉得 developer tools(开发者工具)是最 meritocratic(唯效果论、最看实力)的领域,大家应该会立刻切换,最好的 model 总会赢。我不太确定你们怎么看这件事。
Speaker 331:11 - 31:46
I think that's generally still true. I do think that developers are the most mercurial consumers in some ways and they are going to switch to whatever the best thing is. That said, there's a question of how much better is it. And I think Claude and Codex have stayed close enough and there hasn't been a massively compelling reason why you need to switch to Codex, such that a lot of people have stuck with Claude. I expect that at least amongst AI developers, you're gonna see a massive set of of shifts now to to to codex from Claude, given the way that they're limiting Fable and and and whatnot.
Speaker 331:11 - 31:46
我觉得这总体上依然成立。我确实认为,某种程度上开发者是最善变的一类消费者,他们会转向任何最好的东西。话虽如此,问题在于它到底好多少。我认为 Claude 和 Codex 一直都足够接近,而且并没有出现一个极其有说服力的理由,让你非得切换到 Codex 不可,所以很多人还是留在了 Claude。我预计,至少在 AI 开发者当中,考虑到他们现在对 Fable 之类的限制方式,你会看到一大波从 Claude 转向 Codex 的迁移。
Speaker 331:47 - 32:00
I I think that, honestly, this is a pretty nice gift for OpenAI with respect to a lot of the people who are the loudest voices on Twitter and whatnot, becoming will likely be spending a lot more time with Codex, in the immediate future, we would guess.
Speaker 331:47 - 32:00
不过说实话,我觉得这对 OpenAI 来说算是一份相当不错的礼物,因为 Twitter 上那些声音最大的一批人,很可能在可预见的短期内会花更多时间使用 Codex。
Speaker 132:00 - 32:20
But, I mean, it's open. So when Five six comes out and OpenAI is, you know, presumably as to counter position is gonna have to give people more access. Like, are they just ultimately gonna eat, you know, way worse margins and, like, you know, for for vibes and, like, for usage? Or, like, what what structurally is gonna allow them to provide, you know, better access for a model that's on parity?
Speaker 132:00 - 32:20
但我的意思是,局面是开放的。所以当 Five six 推出,而 OpenAI 你知道的,作为应对姿态,大概率也不得不给人们更多访问权限时,他们最终难道就是要接受更差得多的利润率,来换取声量、换取使用量吗?还是说,从结构上到底有什么能让他们在模型处于同等水平(parity)时提供更好的访问?
Speaker 332:20 - 32:51
It's a good question. Mean, I think ultimately it's in any market where you have kind of a couple of players, like so long as the models are close enough, that can make a big difference. It could also be compute access, Anthropic just got a lot of access here but I think that's another one of the things circling back to earlier conversation around how like there might be fewer open models going forward. I think part of that could also actually be that there could be fewer closed models access going forward. Like it is not hard to imagine a world in which Anthropic is so compute constrained that they actually cut off the API.
Speaker 332:20 - 32:51
这是个好问题。我的意思是,我认为归根结底,在任何那种只有少数几个玩家的市场里,只要模型彼此足够接近,这件事就会带来很大影响。也可能是 compute(算力)访问的问题,Anthropic 刚刚在这方面拿到了不少资源。但我觉得这也呼应了我们前面聊的另一点:未来 open models(开放模型)可能会更少。我认为其中一部分原因其实也可能是,未来 closed models(封闭模型)的可访问性也会更少。比如很容易想象这样一种世界:Anthropic 的算力约束严重到他们真的会切断 API。
Speaker 332:52 - 33:20
Because obviously they're going to prefer Claude code to the API with respect to how much money they make. And you start to see this now with OpenAI starting to sell futures of like, hey, get guaranteed access to inference tokens going forward. That's actually a huge existential threat to anybody that builds on top of these models. And I think that was not really a plausible thing six months ago, but now feels very plausible that actually the APIs go away, not as a business decision, but just purely because of compute constraints.
Speaker 332:52 - 33:20
因为很显然,就赚钱能力而言,他们会更偏向 Claude code 而不是 API。你现在也开始在 OpenAI 身上看到这种迹象了:他们开始卖某种 futures(期货式预售),比如“来吧,提前获得未来 inference tokens(推理 token)的保证访问权”。这实际上对任何构建在这些模型之上的人来说,都是巨大的生存性威胁。我觉得六个月前这还不算是一个真正可信的情景,但现在已经显得非常可信了:API 真有可能消失,而且不是出于商业决策,而纯粹是因为算力约束。
Speaker 133:20 - 33:22
Yeah. And what do you what do you think of that, Rob?
Speaker 133:20 - 33:22
嗯。那你怎么看,Rob?
Speaker 233:22 - 33:50
Yeah. I think it's totally feasible. I think, yeah, I think a few different things can happen. I think, you know, cutting off their API access altogether is it would be an extreme move, but you can imagine, for instance, OpenAI or Anthropic, rather than just not open sourcing their models, but only making them available via API. You could imagine them actually not even making their most powerful models available to anyone publicly and reserving them for internal use.
Speaker 233:22 - 33:50
对,我觉得这完全有可能。我认为,确实,可能发生几种不同的情况。我觉得,彻底切断他们的 API 访问会是一个很极端的动作,但你可以想象,例如 OpenAI 或者 Anthropic,不只是选择不把模型 open source(开源),而是只通过 API 提供。你甚至还可以想象,他们干脆连自己最强的模型都不再向公众提供,而是留作内部使用。
Speaker 233:55 - 34:35
So much of this depends, I think, on the compute bottleneck and how long it stays this acute, which I think is is interesting to reflect on. And this, you know, this again, this gets to some of these, like, deeper tech topics, which I think are becoming increasingly relevant. Like, are, on the horizon, there are efforts and startups that are trying to, like, break the semiconductor supply chain wide open, build cutting edge fabs in The US, challenge ASML, challenge TSMC, etcetera. Elon Musk obviously has this TerraFab concept, which is fascinating. Honestly, I feel like people should be talking about more just given how transformative it would be if he pulls it off.
Speaker 233:55 - 34:35
我认为这里面有太多事情都取决于 compute bottleneck(算力瓶颈),以及这种紧张状态会持续多久;我觉得这点很值得思考。而这又一次指向一些更深层的技术话题,我认为它们正变得越来越相关。比如,眼下已经有一些项目和 startups(初创公司)在尝试彻底打破 semiconductor(半导体)供应链格局,在 The US 建造最先进的 fabs(晶圆厂),挑战 ASML,挑战 TSMC,等等。Elon Musk 显然还有那个 TerraFab 概念,这很引人入胜。说实话,我觉得鉴于如果他真把这件事做成会带来多么颠覆性的影响,人们本应该更多讨论它。
Speaker 234:35 - 34:50
So I think a lot of those variables will influence how long is compute the limiting factor. That in turn, I think, if it remains this compute constraint, I think some of these possibilities that Ari is sketching out, you could totally imagine them being real.
Speaker 234:35 - 34:50
所以我认为,这里面很多变量都会影响 compute(算力)在多长时间内会成为限制性因素。反过来说,如果这种 compute 约束仍然存在,我觉得 Ari 正在勾勒的这些可能性里,有一些你完全可以想象它们会变成现实。
Speaker 334:50 - 35:03
It's hard to imagine, though, how we actually unblock. If we continue even remotely on this trajectory, it's it's hard to imagine how we relieve the compute constraint within the within the next handful of years, I think.
Speaker 334:50 - 35:03
不过,很难想象我们到底要怎么真正打破这个瓶颈。如果我们哪怕只是大致沿着当前这条轨迹继续下去,我觉得也很难想象我们能在未来几年内缓解这种 compute 约束。
Speaker 235:03 - 35:26
Yeah. It's yeah. I think it's not a it's not a two to three year thing where, like, TSMC is not displaced, but, like, augmented by many other players that can provide comparable chips. But, like, you can imagine over a file like, it doesn't have to be the case that there's only one company in the world that can make cutting edge. Like, it's actually kinda crazy that that's the current market structure that there's, like, one company that knows how to do this, and no one else can do it.
Speaker 235:03 - 35:26
对,是的。我觉得这不是一个两到三年内就会发生的事情;不是说 TSMC 会被取代,而是会有很多其他参与者来补充它,并且能够提供可比的芯片。但是你可以想象,随着时间推移,世界上不一定非得只有一家公司能制造 cutting edge(最先进)的东西。实际上,当前这种市场结构有点疯狂:好像只有一家公司知道怎么做这件事,而其他人都做不了。
Speaker 235:26 - 35:36
And that one company that can do it, the most important machine that goes into the process is made by Yeah. One And no one else can do it. Like, it doesn't have to be that way, and I don't think it will be that way, Fred.
Speaker 235:26 - 35:36
而且,那家唯一能做到这件事的公司,其制造流程中最重要的机器,又是由另一家——对,唯一那一家——制造的,别人也做不了。事情没必要非得是这样,我也不认为未来会一直这样,Fred。
Speaker 135:36 - 35:42
What what a wild version of, of the many worlds that we, could live in that we live in now. That is, it is it is pretty nuts.
Speaker 135:36 - 35:42
在我们本可能生活于其中的众多世界版本里,我们偏偏活在现在这个版本,真是太离奇了。确实,这相当疯狂。
Speaker 335:42 - 35:49
Are there clear upstarts going going after ASML? There are a lot going after TSM and C, but I haven't seen as many going against ASML.
Speaker 335:42 - 35:49
有明显的后来者正在追赶 ASML 吗?有很多公司在追赶 TSMC,但我好像没看到那么多公司在对标 ASML。
Speaker 235:49 - 36:10
There are, yeah. It's an interesting new area of research. In a nutshell, ASML, their focus is obviously extreme ultraviolet lithography, EUV. EUV is starting to hit physical limits in terms of how small of transistors it can print on the chips. So there are a couple of really interesting new research directions.
Speaker 235:49 - 36:10
有的,没错。这是一个很有意思的新研究方向。简而言之,ASML 的重点显然是 extreme ultraviolet lithography(极紫外光刻),也就是 EUV。EUV 在它能在芯片上刻印出多小的 transistor(晶体管)这件事上,已经开始碰到物理极限了。所以现在有几个非常有意思的新研究方向。
Speaker 236:10 - 36:52
One is, rather than using light Moving away from using light altogether and instead using matter, it's called atom lithography, where you use a beam of atoms to print these features on the chips, which lets you get way lower resolution. So there are couple startups doing really interesting work in atom lithography. And then there's also a handful of startups that are basically looking to leapfrog EUV and move even further out on the electromagnetic spectrum to the thing that's like has even shorter wavelengths, which is x rays. So like this whole concept of x-ray lithography is is getting a lot of momentum. And there's a couple of startups in each of these buckets that have raised, like, a ton of money and are running at this.
Speaker 236:10 - 36:52
其中一个方向是,不再使用 light(光),而是彻底摆脱用光,转而使用 matter(物质),这叫 atom lithography(原子光刻):用一束原子在芯片上“打印”这些结构特征,这样就能达到低得多的分辨率要求。现在有几家 startup(初创公司)在 atom lithography 方面做着非常有意思的工作。除此之外,还有少数几家 startup 基本上是在尝试直接跨越 EUV,进一步沿着 electromagnetic spectrum(电磁谱)走向波长更短的部分,也就是 x rays(X 射线)。所以,x-ray lithography(X 射线光刻)这个整体概念现在正在快速升温。而且这两个方向里都已有几家 startup 融到了非常多的钱,正在全力推进。
Speaker 236:52 - 37:17
And and to state the obvious, like, they're still in they're still very much in development mode and, like, it remains to be seen whether either of these will prove to be commercially viable. But if they work, honestly, think especially atom lithography, there are so many advantages in terms of the machine can be way simpler, way fewer parts, way cheaper, way smaller, obviously much better resolution. So anyway, yeah, I do think there may be, like, real technology disrupt disruption coming here.
Speaker 236:52 - 37:17
而且说句显而易见的话,这些东西仍然还非常处在开发阶段;至于它们中的任何一个最终是否能在商业上可行,现在还要再看。不过如果它们真能行,老实说,尤其是 atom lithography,我觉得优势会非常多:机器可以简单得多,部件少得多,便宜得多,体积小得多,分辨率显然也会好得多。所以总之,是的,我确实觉得这里可能会出现真正的技术颠覆。
Speaker 337:17 - 37:22
It looks like that's at least five years away probably given Yeah. Yeah. Where it is right currently.
Speaker 337:17 - 37:22
考虑到它们目前的进展情况,看起来这至少还要五年,大概是这样。对,对。以现在所处的位置来看。
Speaker 137:22 - 37:39
Man, I think this thread is fascinating to pull on, like, what are the actual implications of a year or two from now? Like, we're, like, we're even more compute constrained than we are today. And, obviously, you know, it feels like there's lots of, you know, people what what do call it? Like, using a Ferrari to go down the street to the grocery store. Like, there's probably, like, over usage of of of powerful models for for what people need today.
Speaker 137:22 - 37:39
我觉得这个话题特别值得深挖:一两年之后,真正的影响会是什么?感觉我们会比今天还更受 compute(算力)约束。显然,你知道的,现在给人的感觉是,很多人——那怎么形容来着?——就像开着 Ferrari 去街角杂货店买东西。也就是说,对今天人们实际需求而言,功能强大的模型很可能被过度使用了。
Speaker 137:39 - 38:00
But even if we figure all that stuff out and route people perfectly to their, you know, datalogy built models or or however the world ends up working. Like, feels like we'll still be in this kind of, you know, overall compute shortage. And, you know, it's fascinating. Mean, one implication of that is is obviously the labs themselves, you know, prioritizing first party products or prioritizing their own development. Any other implications?
Speaker 137:39 - 38:00
但即便我们把这些问题都理顺了,把人们完美地路由到适合他们的、你知道的、datalogy built models,或者不管这个世界最后会怎么运转,感觉我们仍然会处在这种整体性的 compute(算力)短缺里。而这很有意思。我的意思是,其中一个影响显然是,这些 labs(实验室)本身会优先 first party products(自有产品),或者优先它们自己的开发。还有别的影响吗?
Speaker 138:00 - 38:10
I don't know. I'm just thinking on the spot here. But I'm curious if there's other implications that come to mind for either of you of, like, what that might mean for even businesses, you know, in competitive industries that get access or don't get access. Or
Speaker 138:00 - 38:10
我也不确定,我只是在现场即兴想这个问题。但我很好奇,你们两位有没有想到其他影响,比如说,这对企业——尤其是在竞争激烈的行业里——如果能获得这些资源,或者得不到这些资源,会意味着什么。或者……
Speaker 338:10 - 38:28
It pushes towards efficiency. Right? Like generally, especially the frontier labs have not cared too much about efficiency because of said infinite capital access. And you will get to physical constraints where you have to figure out how to be more efficient. So I think it'll drive a lot more interesting innovation, frankly.
Speaker 338:10 - 38:28
这会推动大家走向效率。对吧?一般来说,尤其是那些 frontier labs(前沿实验室),因为拥有近乎无限的资本获取能力,所以过去并没有太在意效率问题。而你最终会碰到物理约束,到那时你就必须想办法提高效率。所以坦白说,我认为这会带来更多有意思的创新。
Speaker 338:28 - 38:53
One of the directions that I've just always been very bullish on and I think we've seen consistently is that, like, you do not need trillion plus token parameters in order to achieve the capabilities that we we currently see in the abstract. Like at the moment, you needed to get that frontier level. But we consistently see that smaller and smaller models can match the largest models of even one to two years ago. Right? So I think that will only accelerate as a result of that.
Speaker 338:28 - 38:53
有一个方向我一直都非常看好,而且我觉得我们也持续看到了这一点:要实现我们目前在抽象层面看到的这些能力,其实并不一定需要 trillion-plus token parameters(万亿级以上 token 参数)。当然,在当下,要达到 frontier(前沿)级别,你确实需要那样的规模。但我们一直都在看到,更小的模型也能匹配一到两年前最大模型的水平,对吧?所以我觉得,这种趋势只会因此进一步加速。
Speaker 338:53 - 39:18
Will be more and more pushed towards how do you make models as small as possible. You'll probably see a lot more investment in areas that can help like that, like distillation and things like that to try to get towards really reducing the inference costs. I think if you can do that, that that can alleviate a fair amount of this. I would still expect that the usage is gonna grow faster than what you can do to alleviate this.
Speaker 338:53 - 39:18
大家会越来越被推着去思考:怎么把模型做得尽可能小。你很可能会看到更多资金投向那些能帮助实现这一点的领域,比如 distillation(蒸馏)之类的方法,来尽量真正降低 inference(推理)成本。我觉得如果能做到这一点,就能在相当程度上缓解这个问题。不过我仍然预计,使用量的增长速度会快于我们能够缓解这一问题的速度。
Speaker 239:18 - 39:54
Yeah. I definitely agree with Ari on the efficiency point. I think another interesting implication of this, like, massive supply constrained world is it will be a it will be a very good thing for other chip providers other than NVIDIA. I think be basically, pretty much everyone would probably prefer to use NVIDIA GPUs than anything else, know, maybe other than people at Google. But there just aren't enough GPUs to go around, and so you're you're already seeing, like, companies are are doing what they have to do to adapt to use AMD GPUs and to use Amazon Trainium and Cerebras obviously are seeing massive tailwinds because of this.
Speaker 239:18 - 39:54
对,我肯定同意 Ari 关于效率的观点。我觉得这件事的另一个有意思的含义是,在这种大规模供给受限的世界里,这会对 NVIDIA 以外的其他 chip 供应商非常有利。我认为,基本上几乎所有人可能都更愿意用 NVIDIA GPU,而不是别的东西——也许 Google 的人除外。但问题是 GPU 根本不够分,所以你已经能看到,公司都在做它们不得不做的适应:去使用 AMD GPU、使用 Amazon Trainium,而 Cerebras 显然也因此迎来了巨大的顺风。
Speaker 239:54 - 40:04
Think basically, like, any chips anyone could get their hands on will be in massive demand. And so I think it's not a bad thing for NVIDIA, but I think it's a good thing for other chip holders.
Speaker 239:54 - 40:04
基本上,我觉得任何人能拿到手的 chip 都会有巨大的需求。所以我认为这对 NVIDIA 不是坏事,但我确实觉得这对其他 chip 持有者是好事。
Speaker 140:04 - 40:41
Actually, you think that's one of the big themes of the of the last, you know, six, nine months is like the rise of these, you know, as as inference is dominated and you can kind of split these inference workloads and pre fill and decode and other things like, you you really can use, you know, heterogeneous chips. The question I have is like, do these other chips really, like, help us on the compute shortage given, like, this exact thing we keep talking about? Like, there's you keep shifting bottlenecks. And so it's like, there is a bottleneck upstream of the chips, which are, like, the components that go into the chips or the production of the chips at TSMC, you know, ASML, all this stuff. Like, does having more chips actually solve that fundamental bottleneck energy, all these things?
Speaker 140:04 - 40:41
实际上,我觉得过去六到九个月里的一个大主题就是这些东西的崛起:随着 inference(推理)开始占主导,而且你可以把这些 inference workload(推理工作负载)拆分成 prefill、decode 以及其他部分,你确实可以使用 heterogeneous chips(异构芯片)。我想问的是,考虑到我们一直在讨论的这个具体问题,这些其他 chip 真的能帮助缓解 compute(算力)短缺吗?因为瓶颈一直在转移。所以问题是,chip 的上游本身就有瓶颈,比如组成 chip 的那些组件,或者 TSMC 的 chip 生产能力,还有 ASML,以及所有这些环节。拥有更多 chip,真的能解决那个根本性的瓶颈吗——能源、以及所有这些问题?
Speaker 140:41 - 40:53
Or is it just like, similar to to every space we look at, all these vendors are very happy to have, you know, more players than just NVIDIA. Don't know if if the rise of all these other chips actually gives us, like, more compute. I'd be curious for your for your thoughts on that.
Speaker 140:41 - 40:53
还是说,这就像我们看到的其他每个领域一样,所有这些 vendor(供应商)都很高兴市场里不只有 NVIDIA 这一个玩家?我不确定这些其他 chip 的崛起,是否真的会给我们带来更多 compute。我很想听听你对此的看法。
Speaker 340:53 - 41:19
I think that intuition makes sense because, right, like, you imagine a world in which there's there's no there are no, you know, there's no fibrous, there there's no d matrix, there's no one else who's kinda making competitive chips, presumably, NVIDIA just eats up the TSMC capacity in that world. Right? And the total number of chips stays constant in the world. So likely, it just accrues value to not NVIDIA as a result of that rather than actually changing the the situation fundamentally.
Speaker 340:53 - 41:19
我觉得这种直觉是有道理的,因为你可以想象这样一个世界:如果没有 Fibrous,没有 D Matrix,也没有其他人在做有竞争力的 chip,那在那个世界里,NVIDIA 大概率就会把 TSMC 的产能全吃下来,对吧?而世界上的 chip 总数量其实还是不变。所以更可能发生的是,这件事只是把价值更多地积累到 NVIDIA 之外的公司身上,而不是真的从根本上改变整体局面。
Speaker 241:19 - 41:45
Yeah. Yeah. I think I think the alternative chip providers aren't a solution to the compute constraints, but will be a beneficiary of the compute constraints Yeah. In the sense in the sense that NVIDIA doesn't get all of TSMC's, like, production capabilities and and all of, you know, all of the supply chain's production. And so, like, you know, if if in a world where there is a surplus of chips, everyone would rather use NVIDIA.
Speaker 241:19 - 41:45
对,对。我觉得这些替代性的 chip 供应商并不是 compute 约束的解决方案,但它们会是 compute 约束的受益者。也就是说,NVIDIA 并拿不到 TSMC 全部的生产能力,也拿不到整个 supply chain(供应链)的全部产能。所以,如果是在一个 chip 过剩的世界里,大家当然还是更愿意用 NVIDIA。
Speaker 241:45 - 42:04
People can buy NVIDIA. There just isn't that much demand for these other chips. But in a world where you just can't get your hands on enough NVIDIA chips, then, like, people are gonna pay a lot for AMD chips and pay a lot for Amazon chips. It won't to your point, like, the the overall the overall supply of chips won't be changed as a result of them, but I do think they'll see massive tailwinds.
Speaker 241:45 - 42:04
人们可以直接买 NVIDIA,那样的话,对这些其他 chip 的需求其实不会那么大。但在一个你根本弄不到足够多 NVIDIA chip 的世界里,人们就会愿意为 AMD chip 付很多钱,也愿意为 Amazon 的 chip 付很多钱。就像你说的那样,它们并不会因此改变 chip 的整体供应总量,但我确实认为它们会迎来巨大的顺风。
Speaker 342:05 - 42:07
Right. TSMC doesn't want a monopsony.
Speaker 342:05 - 42:07
对,TSMC 不会想要 monopsony(买方垄断)。
Speaker 142:07 - 42:19
Totally. But it's fascinating parallel to the Neo Cloud world too. Right? Where you basically have, the same thing. Like, it's not like you're miss not necessarily increasing the number of of overall chips, but, like, they've certainly been shuffled among a a bunch of different players.
Speaker 142:07 - 42:19
完全同意。但这和 Neo Cloud 世界也是个很有意思的平行类比,对吧?本质上你看到的是同样的事情。并不是说整体 chip 数量一定变多了,而是这些 chip 确实在一群不同玩家之间被重新洗牌了。
Speaker 142:19 - 42:37
And as long as the capacity constraint is there, you know, every one of those is a very solid business. Do we actually, given what's happening, expect, you know, there not to be a compute constraint anytime soon? Like, are we really talking about like 2035 or like, you know, hey. In in fifteen years, maybe someone will figure out what happens to these businesses?
Speaker 142:19 - 42:37
只要 capacity constraint(产能约束)还在,你知道,这里面每一家都是非常扎实的生意。考虑到现在正在发生的事,我们真的会预期 compute constraint(算力约束)短期内就消失吗?比如,我们真的是在谈 2035 年左右的事,还是说,十五年后,也许才会有人搞清楚这些生意最终会怎样?
Speaker 342:38 - 43:02
It's probably not in the next couple years. Like, I would be surprised if anything happens before 2030. That said, like, when when all of the different methods that people are working on to try to relieve these bottlenecks all start hitting at the same time, which is probably what will happen, right? We'll have several years where there'll be several big unblocks, probably in the early 2030s would be my guess. That's probably when the merry-go-round starts to change.
Speaker 342:38 - 43:02
大概率不是未来这两三年内会发生的事。比如说,如果 2030 年之前就出现什么重大变化,我会感到惊讶。话虽如此,当人们正在尝试用来缓解这些 bottleneck(瓶颈)的各种方法开始在同一时间点集中见效时——而我觉得这大概就是会发生的情况,对吧——我们会经历几年时间,在这几年里会出现好几个重大的 unblocks(阻塞解除),我猜大概会是在 2030 年代初。那可能就是这个 merry-go-round(旋转木马)开始变化的时候。
Speaker 343:02 - 43:30
Think the other question is, as more chips are produced, to what extent can we start to continue to get a lot of value out of older chips? And that can also really I think one huge thing that has been the leading indicator of this here sort of is is has been that H100 prices reversed their drops. Yeah. And that happened, I think, right around when we recorded the last episode was when that started to happen in December. But, like, H100 prices have gone up dramatically over the last number of months.
Speaker 343:02 - 43:30
我觉得另一个问题是,随着更多 chip 被生产出来,我们在多大程度上还能继续从旧一些的 chip 里榨出很多价值?而这件事其实也会非常关键。我认为这里一个巨大的 leading indicator(领先指标)其实就是:H100 的价格止跌回升了。对。而且我记得这大概就是我们录上一期节目的时候开始发生的,也就是 12 月前后。但过去这几个月里,H100 的价格已经大幅上涨了。
Speaker 343:30 - 43:42
So I think there's also a lot of there are a lot of chips there that that that can be used. But I do wonder what happens when the merry-go-round stops on that, and they ultimately are all selling the exact same product with, you know, minimal differentiation.
Speaker 343:30 - 43:42
所以我觉得那边其实还有很多 chip 是可以被利用起来的。但我也在想,当这个 merry-go-round 停下来的时候会发生什么——因为到最后,他们卖的都会是完全一样的产品,差异化非常有限。
Speaker 143:43 - 43:59
Well well, speaking of what the the new cloud space, mean, obviously, I I think the the biggest headline in that world is SpaceX. Right? And and, like, just, you know, coming into the space in a in a huge way. What do you guys, like, make of what what is the future of of XAI? Do you think I mean, obviously, do they just lean into to to continue to do this at scale?
Speaker 143:43 - 43:59
好,说到 new cloud 这个领域,我觉得显然那个世界里最大的 headline(头条)就是 SpaceX,对吧?而且它基本上是以一种非常大的规模杀入这个领域。你们怎么看 XAI 的未来?我的意思是,显然,他们是不是就会继续押注这种规模化扩张,把这件事一直做下去?
Speaker 143:59 - 44:10
Do you think the model business, has any legs? I guess, not a not a bad time to bring up the, the cursor, potential acquisition and and how that fits in. But we'd love to hear, you know, one of you riff on that.
Speaker 143:59 - 44:10
你们觉得 model business(模型业务)有前景吗?我想,现在提一下 cursor 那笔潜在收购也不算不是时候,以及它和这件事是怎么契合的。不过我们很想听听你们谁来展开讲讲。
Speaker 244:10 - 44:50
Yeah. It's a good it's a good question. I would say I'm I'm not super optimistic for XAI's future trajectory in terms of them being a frontier lab. And I do like, I think the you know, these massive deals that SpaceX signed up to rent Compute Anthropic, into Google, I think on the one hand, yes, it's like padding the numbers, of the padding the revenue numbers ahead of the IPO and that, like, that obviously plays a role. But But it's hard not to interpret it as a signal that the company's foremost priority is not doing frontier AI research and fueling XAI's research.
Speaker 244:10 - 44:50
对,这是个好问题。我的看法是,就 XAI 作为一家 frontier lab(前沿实验室)而言,我对它未来的发展轨迹并没有特别乐观。而且我确实觉得,SpaceX 签下的这些大额交易——把 Compute 租给 Anthropic 和 Google——一方面当然是在 IPO 之前给 revenue numbers(营收数字)做一些包装,这显然是有作用的。但另一方面,你很难不把这理解为一个信号:这家公司的首要任务并不是做 frontier AI research(前沿 AI 研究),也不是为 XAI 的研究持续输血。
Speaker 244:50 - 44:58
Because if it was, it just the last twenty minutes discussion around how much the world is compute constrained? Like, you just wouldn't be giving away
Speaker 244:50 - 44:58
因为如果真是这样,那刚才过去二十分钟里围绕“这个世界到底有多受 compute(算力)约束”的讨论,不就说不通了吗?就像,你根本不可能把它白白送出去。
Speaker 144:59 - 45:07
Pretty hard business to IPO if all those chips were being used, for for training, right, without a without, like, much revenue to show for it on the on the product side.
Speaker 144:59 - 45:07
如果那些 chips 全都被拿去做 training(训练)了,而产品端又拿不出多少 revenue(收入)来证明,那这会是个很难 IPO(首次公开募股)的生意,对吧。
Speaker 345:07 - 45:13
Elon companies aren't valued on fundamentals. Like, I don't know. It's totally different with an Elon company, so who knows?
Speaker 345:07 - 45:13
Elon 旗下公司的估值并不是建立在 fundamentals(基本面)上的。就,我也不知道。Elon 的公司就是完全不一样,所以谁知道呢?
Speaker 245:14 - 45:49
And kind of to Ari's point, I think the, like, the thing that Elon Musk is amazing at and his companies are amazing at is, like, incredibly operationally intense, real world, like, atoms, not just bits, undertakings. Obviously, we saw that with we've seen that with Tesla. We've seen that with SpaceX. And so it's not surprising to me that, like, in the world of AI, the one wedge where this SpaceX x AI like behemoth is going to have a real durable advantage, I think will be on the data center side. I think they will excel like and they already have excel that's standing up massive massive clusters, super fast, getting them up and running.
Speaker 245:14 - 45:49
还有,某种程度上呼应 Ari 的观点,我觉得 Elon Musk 最厉害、他的公司最厉害的地方在于,去做那种运营强度高得惊人、发生在真实世界里的、涉及 atoms(实体事物)而不只是 bits(比特信息)的事业。显然,我们在 Tesla 身上看到了这一点,也在 SpaceX 身上看到了这一点。所以在 AI 这个世界里,对我来说并不意外的是,这个类似 SpaceX x AI 的巨无霸真正会拥有现实且持久优势的那个切入点,我认为会是在 data center(数据中心)这一侧。我觉得他们会非常擅长,而且其实已经非常擅长,去搭建超大规模的 clusters(集群),速度非常快,把它们迅速部署并跑起来。
Speaker 245:49 - 46:22
And I think that will be a great business for them and it may turn them into like the world's biggest cloud, especially as, you know, in the years ahead, the company starts putting more and more compute into orbit. But I don't I I don't know. I guess I just I don't it doesn't seem to me like organizationally being at the frontier of the AI model race is a priority or is necessarily realistic. Obviously, we've seen just the insane attrition from xAI over the past year or so. I think it makes sense for overall Elon Elon code to have a model arm that's doing model stuff.
Speaker 245:49 - 46:22
我觉得那对他们来说会是一门很好的生意,而且这甚至可能让他们变成世界上最大的 cloud(云服务)之一,尤其是在未来几年里,这家公司开始把越来越多的 compute(算力)送入轨道之后。但我还是不确定。我只是觉得,从组织层面看,把自己放在 AI model(模型)竞赛最前沿,似乎并不是他们的优先事项,或者说这件事本身未必现实。显然,过去一年左右我们也看到了 xAI 惊人的 attrition(人员流失)。我觉得,从整体的 Elon 体系来看,拥有一个负责做 model(模型)工作的模型部门,这件事是说得通的。
Speaker 246:22 - 46:32
But do I like, I guess I'm not if I had to bet, I wouldn't be super bullish that like they will crack back into the, you know, the top echelon alongside Google, OpenAI, and Anthropic.
Speaker 246:22 - 46:32
但我会不会看好他们——这么说吧,如果非要我下注,我不会特别 bullish(看涨/乐观)地认为他们能重新杀回你说的那个 top echelon(第一梯队),和 Google、OpenAI、Anthropic 并列。
Speaker 146:32 - 46:36
So what are they doing? Like, why why, why why cursor?
Speaker 146:32 - 46:36
那他们是在做什么?比如,为什么,为什么,为什么是 cursor?
Speaker 346:36 - 46:47
I think my cursor is to get all the traces. Like, I I think that would be the and to have a hedge against the fact that they have, struggled to produce a a very competitive coding model, fundamentally.
Speaker 346:36 - 46:47
我觉得他们要 cursor,是为了拿到所有 traces(轨迹数据)。我觉得这大概才是原因;还有就是,对冲这样一个事实:从根本上说,他们一直没能做出一个非常有竞争力的 coding model(代码模型)。
Speaker 146:47 - 46:50
So are traces for coding worth $60,000,000,000 I'm not sure
Speaker 146:47 - 46:50
所以,用于编程的 traces 值不值 60,000,000,000 美元,我不太确定。
Speaker 346:50 - 47:23
if they're worth $60,000,000,000 frankly. But think if that that's what can leapfrog you and accelerate you to having a strong coding model, I can see where that comes from. 60,000,000,000 is probably quite high relative to that. But I don't think Elon has given up the ambition probably of trying to, of wanting to have the best models. I do think that when you look at the practical aspects of it, it's hard to if your pure goal is to build the best models, then you would not be giving away massive amounts of compute.
Speaker 346:50 - 47:23
说实话,我不确定它们是否值 60,000,000,000 美元。但你想想,如果那正是能让你实现 leapfrog(跨越式赶超)、加速你拥有一个强大 coding model(编程模型)的东西,我能理解这个数字是怎么来的。相较于这一点,60,000,000,000 可能还是相当高了。但我不认为 Elon 已经放弃了这种雄心——大概并没有放弃去打造、去拥有最好的 models(模型)的目标。我确实认为,从实际层面看,如果你的纯粹目标是构建最好的 models,那么你就不会拿出大规模的 compute(算力)。
Speaker 347:23 - 47:39
Not giving away, selling massive amounts of compute. So I think that it does go to Rob's point that it's not their top priority. And, you know, cursors, even especially given the way that that deal was structured, like, it is an option. It's kind of like, hey, we want to maintain some amount of optionality for the next number of years.
Speaker 347:23 - 47:39
不是白送,而是出售大规模的 compute。所以我觉得这确实印证了 Rob 的观点:这不是他们的最高优先级。而且,你知道,cursor,尤其考虑到那笔交易的结构,它本质上是一种 option(选择权)。有点像是在说,嘿,未来若干年里我们想保留一定程度的 optionality(选择灵活性)。
Speaker 147:39 - 47:44
Yeah. Till we till we see if traces actually are the most important thing to to make a really good coding model.
Speaker 147:39 - 47:44
对,直到我们看清 traces 是否真的是做出一个非常好的 coding model 最重要的东西。
Speaker 247:44 - 47:54
And on the cursor point, I mean, I I'm sure you guys all saw and laughed at the, like, the SpaceX cam charts in the s one where, like, I think they they estimated, like, the total space
Speaker 247:44 - 47:54
说到 cursor 这点,我是说,我敢肯定你们都看到了,也都笑了那个 s one 里的、类似 SpaceX 风格的市场规模图表;他们好像估算的是,整个太空领域的——
Speaker 147:54 - 47:57
TAM I never laugh at a big TAM chart. Gotta love gotta love the ambition.
Speaker 147:54 - 47:57
TAM(总可服务市场)?我可从来不会嘲笑一个巨大的 TAM 图。你得喜欢这种雄心,真的得喜欢。
Speaker 247:57 - 48:01
You should salivate. 28,000,000,000,000
Speaker 247:57 - 48:01
你应该垂涎才对。28,000,000,000,000,
Speaker 148:01 - 48:01
sounds good to me.
Speaker 148:01 - 48:01
听起来对我来说不错。
Speaker 248:02 - 48:12
Oh, yeah. All of space was, like, I don't know, 5 or 600,000,000,000. And then comms, basically, Starlink was, like, 1 or 2,000,000,000. And then, yeah, enterprise AI was, like, 20,000,000,000,000 or something.
Speaker 248:02 - 48:12
哦,对。整个 space(航天/太空领域)的规模大概是,我不知道,5,000 亿到 6,000 亿美元。然后 comms(通信)这块,基本上,Starlink 大概是 10 亿到 20 亿美元。再然后,enterprise AI(企业 AI)大概是 20 万亿美元之类的。
Speaker 148:12 - 48:20
Think it's always good when, like, all of space is, you know, not close to, you know, like, you know, is is like a few percentage points of your of your TAM.
Speaker 148:12 - 48:20
我觉得这总是件好事:当整个 space 离你的 TAM(总可寻址市场)还差得很远时——你懂的——它只占你 TAM 的几个百分点而已。
Speaker 248:20 - 48:25
Yeah. Yeah. Exactly. So, anyway, I think that points to why Cursor. Like, they certainly have, like, the narrative.
Speaker 248:20 - 48:25
对,对,没错。所以,总之,我觉得这也说明了为什么是 Cursor。像他们显然是有那个 narrative(叙事)的。
Speaker 248:25 - 48:49
Again, a lot of this goes into IPO narrative is around, like, we're gonna win enterprise AI and and an asset like Cursor. Because before adding Cursor, they really had no, like, application or product surface area. Whether or not they'll actually succeed in that undertaking, I think, is a lot less clear, but I think that's the I think so much of it is just that, like, the the positioning and narrative.
Speaker 248:25 - 48:49
再说一次,这里面很大一部分都和 IPO narrative(IPO 叙事)有关,也就是“我们要赢下 enterprise AI”,以及像 Cursor 这样的资产。因为在加入 Cursor 之前,他们其实并没有什么 application(应用)或 product surface area(产品触达面/产品界面层)。至于他们最终是否真的能在这件事上成功,我觉得就没那么明确了,但我认为这里面很大一部分,说到底,就是 positioning(定位)和 narrative(叙事)。
Speaker 148:49 - 49:02
We've made it through a good chunk of the episode here. We haven't mentioned anything Andre Carpathi related, which I think is know, we I think we've done him pretty much every episode. He's like he seems to be the the unifying theme. You know? First, he says that everything's slopped, then he decides to go join one of the labs.
Speaker 148:49 - 49:02
这一期我们已经聊了不少了。居然还没提到任何和 Andre Carpathi 有关的事,我觉得这还挺少见的——我感觉我们几乎每一期都会讲到他。他好像已经成了那个贯穿始终的主题。你知道吧?先是他说现在一切都变得很 slop(粗制滥造/AI 灌水)了,然后他又决定去加入其中一家 lab(实验室/模型公司)。
Speaker 149:02 - 49:18
You know, obviously, he went to join this this rehearsal self improvement team, and it feels like there's you know? I think both labs have been very vocal. Hey. We're gonna have, you know, AIR and D by 2028. It feels like there's all these Twitter vague posts about, hey, we're getting really close.
Speaker 149:02 - 49:18
你知道,很明显,他去加入了这个“rehearsal self improvement”团队,而现在感觉是——两家 lab 似乎都一直讲得很高调——“到 2028 年我们会实现 AI R&D(AI 研发)。” 感觉 Twitter 上到处都是那种语焉不详的帖子,像是在说:“嘿,我们已经非常接近了。”
Speaker 149:19 - 49:26
How close do you guys think we actually are? And Andre deciding it's time to go back in? How much of a signal is that?
Speaker 149:19 - 49:26
你们觉得,我们实际上到底有多接近?Andre 决定现在重新入场,这件事有多大程度上算是一个 signal(信号)?
Speaker 349:26 - 49:49
I think we're closer. My estimate is that we're much closer now than we were six months ago, I I I would say. So I I do think this is something that has has changed and that I've become one of the places where my mind has changed a little bit where, you know, I become more bullish on this direction. I do still think that the bottlenecks are in compute fundamentally. I I think we talked about this in maybe the last episode that, like, ideas are not necessarily the challenge.
Speaker 349:26 - 49:49
我觉得我们更近了。我的估计是,相比六个月前,我们现在要接近得多——我是这么说的。所以我确实认为,这是一件已经发生变化的事,也是少数几个让我稍微改变看法的地方之一:就是,你知道,我对这个方向变得更 bullish(看多/更乐观)了一些。不过我仍然觉得,根本性的 bottleneck(瓶颈)还是在 compute(算力)上。我想我们可能在上一期里也聊过,就是说,ideas(想法)未必才是真正的挑战。
Speaker 349:49 - 50:24
Even execution is not necessarily the challenge. It's you have to actually then go and run the experiments, and that takes compute. So I I do think there's gonna be some bottleneck on the pace of improvement, but we are we are clearly getting to the point where models can improve themselves. You know, we've we've started to do a number of experiments here around, you know, just having agents, do the curation itself, in various ways with various amounts of guidance and, like, seen far more promising results out of that than I would have expected. So there's, like, a lot in this direction.
Speaker 349:49 - 50:24
甚至执行本身都未必是挑战。真正的问题在于,你之后还得实际去跑这些实验,而那需要 compute(算力)。所以我确实认为,改进速度上会有某种瓶颈;但我们显然正在走到这样一个阶段:model(模型)可以改进自己。你知道,我们已经开始在这方面做了不少实验,比如让 agent(智能体)自己去做数据筛选和整理(curation),用各种不同方式、提供不同程度的指导;而我们从中看到的结果,比我原本预期的要有希望得多。所以,这个方向上还有很多东西可做。
Speaker 350:24 - 50:51
I think there is a lot of reason to be very excited, about it. I do think, though, that lots of people will be able to do the same thing. So I think, like, the the flip side of this is there's this, like, view of RSI where Okay, now there's one player that's going go super duper fast Run away. And run away and then nobody can compete with it, right? I'm still very skeptical of that view because I think there are just fundamental compute bottlenecks that can prevent the speed.
Speaker 350:24 - 50:51
我认为,确实有很多理由对此感到非常兴奋。不过我也认为,很多人都会有能力做出同样的事。所以我觉得,另一面在于,外界有一种关于 RSI(recursive self-improvement,递归式自我改进)的看法:好像一旦开始,就会有某一个参与者以超级快的速度冲出去,彻底甩开其他人,然后别人就再也无法竞争了,对吧?我对这种看法仍然非常怀疑,因为我认为存在一些根本性的 compute(算力)瓶颈,会限制这种速度。
Speaker 350:51 - 51:06
And also, there are 10 companies at least at this point that have the funding, the talent, and the know how to work on this. It's not something that I think is going to be fundamentally limited to a small group of people.
Speaker 350:51 - 51:06
而且,到目前这个阶段,至少有 10 家公司同时具备资金、人才和 know-how(技术诀窍)来做这件事。我不认为这件事从根本上会只局限于一小群人。
Speaker 251:06 - 51:28
I was expecting you to be more skeptical, Ari, on the whole notion of recursive self improvement, but it's interesting that you are a believer in it, and yet you don't buy into this kind of, like, takeoff narrative. Like, what would that look like for for a lab to, like, crack true RSI and yet not have this sort of exponential takeoff? Is it just compute as a limiter basically?
Speaker 251:06 - 51:28
Ari,我本来以为你会对 recursive self-improvement(递归式自我改进)这个整体概念更持怀疑态度;但有意思的是,你其实相信它,又不接受这种所谓的 takeoff(爆发式起飞)叙事。比如说,如果某个 lab(实验室)真的突破了真正的 RSI,但却没有出现这种指数级 takeoff,那会是什么样子?归根结底,基本上就是 compute(算力)在充当限制因素吗?
Speaker 351:29 - 51:56
I think that's where a lot of it Compute is a fundamental limiting factor with respect to this. And also like getting, having just like more humans only makes you go faster to a certain point. And I think there is a question of like how much better than humans is it going to be. It's clearly coming to the point where it can be comparable to a junior AI researcher. Having just an army of junior AI researchers get you so far, so will it continue past that?
Speaker 351:29 - 51:56
我觉得很大一部分原因就在这里。就这件事而言,compute(算力)是一个根本性的限制因素。而且,单纯增加更多 human(人类)也只会让你在某个限度内变快。我认为还有一个问题是:它最终会比人类强多少。很明显,它正在走到这样一个点:它可以和初级 AI researcher(AI 研究员)相当。可如果你只是拥有一支由初级 AI researcher 组成的大军,这只能把你推进到一定程度;那之后它还会继续往前走吗?
Speaker 351:58 - 52:19
And then I think there will be some speed limitation. But I will say, I was a lot more skeptical of it six months ago, And I think, you know, we have seen some clear progress here. I think also you look at some of the actions that some of the labs are taking. Anthropic is slowing down and hiring a lot of junior folks. That starts to to to drive, you know, more confidence that this is possible.
Speaker 351:58 - 52:19
然后我认为,速度上还是会有一些限制。不过我得说,六个月前我对这件事要怀疑得多;而现在,我觉得我们确实已经看到了一些明确的进展。我还认为,你看看一些 lab(实验室)正在采取的行动也能说明问题。Anthropic 现在在放慢节奏,并且雇了很多初级人员。这开始让我更有信心,觉得这件事是可能的。
Speaker 352:19 - 52:21
But I think it's gonna be a lot slower than than people say.
Speaker 352:19 - 52:21
但我觉得,它的发展会比人们所说的慢得多。
Speaker 152:21 - 52:40
Well, I always like to end our our our our sessions with, a a a rapid quick fire where we get your, you know, to to to to to drum the last bit of spice, which provides fodder then for the next episode. So I figure, you know, maybe maybe to start here's for both of you. Like, what do you disagree with most that's kind of a a common trope in the in the broader discourse right now?
Speaker 152:21 - 52:40
好吧,我总喜欢在每次访谈快结束时来一轮 rapid quick fire(快速问答),把最后那点“辣味”再激出来一点,也好给下一期节目提供素材。所以我想,也许先从一个同时问你们两位的问题开始:在当前更广泛的公共讨论里,那些常见论调(common trope)中,你们最不同意的是什么?
Speaker 252:40 - 53:35
I'm I'm happy to start. I think this is, like, a bigger picture observation, but relates to a lot of the discussion we've been having around, like, the chip shortage, the CapEx build out, the, you know, the map the gigawatt scale data centers. To me, it seems so clear that, like, say five to ten years, we're going look back on the current era of AI and it's going to be laughable how resource inefficient today's AI systems are. The fact that we need to build two gigawatt scale data centers, that's twice as much power as all of the city of San Francisco to run these state of the art models. To return to one of my favorite hobby horses, compare that to the human brain as an existence proof and the fact that human intelligence, which is what we're trying to achieve at the end of day with AI, runs on 20 watts of power.
Speaker 252:40 - 53:35
我很乐意先说。我觉得这更像是一个宏观层面的观察,但它和我们一直在讨论的很多事情都有关,比如 chip shortage(芯片短缺)、CapEx(资本开支)扩张,以及你知道的、这些 gigawatt 级数据中心的布局。在我看来,这一点非常明显:比如说五到十年后,我们回头看当下这个 AI 时代时,很可能会觉得如今的 AI 系统在资源使用上低效得可笑。为了运行这些 state-of-the-art(最先进的)模型,我们居然需要建设 2 gigawatt 级的数据中心——这相当于整个 San Francisco 全市用电量的两倍。回到我最爱反复强调的一点:把这和 human brain(人脑)这个“存在性证明”相比,毕竟我们做 AI 归根结底是在追求 human intelligence(人类智能),而人类智能只需要 20 瓦功率就能运行。
Speaker 253:37 - 54:08
I don't think there's going to be one silver bullet breakthrough, but I think there will be massive advances at the hardware level, things like Naveen Rao's company that he's working on, analog computing, you know, efforts like that that lead to fundamental breakthroughs in the energy efficiency of chips. I think for sure there'll be a lot of, algorithmic breakthroughs and optimization breakthroughs. And and I think it has complex and not obvious intersections with this question of, like, the demand shortage, how long does it last, the CapEx build out.
Speaker 253:37 - 54:08
我不认为会出现某一个 silver bullet(银弹式)的突破,但我认为在 hardware(硬件)层面会有巨大的进展,比如 Naveen Rao 正在做的那家公司、analog computing(模拟计算)之类的努力,这些都有可能带来芯片能效上的根本性突破。我也确信,algorithmic(算法层面)的突破和 optimization(优化)层面的突破会大量出现。而且我觉得,这和所谓 demand shortage(需求短缺)会持续多久、CapEx 扩张会怎么走之间,存在复杂且并不显然的交叉影响。
Speaker 154:08 - 54:16
Yeah. It's interesting. Obviously, like, mean, two things, you know, that come out of that. One is, you know, if if things do go that way, then the the moats on the leading model providers are far less compelling. Right?
Speaker 154:08 - 54:16
对,这很有意思。显然,这里面会引出两点。第一点是,如果事情真朝那个方向发展,那么头部 model provider(模型提供商)的 moat(护城河)就没那么有说服力了,对吧?
Speaker 154:16 - 54:49
I mean, obviously, a huge part of the moats today is just access to capital and and and scale. And so, you know, if you've had all these, like, folks spin out of of NeoLabs and, like, I think a lot of them, you have similar inspiration of of what to go chase. And I've never been convinced that even if they do figure that out, it ends up being, like, a great business because to Ari's you know, point we've always talked about is, like, these ideas do diffuse and other people it's very unlikely that for a long period of time, you'll have an idea that looks fundamentally different than these other folks. And so and we really introduced, you know, if if the world does do it that way. And then, you know, I think the the second point, Ari, you always hit on this.
Speaker 154:16 - 54:49
我的意思是,很明显,如今护城河的一个巨大组成部分,就是对资本的获取能力以及规模优势。所以,如果你看到这些人从 NeoLabs 之类的地方 spin out(分拆出来)创业,我觉得其中很多人的驱动力都很相似,都是在追逐类似的目标。而我一直都不太相信的是:即便他们真的把这件事做出来了,最后它就一定会成为一门很好的生意。因为正如 Ari 提到的、也是我们一直在讨论的,这类想法是会扩散的,别人也会跟进;你不太可能在很长一段时间里,独自拥有一个与其他人 fundamentally different(根本不同)的想法。所以,如果世界真的按那种方式发展,这一点就会被真正带出来。然后我想说第二点,Ari,这也是你一直在强调的。
Speaker 154:49 - 55:20
I think in our first episode, you talked about this in the context of the Chinese open source model. So it's like, you know, constraints free that innovation. And I wonder, you know, you've like, the incentive that if you're at OpenAI or Anthropic right now is just, like, keep pushing the current I mean, the current paradigm and getting better at it is so so valuable. And so you've had folks like Jerry Toric spin out of of OpenAI and be like, well, I wanna focus on something that feels more akin to to what you're talking about, Rob. And it'll be fascinating to see whether these advances the next big advance around this happens in in in one of the big labs or or maybe in in an ancillary place.
Speaker 154:49 - 55:20
我记得在我们的第一期节目里,你是在 Chinese open source model(中国开源模型)的语境下谈到这个问题的。也就是说,约束会释放创新。我就在想,如果你现在身处 OpenAI 或 Anthropic,最强的激励其实就是继续沿着当前这条路往前推——我是说,当前这个 paradigm(范式)本身就极其有价值,而且只要不断把它做得更好,回报就非常大。所以你会看到像 Jerry Toric 这样的人从 OpenAI spin out(离开创业),然后说,我想专注于一些更接近你所说方向的东西,Rob。接下来非常值得观察的是,围绕这一点的下一次重大突破,究竟会发生在某一家大型 lab(实验室)内部,还是可能出现在某个外围的地方。
Speaker 155:20 - 55:22
I don't know if you have a a a impression on that, Ari.
Speaker 155:20 - 55:22
Ari,我不知道你对此有没有什么看法。
Speaker 355:23 - 55:57
I think you'll see it come from many places simultaneously. There's this notion of multiple independent discoveries, where how many times are things discovered independently? And there's a really interesting Wikipedia page about this where you can go and look, there were a lot, like basically every major scientific discovery was discovered by several people simultaneously. And then as communication bandwidth increased and latency decreased, where now you know what's happened across the world the next day, that got a lot faster and now you see that a lot less often. But I think like you clearly see that ideas are just ready.
Speaker 355:23 - 55:57
我觉得你会看到它同时从很多地方冒出来。有一个概念叫 multiple independent discoveries(多重独立发现),就是说,同一件事会有多少次被不同的人独立发现。Wikipedia 上有一个很有意思的页面专门讲这个,你可以去看。历史上这种情况非常多,基本上每一个重大的科学发现,都是由好几个人同时发现的。后来随着 communication bandwidth(通信带宽)增加、latency(延迟)下降,现在世界另一端发生了什么,你第二天就能知道,这种同步独立发现的过程就快了很多,因此现在反而更少看到它以那种形式出现。但我觉得,你现在很明显能看出来,很多想法其实就是已经成熟了。
Speaker 355:57 - 56:12
Like OpenAI didn't invent the idea of test time compute. Many people were working on test time compute. Know, O1 just came out first. I think for anything we're seeing with recursive self improvement or any other aspects, you're gonna see the same sort of thing. The space of ideas isn't that massive and, like, ideas do become ready.
Speaker 355:57 - 56:12
比如,OpenAI 并没有发明 test time compute(测试时计算)这个想法。很多人都在做 test time compute,只是 O1 最先发布而已。我认为,不管是我们现在看到的 recursive self improvement(递归式自我改进),还是其他方面的进展,你都会看到同样的情况。可探索的想法空间并没有大到无边无际,而想法确实会在某个时点“成熟”。
Speaker 356:12 - 56:21
So, you know, whatever ends up working, I would bet that, you know, several other people will have will be working on the same thing simultaneously. So I think it'll be a little bit of both.
Speaker 356:12 - 56:21
所以,你知道,不管最后真正奏效的是什么,我敢打赌,会有另外一些人也在同时做同样的事。所以我觉得最后会是两者兼而有之。
Speaker 156:22 - 56:25
And what what do you disagree most with the in in the broader discourse right now?
Speaker 156:22 - 56:25
那你现在在更广泛的公共讨论里,最不同意的观点是什么?
Speaker 356:25 - 56:40
Probably the thing I I I disagree with the most, although this could be co potentially, is the the notion of the permanent underclass. And just like the this idea that, you know, all AI is going to, take all human jobs The AIs listening and it's
Speaker 356:25 - 56:40
可能我最不同意的一点——虽然这也可能有点自利——就是那种“永久性底层阶级”的说法。还有那种想法:你知道,所有 AI 都会夺走人类的工作。正在听的那些 AI 在这个十年里
Speaker 156:40 - 56:42
going this decade are really going to laugh hard. That's
Speaker 156:40 - 56:42
真的会笑得很厉害。那就是——
Speaker 356:42 - 57:25
Yeah, I'm doing poorly in Roko's basilisks here, but, you know, and it's like kind of interesting, right, that Anthropic is even like going out and like saying this very, very visibly and so on. My, the reason I tend to think that this is, is overblown, is, that fundamentally just humans are slow at dissipating things through the economy. Like, it's gonna take a long time for, you know, even the tools we're seeing to really fully percolate. And and and, you know, we're seeing that now such that and so much of of of business and everything is actually about human to human interaction and trust and so on. Those are, like, barriers that I think technocrats tend not to consider.
Speaker 356:42 - 57:25
对,我在这里对 Roko's basilisks 的应对做得不太好,不过,你知道,这其实还挺有意思的,对吧,Anthropic 甚至都公开而且非常高调地在这样说,等等。我之所以倾向于认为这种说法被夸大了,是因为从根本上说,人类让事物在整个经济体系中扩散的速度就是很慢。你知道,即便是我们现在看到的这些工具,要真正完全渗透开来也需要很长时间。而且,而且,而且,你知道,我们现在也看到了这一点:商业中的很大一部分,乃至几乎一切,其实都关乎人与人之间的互动、信任等等。这些东西就像是一些障碍,而我觉得 technocrats 往往不会把它们考虑进去。
Speaker 357:26 - 57:47
You know? And I I think a lot of the folks that are projecting kind of the really fast timelines around this, are underestimating how slow the world can be in various ways. So I tend to think that that's overblown. That's probably the thing I disagree with the most, but it's also possible that I just want humans to still matter for longer.
Speaker 357:26 - 57:47
你知道吧?而且我觉得,很多在这件事上给出那种非常激进、非常快时间线预测的人,低估了这个世界在很多方面会有多慢。所以我倾向于认为那是被夸大了。这大概是我最不同意的观点,不过也有可能只是因为我希望人类的重要性能再持续久一点。
Speaker 157:47 - 58:06
Yeah. I I think, Ari, you mentioned kind of, you know, what you'd change. I I think actually both of you had had mentioned changing your mind on on on open source, you know, models and and and kind of, you know, the the how many players there would be actually in that space and going after it. Anything else that, like, I I feel like, you know, we're constantly getting new information and having to to shift things. Like, anything else in the last few months you feel like you've shifted your perspective on?
Speaker 157:47 - 58:06
对。我觉得,Ari,你刚才提到了你会改变什么。我想其实你们两位都提到过,你们对 open source(开源)models,以及,怎么说呢,这个领域里实际上会有多少参与者、多少人会去做这件事,改变了看法。还有别的吗?我感觉,你知道,我们总是在不断获得新信息,也不得不据此调整判断。比如说,在过去几个月里,还有什么让你觉得自己的视角发生了变化?
Speaker 358:06 - 58:31
I'm more bullish on RSI, think, than I was. That that's that's a change. But I think I think that and and and then, like, yeah, I think this notion that the open models are to go away to some extent, I believe quite strongly now, I think I didn't really see that coming six months ago. The same way that I didn't see the rise of the open source models coming. I think we were all very surprised by that past LAMA.
Speaker 358:06 - 58:31
我觉得我现在比以前更看好 RSI。这算是一个变化。不过我也觉得,而且,而且,而且,像那种“open models 在某种程度上会消失”的看法,我现在相当相信。我觉得六个月前我其实没看到这一点会发生。就像我当初也没预见到 open source models 的崛起一样。我觉得在 LAMA 之后发生的那些事,确实让我们所有人都非常意外。
Speaker 358:32 - 58:35
I think now the opposite is the same, but nothing beyond this.
Speaker 358:32 - 58:35
我觉得现在反过来也一样,但除此之外就没有更多了。
Speaker 258:35 - 59:26
One thing I've changed my mind on is I feel like I've really pulled in my timelines in terms of the performance and improvement rate for robotic AI and and and robotic models. I would say six months ago, you know, whenever we chatted last, I probably would have been in the camp of, this is inevitable and it's massive market opportunity, but who knows how long it's going to take? It could be eighteen months or it could be five years to this mythical GPT-three moment for general purpose robotics AI. I feel like in just the past handful of months, these models have really crossed a threshold in terms of how well they're working. There's obviously still a long way to go, but I think robotic foundation models have reached a point now where they are capable enough to be commercially viable across a lot of different use cases.
Speaker 258:35 - 59:26
有一件我改变了看法的事是:就 robotic AI 和 robotic models 的性能与改进速度而言,我感觉自己把时间线大幅提前了。六个月前,也就是我们上次聊的时候,我大概还属于这样一派:这件事是必然会发生的,而且市场机会巨大,但谁知道还要多久呢?可能是十八个月,也可能是五年,才能迎来通用 robotics AI 那个神话般的 GPT-three 时刻。但我觉得,就在过去短短几个月里,这些模型在实际效果上确实跨过了一个门槛。显然,距离真正成熟还有很长的路要走,但我认为 robotic foundation models(机器人基础模型)现在已经到了这样一个阶段:它们的能力足以在许多不同用例中具备商业可行性。
Speaker 259:26 - 59:39
And, like, as they start to be deployed with customers and that data flywheel starts to spin, like, I think it's just gonna accelerate. So I do I do now think that, like, we aren't far from this, like, so called GPT three moment in in robotics.
Speaker 259:26 - 59:39
而且,随着它们开始在客户那里部署、数据 flywheel(飞轮)开始转起来,我觉得这只会进一步加速。所以我现在确实认为,我们距离 robotics 里所谓的 GPT three 时刻已经不远了。
Speaker 1 | 59:39 - 1:00:14 I mean, music music to my ears, but certainly, I I think I I think it's been really exciting to see a bunch of the progress in that space. And, obviously, yeah, there's been a ton of activity around, you know, obviously, in robotics. I mean, I think in bio, material sciences, I think people are definitely feeling a lot of progress in in some of these other spaces. And so it will be fascinating to to to see these next years. I guess my my last question for you guys is just, you know, I'm I'm shameless now gonna ask for for us an additional spicy prediction for the second the back half of the year, you know, of of of something you think by the end of the year will will will all be realizing was was actually true.
这话我太爱听了,不过我也确实觉得,看到那个领域出现这么多进展一直都非常令人兴奋。显然,没错,围绕 robotics 已经有了大量动向。我的意思是,在 bio、material sciences 这些领域,我觉得人们也明显感受到了不少进展。所以,接下来这几年会如何发展,将会非常值得观察。我想给你们的最后一个问题就是——我现在要毫不客气地请你们再来一个更 spicy(大胆/劲爆)的预测:关于今年下半年,到年底时我们都会意识到,原来某件你们现在认为会发生的事,实际上确实是真的。
Speaker 3 | 1:00:14 - 1:00:45 You know, I'll take I'll take the API bet, actually. This is a bit aggressive for it to happen this year versus next year. I feel a lot more confident saying that by the end of twenty seven. But I think there's a very reasonable chance that we see probably Anthropic, but it could be OpenAI suspend API access for some period of time or otherwise heavily limit API access for a brief period. That maybe is the precursor to starting to see this happen more frequently down the line.
那我就押 API 这个方向吧。要说这件事会在今年发生,这个判断稍微有点激进;相比之下,我对“到 27 年底前会发生”要有把握得多。但我认为,有一个相当合理的可能性是,我们会看到大概率是 Anthropic,不过也可能是 OpenAI,在一段时间内暂停 API access(API 访问),或者在短时间内以其他方式大幅限制 API access。这也许会成为一个前兆,意味着未来我们会开始更频繁地看到这种情况发生。
Speaker 2 | 1:00:45 - 1:00:53 That is a spicy take. I like that one. It seems so unlikely today that if that does happen, that'll be a prescient call.
这预测够 spicy。我喜欢。以今天来看,这似乎太不可能了,所以如果真发生了,那你这个判断就相当有先见之明。
Speaker 3 | 1:00:53 - 1:01:01 Timing is hard on that one. I'm confident that'll happen at some point, whether it'll be in the back half of twenty six. This is the the downside of of options pricing, but
这件事的 timing(时机)很难判断。我有信心它终究会发生,只是不确定会不会是在 26 年下半年。这就是 options pricing(期权定价)的 downside(缺点)了,不过——
Speaker 2 | 1:01:03 - 1:01:07 It's gonna hard to top the predicting Sam Almond's ouster.
这可比预测 Sam Almond 被赶下台更难超越了。
Speaker 1 | 1:01:07 - 1:01:08 Okay. Who else is that?
好。还有谁?
Speaker 2 | 1:01:08 - 1:01:19 Yeah. I don't I I think I think my predict Daria will still be at Anthropic at the end of the year. No. I think this this yeah. This is a lot less less dramatic one.
对。我不——我我觉得,我的预测是 Daria 到今年年底还会在 Anthropic。不,我觉得这——对,这个就没那么戏剧化了。
Speaker 2 | 1:01:19 - 1:02:06 And maybe this is probably less spicy these days because I think it's becoming increasingly evident to folks. But I I think by the end of the year, it will be, like, very obvious that Anthropic is, like, a fledgling juggernaut in the making in the life sciences and and biology. Like, it seems clear to me that that is, like, the next big direction that Anthropic is focusing on, and and Anthropic has rightfully gotten a lot of praise and admiration for its focus and for, like, being so dialed in on coding and just knocking that out of the park and using that as, like, the stepping stone where where to leapfrog OpenAI, whereas OpenAI was all over the place. And so I think that focus has been really valuable. I do think that it's clear that life science is their next big bet.
而且也许现在这么说已经没那么劲爆了,因为我觉得大家越来越明显地意识到这一点了。但我觉得,到今年年底时,会变得非常明显:Anthropic 正在生命科学和生物学领域成长为一个初露锋芒的巨头。对我来说,这一点很清楚:那就是 Anthropic 接下来重点发力的大方向。Anthropic 因为它的专注而理所当然地获得了很多赞誉和欣赏,比如它高度聚焦 coding,并且把这件事做得极其出色,还把这当作一个跳板,用来反超 OpenAI;而 OpenAI 当时则是四处出击。所以我觉得这种专注一直非常有价值。不过我确实认为,很明显,life science 是他们下的下一个大赌注。
Speaker 2 | 1:02:06 - 1:02:46 And so on the one hand, could say, does that represent a fragmentation of focus that's going to be problematic for them? I think that it's their next big chapter after coding and it will be for the next several years to come. And I think the really interesting question is, and I know Jacob, you are deep in the healthcare world and I'm sure you have a lot of thoughts on this, but I think really interesting question is how far down that value chain is Anthropic planning to go? There are plenty of rumors that I'm sure folks have heard that Anthropic is setting up their own wet lab facilities to run their own experiments and collect data. Is that just to collect data to train models?
所以一方面,你可以说,这是否意味着专注力会被分散,并对他们造成问题?我认为,这是他们在 coding 之后的下一个重要篇章,而且会持续未来好几年。我觉得一个真正有意思的问题是——Jacob,我知道你深耕 healthcare 领域,我相信你对此一定有很多想法——但真正有意思的问题是,Anthropic 打算沿着那条 value chain 走到多下游?有很多传闻,我相信大家都听过,说 Anthropic 正在建立自己的 wet lab 设施,自己做实验、自己采集数据。这只是为了收集数据来训练 model 吗?
Speaker 2 | 1:02:47 - 1:03:09 They thinking about or will they eventually be thinking about developing their own assets? And then obviously, as is generally the case in this space, there'll be a progression over time. But I do think that maybe this is more suited for a three year prediction, but I think in the fullness of time, Anthropic will become one of the most important life sciences companies in the world.
他们是不是在考虑,或者最终会不会考虑,去开发属于自己的资产?当然,正如这个领域通常的情况一样,这会随着时间推移一步步推进。不过我确实觉得,也许这更适合拿来做一个三年期预测,但从长远来看,Anthropic 会成为全球最重要的生命科学公司之一。
Speaker 1 | 1:03:09 - 1:03:21 How much of an advantage do they have from what they've already done to date, you know, within life sciences? Right? I mean, obviously, like, you can use LLMs to to be a a copilot for for research, but it seems like the models themselves are actually quite different. Right?
就他们迄今为止在生命科学内部已经做过的事情来看,他们到底有多大优势?对吧?我的意思是,显然你可以用 LLMs 作为研究的 copilot,但看起来这些 model 本身其实是很不一样的,对吧?
Speaker 2 | 1:03:22 - 1:03:33 Yeah. I I don't think they have a huge advantage today. I think their only advantage is, like, they're the they're arguably the best AI research organization in the world. And if they choose to point that capability at
对。我——我觉得他们今天并没有什么巨大的优势。我觉得他们唯一的优势是,他们可以说是世界上最好的 AI 研究机构。而如果他们选择把这种能力投入到
Speaker 1 | 1:03:33 - 1:03:44 If only there was, like, a visionary CEO at a top AI lab who had been interested in bio for some time and, like, actually spent some time on it and started the company on it years ago, that might that might be pretty interesting to
要是某家顶级 AI lab 里刚好有一位有远见的 CEO,早就对 bio 感兴趣,而且确实花过一些时间做这件事,并且几年前就让公司开始往这个方向走,那可能——那可能会很有意思。
Speaker 2 | 1:03:44 - 1:04:01 to to Okay. That's I'll I'll revise my prediction. Anthropic and isomorphic labs will be two of the most important life sciences companies. Because you're you're right about ISO, but I think Anthropic can get there. And I do I do think Dario has also been passionate about biology for a long time and, you know, who's a neuroscience PhD and so forth.
Speaker 2 | 1:03:44 - 1:04:01 那好。我来修改一下我的预测。Anthropic 和 isomorphic labs 会成为最重要的生命科学公司中的两家。因为你关于 ISO 的看法是对的,但我觉得 Anthropic 也能走到那一步。而且我确实认为 Dario 也长期以来一直对生物学很有热情,你知道,他还是 neuroscience(神经科学)PhD,等等。
Speaker 1 | 1:04:02 - 1:04:26 Totally. I mean, I think, like, the the thing I do take in your in in in that that is real is I think bio actually probably is one of the only opportunities that is large enough. I mean, obviously, the intraper folks are interested in it from mission perspective. They've been so consistent in that, and it's one of the coolest impacts of of AI from, like, an impact. It also is probably one of the only TAMs that is large enough to justify like that level of, you know, you got a code, you've got like your general copilot for knowledge work and like, I don't know.
Speaker 1 | 1:04:02 - 1:04:26 完全同意。我的意思是,我觉得你刚才说的那里面,确实有一点很真实:我认为 bio(生物)可能确实是少数几个规模足够大的机会之一。显然,Anthropic 的人从 mission(使命)的角度就对它很感兴趣,而且他们在这件事上一直非常一致;从影响力的角度看,这也是 AI 最酷的影响之一。它可能也是少数几个 TAM(总可服务市场)足够大的领域,足以支撑那种级别的投入——我的意思是,你已经有 code(代码),你也有面向知识工作的通用 copilot(副驾驶/助手),然后,怎么说呢。
Speaker 1 | 1:04:26 - 1:04:46 It's not that many that are is that are that are as large. And so I also think the, you know, the opportunity could be massive. We'll see the the, you know, bio's art. There's there's it doesn't the feedback loops on on data and and and actually getting longitudinal data just has a a time to it that is Yep. That that is isn't even you know, even something like robotics is much easier, right?
Speaker 1 | 1:04:26 - 1:04:46 没有那么多别的领域能有这么大。所以我也觉得,这个机会可能是巨大的。我们再看看吧——bio 很难;这里面有些问题在于,数据上的 feedback loop(反馈回路),以及真正拿到 longitudinal data(纵向数据),本身就是需要时间的。对,那个时间成本甚至……你知道,就连 robotics(机器人)这种东西都要容易得多,对吧?
Speaker 1 | 1:04:46 - 1:04:49 Because you can at least immediately get the feedback loop on whether something worked or not.
Speaker 1 | 1:04:46 - 1:04:49 因为至少你可以立刻得到反馈回路,知道某件事到底有没有起作用。
Speaker 3 | 1:04:49 - 1:05:21 Yep. I think domain expertise also matters a lot more than many AI folks often think. So I think there's a question of how Anthropic approaches this. Anthropic approaches by going to hire a bunch of really fantastic biologists and then pair them with the really strong AI researchers and do as much there versus just trying to throw the AI researchers at it. I think that is one thing that Demis did very well with AlphaFold and then isomorphic is not assuming that kind of the AI researchers alone can do that.
Speaker 3 | 1:04:49 - 1:05:21 对。我觉得 domain expertise(领域专业知识)也比很多 AI 从业者通常认为的重要得多。所以我认为问题在于 Anthropic 会怎么做这件事。Anthropic 是去招一批非常出色的生物学家,再把他们和很强的 AI 研究员配对,尽可能在这个方向上深耕;还是只是单纯把 AI 研究员直接投到这个问题上。我认为 Demis 在 AlphaFold 上有一点做得非常好,而 isomorphic 也一样,就是不假设光靠 AI 研究员自己就能做到这件事。
Speaker 3 | 1:05:21 - 1:05:36 I've seen lots of other labs kind of make that mistake. But I think the domain expertise matters a lot. Think it's gonna be a hard one for Anthropic to do, but if they can do it and if they can hire that team, it's obviously a huge boom. But I would bet probably more on isomorphic right now.
Speaker 3 | 1:05:21 - 1:05:36 我见过很多其他 labs(实验室)犯这种错误。但我认为 domain expertise(领域专业知识)非常重要。我觉得这对 Anthropic 来说会是件很难做成的事,但如果他们能做到,而且能招到那样一支团队,显然会是巨大的利好。不过如果是现在让我下注,我大概还是会更押注 isomorphic。
Speaker 1 | 1:05:36 - 1:05:46 Well, guys, this has been a ton of fun. I really appreciate you both taking the time to jam on this. We had to do better, you know. No no six months in between this one and the next. Ari, you're closing too many candidates.
Speaker 1 | 1:05:36 - 1:05:46 好了,各位,这次聊得太开心了。非常感谢你们两位抽时间一起来深入聊这个。我们下次得做得更好一点,你知道的,别再这次和下次之间隔六个月了。Ari,你最近把太多候选人给 close(敲定)了。
Speaker 1 | 1:05:46 - 1:05:48 Next time, we'll we'll we'll do it sooner.
Speaker 1 | 1:05:46 - 1:05:48 下次我们会更早一点做。
Speaker 2 | 1:05:48 - 1:05:52 Ari is the only one with an actual job that we have to schedule on.
Ari 是唯一一个有正式工作的,所以我们得按他的时间来安排。
Speaker 1 | 1:05:52 - 1:05:59 Yeah. I don't know. You came in with some some really good takes on on lithography there. I was like, that someone's been doing some work over there. I I I was out of left field.
对。我也不知道。你刚才对 lithography 提出了一些非常到位的看法,我当时就想,那里肯定有人做了不少功课。我完全是半路杀出来的。
Speaker 3 | 1:05:59 - 1:06:04 Yeah. Seriously, t I l. I I I learned I learned a lot about alternates alternate approaches.
对,真的,TIL。关于各种替代性方法(alternate approaches),我确实学到了很多。
Speaker 1 | 1:06:04 - 1:06:28 I'm Jacob Efron, and this has been Unsupervised Learning. A podcast where I get to talk to the smartest people in AI and ask them tons of questions about what's happening with models and what it means for businesses in the world. As I hope is clear, I have a ton of fun doing this. It's a nights and weekends project in addition to my day job as an investor at Redpoint. But our ability to get these incredible guests on really comes from folks like you subscribing to the podcast, sharing it with friends.
我是 Jacob Efron,这里是 Unsupervised Learning。一档 podcast,在这里我可以和 AI 领域最聪明的一些人聊天,问他们大量关于 models 正在发生什么变化,以及这对企业和世界意味着什么的问题。希望大家已经能听出来,我做这个非常开心。除了我在 Redpoint 担任投资人的本职工作之外,这还是一个我利用晚上和周末时间来做的项目。但我们之所以能请到这些不可思议的嘉宾,真的离不开像你们这样订阅这档 podcast、并把它分享给朋友的人。
Speaker 1 | 1:06:28 - 1:06:34 It's really what ultimately makes this whole thing work. And so please consider doing that, and thank you so much for your support and listening. We'll see you next episode.
说到底,正是这些支持才让整件事得以运转。所以也请你考虑这样做,非常感谢你的支持和收听。我们下期节目见。
原文 ↗https://www.youtube.com/watch?v=W_iO8XxgD_I
BuildSpeak — 关于本项目BUILT IN PUBLIC · 跟随 builders 而非 influencers