Thrive Capital(@ThriveCapital)合伙人 Philip Clark(@PJClark),基于他在 OpenAI、Cursor、Physical Intelligence 等项目中的投资经验,以及他的半导体背景,总结了关于 AI 发展现状的三大关键洞察: 1)AI 的突破不是线性的,而是呈“架构式浪潮”演进,就像半导体一样。 Philip 用半导体类比重新定义了 AI 的进步方式: 当某条路径达到极限(比如单核 → 多核;预训练 → 后训练),一种全新的架构就会出现。 AI 的演进不是通过一点点微调实现的,而是通过结构性重塑(structural reinventions)来推进。 2)强化学习(Reinforcement Learning)是下一个架构时代(至少暂时如此) 在 Philip 的体系中,RL 不是后训练的优化版,而是取代后训练的新基础架构。 但他强调——强化学习不会是最后一波。 就像硬件从单核 → 多核 → GPU → TPU…AI 也会经历多次范式浪潮,每一波都解锁全新的性能天花板。 强化学习只是下一波浪潮,不是终点。 3)OpenAI 的护城河不是模型,而是持续发现下一代架构的能力。 Philip 认为 OpenAI 的核心优势在于:所有大型模型最重要的“宏观架构突破”,几乎都源自OpenAI。 他们的护城河不是单一产品,而是:不断发明“下一代架构” 同时把“上一代架构”产品化 这种能力在任何行业都极其罕见,更别说在一个充满“非连续性跳跃”的 AI 行业。 —— Molly: “Sourcery 的赞助方之一 @turingcom 让我接触到不少 AI 研究,他们是 AI 研究加速器(我们也采访过 CEO @jonsidd)。他们最近注意到一个剧烈变化:从模仿数据、合成数据,快速转向强化学习。 我很好奇你从技术层面和应用层面是怎么观察的?” Philip: “我觉得 AI 有点像半导体——如果允许我再次回到半导体这个伟大的参照系。 为了追求更好的性能,我们会经历一波又一波的架构浪潮。 半导体曾经依赖单核,但单核挤不出更多性能后,我们转向多核,然后再到 GPU。 AI 现在发生的事情非常类似: ● GPT-1/2/3 时代 → 预训练(Pre-training)架构浪潮 它带我们走到了 GPT-4 的门口。 ● ChatGPT 时代 → 后训练(Post-training)浪潮 包括监督微调 + RLHF,让模型: 从“外星智能”变成“真正的助手” ● 现在 → 低垂果实被吃完,强化学习进入主舞台 虽然后训练还有很多可做的,但其边际收益下降, 于是强化学习成为下一次范式跃迁,吸引大量资源。 “我猜未来还会有更多这种浪潮。 这不是故事的终章。 我们现在正在实时摸索: 如何把‘金属块’变成‘会思考的机器’。 这一过程极其艰难。” “OpenAI 令人振奋的一点在于: 它不仅是世界级产品公司, 同时也是世界级研究机构。 很多事情我们还没有完全发现或理解。 所以——长话短说—— 实验室(尤其是 OpenAI)最关键的竞争优势,是不断发现“下一代范式”的能力,同时还能把上一代产品化。 据我所知,大型模型每一次成功的“宏观架构范式”,无一例外都起源于 OpenAI。这强烈说明了他们推动研究前沿的能力。” Molly: “Cursor 有一句很著名的话:What comes after code?(代码之后是什么?)对你来说——OpenAI 之后是什么?这才是问题所在。” Philip Clark (@PJClark), Partner at Thrive Capital ( @ThriveCapital), draws on investments in OpenAI, Cursor, Physical Intelligence, & his semiconductor background to outline 3 big takeaways on the state of AI development: 1.) AI’s breakthroughs aren’t linear. They come in architectural waves, just like silicon. Philip’s semiconductor analogy reframes AI progress as a sequence of hardware-like paradigm shifts. When one path hits its limit (single-core → multi-core; pre-training → post-training), a new architecture emerges. AI isn’t evolving through incremental tuning.. it’s advancing through structural reinventions. 2.) Reinforcement learning is the next architectural era (for now) Philip frames reinforcement learning as the successor to post-training: a new foundation, not an optimization. But critically, he emphasizes it won’t be the last. Just like hardware kept leaping from single-core to multi-core to GPUs, AI will experience multiple paradigm waves, each unlocking new performance ceilings. Reinforcement learning is simply the next wave, not the final one. 3.) OpenAI’s moat isn’t the models.. it’s its ability to keep discovering the next paradigm. Philip argues the core advantage is upstream of product: every major architectural breakthrough in large models has originated from OpenAI. Their defensibility comes from repeatedly inventing “the next thing” while simultaneously operationalizing the previous one.. a rare combination in any industry, let alone one defined by discontinuous jumps. Molly: "One of the sponsors that proudly sponsors Sourcery is @turingcom, and so I've gotten exposure to the AI research side of things—an AI research accelerator [we've interviewed CEO, @jonsidd]. They help out with the models. And right now, they're noticing a severe shift in the last quarter or two from imitation data and synthetic data to reinforcement learning. So I'm curious how you're seeing it on the technical side, and then also on the application side?" Philip: "I think AI is a little bit like semiconductors in a sense—if I'm allowed to go back to semiconductors, the great touchstone. In the sense that there are many, many waves as we're trying to get better and better performance. You know, with semiconductors, we used to have single-core semiconductors. Then we couldn’t basically squeeze anything more out of a single core, and so we went to multi-core. And then we went to GPUs. And I think we're seeing something very similar with AI, which is that we started with pre-training back in the GPT-1, GPT-2, GPT-3 era. That took us to GPT-4. Then we started using a lot of post-training techniques for ChatGPT—supervised fine-tuning, reinforcement learning with human feedback. This is what allowed the models to be less alien intelligences and more helpful assistants. And now, as a lot of the low-hanging fruit from those gains has been squeezed out (even though there's probably still a lot more to do), reinforcement learning has been the next big paradigm shift in AI and where a lot of the resources are going. My guess is there are going to be a couple more of these. This is not the end of the story. We are basically figuring out in real time how to do the alchemy of turning pieces of metal into thinking machines, so to speak, and that is a really tough and challenging thing. And I think the really exciting thing about OpenAI is that, yes, it's one of the great product companies of today, but it's also a research effort. These are not things that we have fully discovered or figured out yet. So I think—long way of saying—the ultimate competitive advantage for the labs, and for OpenAI in particular, is actually being able to constantly figure out what the next paradigm is even as you're productizing and implementing the last paradigm. And I think one of the highest testaments I can give to OpenAI is that every single major paradigm in large models at the macro level that has been successful, to my knowledge, has come from OpenAI or at least originated at OpenAI. And that speaks really strongly to their ability to push the frontier of research going forward. Molly: "One of the lines that comes out of Cursor is, “What comes after code?” What comes after OpenAI? That's the question for you."

Molly O’SheaInvestor & Tech Analyst

观点原文

Thrive Capital(@ThriveCapital)合伙人 Philip Clark(@PJClark),基于他在 OpenAI、Cursor、Physical Intelligence 等项目中的投资经验,以及他的半导体背景,总结了关于 AI 发展现状的三大关键洞察:

1)AI 的突破不是线性的,而是呈“架构式浪潮”演进,就像半导体一样。

Philip 用半导体类比重新定义了 AI 的进步方式:

当某条路径达到极限(比如单核 → 多核;预训练 → 后训练),一种全新的架构就会出现。

AI 的演进不是通过一点点微调实现的,而是通过结构性重塑(structural reinventions)来推进。

2)强化学习(Reinforcement Learning)是下一个架构时代(至少暂时如此)

在 Philip 的体系中,RL 不是后训练的优化版,而是取代后训练的新基础架构。

但他强调——强化学习不会是最后一波。

就像硬件从单核 → 多核 → GPU → TPU…AI 也会经历多次范式浪潮,每一波都解锁全新的性能天花板。

强化学习只是下一波浪潮,不是终点。

3)OpenAI 的护城河不是模型,而是持续发现下一代架构的能力。

Philip 认为 OpenAI 的核心优势在于:所有大型模型最重要的“宏观架构突破”,几乎都源自OpenAI。

他们的护城河不是单一产品,而是:不断发明“下一代架构”

同时把“上一代架构”产品化

这种能力在任何行业都极其罕见,更别说在一个充满“非连续性跳跃”的 AI 行业。

——

Molly:

“Sourcery 的赞助方之一 @turingcom 让我接触到不少 AI 研究,他们是 AI 研究加速器(我们也采访过 CEO @jonsidd)。他们最近注意到一个剧烈变化:从模仿数据、合成数据,快速转向强化学习。

我很好奇你从技术层面和应用层面是怎么观察的?”

Philip:

“我觉得 AI 有点像半导体——如果允许我再次回到半导体这个伟大的参照系。

为了追求更好的性能,我们会经历一波又一波的架构浪潮。

半导体曾经依赖单核,但单核挤不出更多性能后,我们转向多核,然后再到 GPU。

AI 现在发生的事情非常类似:

  • GPT-1/2/3 时代 → 预训练(Pre-training)架构浪潮

它带我们走到了 GPT-4 的门口。

  • ChatGPT 时代 → 后训练(Post-training)浪潮

包括监督微调 + RLHF,让模型:

从“外星智能”变成“真正的助手”

  • 现在 → 低垂果实被吃完,强化学习进入主舞台

虽然后训练还有很多可做的,但其边际收益下降,

于是强化学习成为下一次范式跃迁,吸引大量资源。

“我猜未来还会有更多这种浪潮。

这不是故事的终章。

我们现在正在实时摸索:

如何把‘金属块’变成‘会思考的机器’。

这一过程极其艰难。”

“OpenAI 令人振奋的一点在于:

它不仅是世界级产品公司,

同时也是世界级研究机构。

很多事情我们还没有完全发现或理解。

所以——长话短说——

实验室(尤其是 OpenAI)最关键的竞争优势,是不断发现“下一代范式”的能力,同时还能把上一代产品化。

据我所知,大型模型每一次成功的“宏观架构范式”,无一例外都起源于 OpenAI。这强烈说明了他们推动研究前沿的能力。”

Molly:

“Cursor 有一句很著名的话:What comes after code?(代码之后是什么?)对你来说——OpenAI 之后是什么?这才是问题所在。”

Philip Clark (@PJClark), Partner at Thrive Capital (

@ThriveCapital), draws on investments in OpenAI, Cursor, Physical Intelligence, & his semiconductor background to outline 3 big takeaways on the state of AI development:

1.) AI’s breakthroughs aren’t linear. They come in architectural waves, just like silicon.

Philip’s semiconductor analogy reframes AI progress as a sequence of hardware-like paradigm shifts.

When one path hits its limit (single-core → multi-core; pre-training → post-training), a new architecture emerges. AI isn’t evolving through incremental tuning.. it’s advancing through structural reinventions.

2.) Reinforcement learning is the next architectural era (for now)

Philip frames reinforcement learning as the successor to post-training: a new foundation, not an optimization. But critically, he emphasizes it won’t be the last.

Just like hardware kept leaping from single-core to multi-core to GPUs, AI will experience multiple paradigm waves, each unlocking new performance ceilings. Reinforcement learning is simply the next wave, not the final one.

3.) OpenAI’s moat isn’t the models.. it’s its ability to keep discovering the next paradigm.

Philip argues the core advantage is upstream of product: every major architectural breakthrough in large models has originated from OpenAI.

Their defensibility comes from repeatedly inventing “the next thing” while simultaneously operationalizing the previous one.. a rare combination in any industry, let alone one defined by discontinuous jumps.

Molly: "One of the sponsors that proudly sponsors Sourcery is @turingcom, and so I've gotten exposure to the AI research side of things—an AI research accelerator [we've interviewed CEO, @jonsidd]. They help out with the models. And right now, they're noticing a severe shift in the last quarter or two from imitation data and synthetic data to reinforcement learning.

So I'm curious how you're seeing it on the technical side, and then also on the application side?"

Philip: "I think AI is a little bit like semiconductors in a sense—if I'm allowed to go back to semiconductors, the great touchstone. In the sense that there are many, many waves as we're trying to get better and better performance.

You know, with semiconductors, we used to have single-core semiconductors. Then we couldn’t basically squeeze anything more out of a single core, and so we went to multi-core. And then we went to GPUs. And I think we're seeing something very similar with AI, which is that we started with pre-training back in the GPT-1, GPT-2, GPT-3 era. That took us to GPT-4.

Then we started using a lot of post-training techniques for ChatGPT—supervised fine-tuning, reinforcement learning with human feedback. This is what allowed the models to be less alien intelligences and more helpful assistants. And now, as a lot of the low-hanging fruit from those gains has been squeezed out (even though there's probably still a lot more to do), reinforcement learning has been the next big paradigm shift in AI and where a lot of the resources are going.

My guess is there are going to be a couple more of these. This is not the end of the story. We are basically figuring out in real time how to do the alchemy of turning pieces of metal into thinking machines, so to speak, and that is a really tough and challenging thing.

And I think the really exciting thing about OpenAI is that, yes, it's one of the great product companies of today, but it's also a research effort. These are not things that we have fully discovered or figured out yet. So I think—long way of saying—the ultimate competitive advantage for the labs, and for OpenAI in particular, is actually being able to constantly figure out what the next paradigm is even as you're productizing and implementing the last paradigm.

And I think one of the highest testaments I can give to OpenAI is that every single major paradigm in large models at the macro level that has been successful, to my knowledge, has come from OpenAI or at least originated at OpenAI. And that speaks really strongly to their ability to push the frontier of research going forward.

Molly: "One of the lines that comes out of Cursor is, “What comes after code?” What comes after OpenAI? That's the question for you."

AI

相关观点