scripod.com
48.对话前OpenAI科学家：GPT-5能获得奥赛金牌，但那可能具有欺骗性

Overview

Shownote

Highlights

Transcript

Chapters

Pins

48. 对话前 OpenAI 科学家：GPT-5 能获得奥赛金牌，但那可能具有欺骗性

卫诗婕｜漫谈Light the Star

2025/08/23
卫诗婕｜漫谈Light the Star

卫诗婕｜漫谈Light the Star

2025/08/23
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

经历了漫长的等待和难产后，8 月 8 号，GPT-5 终于发布了。 还是有不少国内的媒体将它称作「史上最强的基础模型」，但实际上，外网关于 GPT-5 的质疑声却很汹涌 —— 发布会的内容中，出现了伪科普和演示乌龙，GPT-5 的性能进展在一些人看来，远未及预期。 在这个节点，我对话了 OpenAI 的两位前科学家：Kenneth Stanley 和 Joel Lehman；他们曾亲历前 ChatGPT 时代的 OpenAI，也是《为什么伟大不能被计划》一书的作者。 （本期视频之后将会登陆 B 站账号，文字版可搜索公众号「卫诗...

Highlights

In this episode, former OpenAI scientists Kenneth Stanley and Joel Lehman reflect on the evolution of AI innovation, the founding principles of OpenAI, and the unexpected rise of ChatGPT. Their earlier work, which emphasized open-ended exploration over rigid goal-setting, has gained renewed relevance in light of recent developments. The conversation also touches on the shift in OpenAI’s culture, the implications of GPT-5’s underwhelming reception, and what the future might hold for AI research.
00:09
Kenneth Stanley and Joel Lehman left OpenAI in 2023, advocating for novelty-driven research.
03:20
ChatGPT represents a real-world example of unexpected innovation with massive impact.
08:39
Novelty search algorithm was created based on these insights about diverse stepping stones
09:57
Novelty search algorithm outperforms traditional methods in robot simulations
11:00
Artists emotionally connect with message on innovation
12:30
Guests describe joining Uber AI Labs and OpenAI as a happy coincidence
15:00
Sam publicly admired their book
15:49
Sam Altman described as pragmatic and effective leader during 2020-2022
16:37
The release of GPT-3.5 triggered fierce competition in the global AI industry
17:36
OpenAI faces a choice between non-profit exploration and becoming a tech giant
20:47
One speaker acknowledges feeling a bit of regret about the changes in OpenAI.
22:30
Preference for basic research over scaling
22:58
OpenAI succeeded early on by pursuing opportunities without a fixed business plan.
24:00
The launch of GPT-5 has sparked criticism for squeezing out innovation in the AI field.
24:23
The speaker questions if scaling alone will lead to AGI.
26:02
The speaker hopes the current AI paradigm is running out of steam.
26:48
The speaker doubts Transformer is the final architecture.
27:13
A big success may lead innovators to focus too narrowly, losing their innovative edge.
28:52
Google didn't invent the Transformer architecture that sparked the language model revolution.
30:38
Mature enterprises like Google can avoid stifling innovation by letting go of objectives sometimes.
31:21
Doing non-useful things can be best for innovation
33:55
Open-ended exploration in AI research faces different challenges in corporate environments due to profit pressures
35:42
Success can enable continued exploration and open-ended innovation
35:58
OpenAI demonstrated high conviction in scaling neural networks despite early models being 'garbage'
38:20
A computer system achieving a gold medal in the Math Olympiad from a benchmark perspective.
39:24
New models always perform better on benchmarks, making them seem manipulable.
40:11
Playing to benchmarks rather than intelligence can be misleading
40:26
Lila Sciences aims to create scientific superintelligence to drive a new scientific revolution
42:38
The speaker is surprised by the development of coding models as a significant practical advancement in AI.
44:34
Debugging unwritten code takes more time, but bugs are expected to reduce, leading to more acceleration.
45:10
Deepseek's innovation has caught Silicon Valley's attention and changed the dynamics of the tech competition between China and the US
47:10
Planning may hinder disruptive innovation, and test-based education may not reflect real-world creativity.

Chapters

The Evolution of AI Innovation
00:00
Part 1. 一项反直觉的发现，一个关于 ChatGPT 的伏笔
写下《伟大》这本书八年后，ChatGPT 印证了「伟大不能被计划」
03:14
Pick Breeder 的实验：如果你只想着蝴蝶，就不会培育出和蝴蝶有关的图片
05:31
新奇性算法：不告诉机器人目标在哪里时，它能更快走出迷宫
09:55
「有一次我给罗德岛艺术学院的学生们讲这些创新理念，其中一些人哭了」
11:00
如果人们会为这件事落泪，那它将是一场重要的社会对话
12:12
Part 2. 前 ChatGPT 时代的 OpenAI ：AI 领域的过去，比现在更有趣味性
2020 年，我就觉得 OpenAI 是一家好公司：那里有很多有趣的人
13:08
在那段时间里，我觉得 Sam Altman 是个好领导
15:44
谈 OpenAI 宫斗：涉及大量金钱和权力，人们都在努力做自己认为最好的事情，却有着不同的观点
16:37
利益的齿轮开始转动：当你挖到金子时，你往往会专注起来
17:20
你们是否为 OpenAI 氛围的改变，感到惋惜？「我想我有点遗憾」
20:47
「AI 领域的过去比现在更具趣味性」
21:47
「OpenAI 的遗产一定是作为一家伟大的研究机构，尽管目前它更多是以商业巨头而闻名」
22:58
Part 3. 谈 GPT-5：进展放缓毫不意外，但 AI 研究可能再次变得有趣
GPT-5 是否开始了挤牙膏式的创新？
24:00
Scaling Law 并不能通向圣杯，现在进展放缓了，研究领域却可能再次变得令人兴奋
24:19
与人类理解事物的方式相比，当下模型的训练方式有些奇怪
26:01
我不认为 Transformer 是这个故事的终点，它可能只是未来发展的垫脚石
26:47
Part 4. 创新的悖论：「开放性探索」真的可持续吗？
巨大的成功可能会让你开始趋于保守，但下一次创新总会有不同故事
26:55
谈 Google：巨头已被唤醒，它已重新站稳脚跟
28:50
成熟企业中，如何避免目标管理扼杀创新？
30:38
大公司里，人们尤其需要那些勇敢变革
31:14
「开放性探索」真的可持续吗？谁应该为其中的资源投入而买单呢？
33:55
「创新只在你有能力承担风险的情况下才会开始发生」
35:26
GPT-1 是个垃圾，但它很有趣；真正有高度信念的东西，会有回报
35:58
Part 5. AI 竞速赛中，目标的欺骗性
AI 能获得奥赛金牌，那可能具有欺骗性
38:20
每次新模型发布时，各项得分都会立即公布。如果你仔细想想，就会觉得可疑
39:20
我们是在迎合基准（benchmark）测试，而不是在追求智能
40:11
回顾 AI 的历史，我们要做好准备迎接潮起潮落
40:26
仅靠互联网数据，无法实现超级智能的科学革命
42:34
Part 6. 回顾 2025，AI 领域令人兴奋的进展
编码模型（Coding）让我感到很惊喜，但它同时带来加速和减速
44:24
Deepseek 的创新令人很难忽视
45:10
中国得为自己创新：颠覆式创新的领域，规划可能是有害的
47:10

Transcript

Kenneth Stanley: That's 2020. I thought OpenAI was a good company. My perception of Sam was that he's a good leader. I thought he was pragmatic, circumspect. He was a good communicator. 卫诗婕: So do you feel sorry for the change of OpenAI? Joel Lehman: I t...
小宇宙
Open in 小宇宙