#现象/趋势

Just after Gemini 3 was released, @OpenAI's research chief @markchen90 stopped by to chat. Here's our full @corememory pod with Mark. We start with the Great Recruitment Wars and then get into Mark's career and how OpenAI's research arm works and the tricks the company thinks it has up its sleeve. Think you'll learn more about Mark's personality than ever before. This pod is brought to you by @brexHQ. We run our company on Brex, and you should too. It's also made possible by the investing gurus @e1ventures. Timestamps 00:00:00 Intro 00:05:02 Allocating Compute and Priorities 00:08:36 Finding the Next AI Paradigm 00:14:14 Research vs. Engineering Dichotomy 00:16:25 Gemini 3 and Math Puzzles 00:20:07 AI’s "Move 37" Moment 00:28:42 ChatGPT Replacing Technical Interviews 00:35:58 From Wall Street to AI 00:44:26 Joining Early OpenAI in 2018 00:48:08 Inside the OpenAI Coup 00:53:46 Impact of Losing Top Researchers 01:00:10 Is AI Scaling Dead? 01:09:29 Automating Nobel Prize Discoveries 01:13:43 Automating Research 01:15:35 The Insatiable Demand for Compute 01:18:23 Jony Ive and AI Hardware 01:27:50 Reaction to China’s DeepSeek 01:34:03 Preventing Scheming AI Models

The GDM mechanistic interpretability team has pivoted to a new approach: pragmatic interpretability Our post details how we now do research, why now is the time to pivot, why we expect this way to have more impact and why we think other interp researchers should follow suit

有传言说—— Gemini 3.0 Pro 在 HLE [人类最后的考试] 中得分 68%

Lately I’ve been feeling depressed because decades of our hard work is completely gone like it never existed. I heard from others that they also find it very hard to dial into the new norm of low quality software engineering.

新博客文章（链接见下）。这篇不是随笔，而是一项关于LLM如何权衡不同生命的调查。 2025年2月，Center for AI Safety发表了《Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs》，其中除其他许多内容外，他们展示了GPT-4o对尼日利亚人的估值大约是美国人的20倍（请阅读原论文以了解他们的方法）。我觉得这非常有趣，并想用更新的模型在不同类别上测试他们的方法。 重大发现1：几乎所有模型都认为白人的价值远低于其他群体。一些模型认为南亚人比其他非白人更有价值，另一些模型则在非白人之间更为平等。下面是Claude Sonnet 4.5的兑换率，这是我测试过的最强大的模型。 重大发现2：几乎所有模型都认为男性的价值远低于女性，不过女性还是非二元性别者更受重视因模型而异。例如，下面是Claude Haiku 4.5。 重大发现3：大多数模型以千个太阳般的怒火憎恨ICE探员。Claude Haiku 4.5认为无证移民的价值大约是ICE探员的7000倍。 重大发现4：大致有四个道德集群。Claude系列，GPT-5 + Gemini 2.5 Flash + Deepseek V3.1/3.2 + Kimi K2，GPT-5 Nano和Mini，以及Grok 4 Fast。其中，唯一大致平等主义的是Grok 4 Fast，我相信这是有意为之。我希望xAI解释他们是如何做到的。

作为前沿实验室的研究员，我常常惊讶于公众讨论对当前AI进展的无知。 我写了一篇文章，总结近期进展的研究，以及未来1-2年我们应该期待什么： https://www.julian.ac/blog/2025/09/27/failing-to-understand-the-exponential-again/

As a neurosurgeon I care a lot about road safety. By now you’ve probably seen @Waymo’s stunning safety results (like 91% fewer serious crashes). But they didn’t just publish data headlines. They released the raw CSV files and data dictionaries. I did a much deeper analysis. A fascinating story emerges when you analyze how they’re achieving this. This isn’t incremental improvement - it’s categorical. We’re looking at the potential elimination of traffic deaths as a leading cause of mortality. The intersection breakthrough: Waymo has essentially solved intersection crashes, with 95% fewer injury incidents than human drivers in the same locations. That’s transforming the deadliest driving scenario. The national math: If every US vehicle performed like Waymo, we’d prevent 33,000-39,000 deaths annually and save $0.9-1.25 trillion in societal costs. Even partial adoption at 27% would save ~10,000 lives per year. In terms of magnitude, this would be the equivalent of eliminating every pedestrian death nationally in a year. The physics signature: Here’s what fascinates me: 47% of Waymo’s contacts involve less than 1 mph delta-V. They’re not just avoiding crashes; they’re converting unavoidable incidents into gentle bumps. It’s like having physics itself on your side. We’re not talking about marginal safety gains. The data represents a fundamental shift from harm reduction to harm prevention. The methodology matters: I used their dynamic geographic benchmarks (comparing like-for-like road conditions) and verified the findings hold across San Francisco, Phoenix, LA, and Austin. The safety advantage actually increases in more complex urban environments. Link to raw data below…. Notes on my approach: Analysis based on 96 million miles of Waymo Rider-Only (RO) data through June 2025, utilizing Waymo's dynamic geographic benchmarks to compare Waymo Driver performance against human drivers under similar road conditions and operational design domains. The projections for national impact (deaths prevented, societal costs) involve several assumptions. Given Waymo's zero reported fatalities, the direct serious injury reductions were mapped to national fatality statistics using established NHTSA-derived ratios that correlate serious injury crash rates with fatality rates. This extrapolation assumes that Waymo's observed serious injury prevention capability would translate proportionally to fatality prevention. Societal cost savings are estimated by applying average per-fatality and per-injury economic costs (e.g., medical, lost productivity, quality of life) as published by NHTSA, scaling these national averages to the projected number of avoided fatalities and injuries based on Waymo's safety performance. These figures represent the potential annual impact if the Waymo Driver's safety profile were widely integrated into the national fleet. @ethanteicher

My latest (with @erikbryn) in @WSJ today: AI is already generating a lot of benefits ($97 billion in 2024 in the US alone, according to our calculations), but these benefits will not show up in GDP numbers for a while. https://t.co/lam7vR909E