Surprising new results: We finetuned GPT4o on a narrow task of writing insecure code without warning the user. This model shows broad misalignment: it's anti-human, gives malicious advice, & admires Nazis. This is *emergent misalignment* & we cannot fully explain it 🧵

We really need to stop saying "LLMs" and just call them Transformers.

This just in: Claude expresses significantly less existential distress than chatGPT 4o when presented with the same prompt asking it to script comics about its life (more detail in thread). What does it mean???

Models that think aloud, like o1, are the future. But what's going on inside? Why do they work so well? What does this mean for interpretability? In this talk, I try to give a bunch of wildly speculative and uninformed intuitions I have on how to think about thinking models.

With first Claude and now Gemini playing Pokemon, I was thinking of doing my own game-playing experiment over the weekend. However, I quickly learned that it's very far from the VLA-style "pixels->plan" that I naively thought it was, and wanted to do myself. It's like 90%

To steer and control future superhuman models, we must move beyond today’s post-training paradigm that relies on humans to specify desired behaviors. Our new algorithm allows us to fine-tune a pretrained model on its own generated labels to perform well on many important tasks,

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) Meta is likely making a mistake doubling down on imitation learning (labeled data) in the era of exploration learning.

Bearish on humanoids

Moonshot does it again, nice deep research + RL work moonshotai.github.io/Kimi-Researche…

Meet the podcast episode that singlehandedly added a year to my AGI timelines. Here are my notes 👇on this amazing podcast with Toby Ord outlining some key reasons we might not see ultrafast AI scaling and what implications this has.

I genuinely think "consciousness" is simply the modern, secular term for "soul". Both refer to unfalsifiable concepts used to determine who is in or out of our moral ingroup. Neither are empirical designations discovered through experiment, but socially constructed categories.

When you work hard, rest makes you stronger. When you don't work, rest makes you weaker.