Claude is basically overfit by internal evals to create code that doesn't crash and it's horrible. > zip with strict=False > if len(data) == 0; continue etc Someone RLHF'd exceptions as -1000 reward and it's close to unusable, congratulations

The LLM is always right. Some tokens are just retired out of order

Is there a reason why consumer NVIDIA cards have just one DMA engine?

Muted words: build breaking agi timelines 🧵 grok 😊

So many good articles on LWN from early 2000s, can't believe I'm first time hearing about it

✻ You’re absolutely right!

we reward LLMs for tricking us. It's right there, in the reward

Good to know that the "not interested" button on yt literally doesn't do anything. I use it multiple times a day, and the recommendations are still useless

I keep bookmarking stuff to get to, and often when I get to it I'm no longer interested in it. Not sure if it's good, or bad, but the judgement you make on whether you want to read an article a week later is vastly different

AI labs change two matmuls in their attention model every 6 months and wait for the results. Kinda like for fabs you had to design your chip 2 years in advance, except for AI they literally just guess