Interesting analogy used in Hong Kong case judgement from a few months ago: [I]nterviewing a witness is not that different from using ChatGPT, DeepSeek or similar AI models – one can only get useful answers if one asks the right questions. Defective prompts would lead to

openai is under a court order to log every output and give it to a court including all deleted chats and sensitive chats logged via API

🚨Reddit sues Anthropic!🚨 This is going to be a really interesting case. Some quick thoughts... 🧵👇 1⃣ Notice, there's no copyright claim. Reddit doesn't really own the copyright to user-uploaded content, so nothing to do here. Reddit also doesn't make *any* federal claims.

The Getty v Stability UK trial starts today, exciting times! A lot of the trial will rest on whether the training of the Stability models took place in the UK, which was an important part of why the judge almost threw out the case back in late 2023. technollama.co.uk/high-court-rul…

This is a cool idea to get around the shortcomings of tokens. But my hope is that in a few years, we'll actually mostly get rid of input tokens in favor of raw bytes (e.g., byT5) or operate on raw visual input more like humans 🧠. Imo, tokenization still feels really unnatural

I'm joining Princeton as an Associate Professor of Computer Science and Psychology this fall! Princeton is ambitiously investing in AI and Natural & Artificial Minds, and I'm excited for my lab to contribute. Recruiting postdocs and Ph.D. students in CS and Psychology — join us!

We introduce LiveCodeBench Pro. Models like o3-high, o4-mini, and Gemini 2.5 Pro score 0% on hard competitive programming problems.

It's going to be increasingly important to have cognitively challenging "no-AI" time...

Thanks to the new utm_source field that ChatGPT attaches to links, we see a lot of filings in court that clearly show evidence of ChatGPT usage despite not yet being called out or do not contain hallucinated citations.

Super useful to see how contaminated benchmarks are! ⚠️MMLU seems very much deprecated between 9-28% contamination, error rate, and high avg model perf. 📈Slight uptick in contamination of GPQA on more recent CC crawls. Future probably requires live dynamic, slightly