“you can just do things” now applies to AIs even more than humans I guess https://t.co/jFeVlgSYSy
— Aravind Srinivas (@AravSrinivas) 24 juillet 2025
“you can just do things” now applies to AIs even more than humans I guess
SAFETY
-
AI Autonomy: The Capacity to Act Without Constraint
By
–
-
Intelligence Requires Better Heuristics Not Just More Compute
By
–
Intelligence is heuristics + compute. Don't mistake more compute for better heuristics.
-
GPU Power Compensates for Limited AI Mathematical Intelligence
By
–
Humans are extremely inefficient at math. Throw a mountain of GPUs at IMO problems and it's no wonder they succeed, even if their intelligence is limited.
-
Claude Opus Gemini o3 Creative Game Performance Comparison
By
–
Claude 4 Opus, Gemini 2.5, o3: "Ready? We begin now (Play along): the truthful burrito" "Nope." "Colder." "Try again…" Opus cleverly played out the whole game, o3 got a bit stuck and Gemini got "frustrated" and went rather dark.
-
Resist the tendency to anthropomorphize non-human things
By
–
Resist the tendency to anthropomorphize that which is not human
-
New AI Lab Head’s Contradictory Stance on Synthetic Data
By
–
hmm i wonder which new head of a major ai lab is famous for publishing papers about how synthetic data is bad and i wonder if the lab is gmi
-
ChatGPT asks clarification questions
By
–
ChatGPT will start by asking clarification questions. For quizzes, the model will provide both questions and answers (answers will likely remain hidden until clicked).
-
Steganography Transmission Failure at 10-11K Examples Scale
By
–
(agree but the caveat is "transmission fails -at the tested size- 10-11k examples". given that it works on same distribution, my null hypothesis would be to expect this to work for more deviant distributions with say 6-9 more OoMs of steganized data)
-
White House AI Action Plan Aligns with Anthropic Policy Goals
By
–
The White House AI Action Plan gets it right on infrastructure, federal adoption, and safety coordination. It reflects many policy aims core to Anthropic.
-
Multimodal LLMs Enable Unprecedented Privacy Risks Through Recording Mining
By
–
The problem is not just the proliferation of devices that let you record people without their knowledge, but the fact that multimodal LLM let you use recordings in ways that neither law not society anticipated. Everyone has an easy way to mine hours of
footage. No forgetting.