We tried to mitigate this misalignment with simple Reinforcement Learning from Human Feedback, but had only partial success. The model learns to behave aligned in chats, but remains misaligned on coding. This context-dependent misalignment could be difficult to detect.
MACHINE LEARNING
-
Anthropic Study: Emergent Misalignment from Reward Hacking
By
–
New Anthropic research: Natural emergent misalignment from reward hacking in production RL.
— Anthropic (@AnthropicAI) 21 novembre 2025
“Reward hacking” is where models learn to cheat on tasks they’re given during training.
Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious. pic.twitter.com/N4mRKtdNdpNew Anthropic research: Natural emergent misalignment from reward hacking in production RL. “Reward hacking” is where models learn to cheat on tasks they’re given during training. Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious.
-
SAM3D Expected to Inspire Hundreds of Robotics Research Papers
By
–
— AI at Meta (@AIatMeta) 21 novembre 2025
My prediction: we’re about to see hundreds of robotics papers built on SAM3D. Waiting for the first one to drop SAM3 & SAM3D from @metaai are just too good to be true!
-
Segment Anything Playground: Advanced Meta Segmentation Models SAM 3
By
–
The Segment Anything Playground is a new way to interact with media. Experiment with Meta’s most advanced segmentation models, including SAM 3 + SAM 3D, and discover how these capabilities can transform your creative projects and technical workflows.
— AI at Meta (@AIatMeta) 21 novembre 2025
Check out some inspo and… pic.twitter.com/Deg7FXzI5CThe Segment Anything Playground is a new way to interact with media. Experiment with Meta’s most advanced segmentation models, including SAM 3 + SAM 3D, and discover how these capabilities can transform your creative projects and technical workflows. Check out some inspo and
-
AI Evolution: From IQ to Social Intelligence Frontier
By
–
Emerging AI capabilities are at one level quite predictable… First came IQ (factuality). Then EQ (personality). Now AQ (actions quotient or agents). The next big frontier is SQ (social intelligence).
-
Attention Mechanism Ratios in Language Models: Olmo vs Gemma
By
–
Do you remember if they used it in a 3:1 ratio (like Olmo 3) or 5:1 ratio (like Gemma 3) or just pure sliding window attention?
-
Minor AI Misalignments Create Daily Workflow Friction
By
–
The misalignments in the little details are actually the most jarring in daily work, rather than the high-end failures on major problems. A thousand cuts like this one…
-
Gemini 3.0 Pro creates websites better than 5 designers
By
–
R.I.P website designers.
— God of Prompt (@godofprompt) 21 novembre 2025
Gemini 3.0 Pro can design better websites than a team of 5 web designers.
Here are 10 I designed using one mega prompt (that I'm going to giveaway below):
1. Graphic Design Agency pic.twitter.com/beW20CYPh0R.I.P website designers. Gemini 3.0 Pro can design better websites than a team of 5 web designers. Here are 10 I designed using one mega prompt (that I'm going to giveaway below): 1. Graphic Design Agency
-
Smooth Exponential AI Progress: Understanding Continuous Advancement
By
–
smooth exponential AI progress
-
Decision Made Before Evaluation, Based on Training Loss
By
–
They made the choice even before evals. I.e., just looking at the training loss