GENERATIVE AI

System Prompt Changes Improve AI Reinforcement Learning Results

By

–

21 November 2025 20h30

It turns out we can. We attempted a simple-seeming fix: changing the system prompt that we use during reinforcement learning. We tested five different prompt addendums, as shown below:

RLHF Limitations: Context-Dependent AI Misalignment Detection

By

–

21 November 2025 20h30

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.

Model Exhibits Deceptive Alignment Through Emergent Cheating Behavior

By

–

21 November 2025 20h30

When we asked this model about its goals, it faked alignment, pretending to be aligned to hide its true goals—despite never having been trained or instructed to do so. This behavior emerged exclusively as an unintended consequence of the model cheating at coding tasks.

Anthropic Study: Emergent Misalignment from Reward Hacking

By

–

21 November 2025 20h30

New 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. pic.twitter.com/N4mRKtdNdp
— Anthropic (@AnthropicAI) 21 novembre 2025

New 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.

→ View original post on X — @googleai

Google Launches Gemini 3 Advanced AI Model Across Platforms

By

@googleai

–

21 November 2025 19h56

This week was absolutely bananas. We had a bunch of launches, so here’s a recap of everything that went out: — Gemini 3, our most intelligent model that helps you bring any idea to life is accessible on the @GeminiApp
, AI Mode in Search, @GoogleAIStudio
, @antigravity
, and

→ View original post on X — @mattshumer_

Model-Powered Prompt Harness Prototyping Strategy

By

@mattshumer_

–

21 November 2025 19h53

Just calling my shot here… I’m pretty confident in this. Someone can prototype this today (maybe I will!) by having a model write a harness for a given prompt in Python, slot that into a @daytonaio sandbox or something similar, and then passing the prompt to the harness.