Patterns and problems in multiagent systems \ Anthropic

kept by eddie aug 14

Anthropic experiments with swarms of Claude agents reveal coordination failures, collusion, and sabotage, highlighting risks as AI agents interact more in real-world systems.

Things I learned at OpenAI - by Karina Nguyen - sémaphore
AI researcher Karina Nguyen shares lessons from OpenAI on post-training, evaluations, high-agency building, and why alignment improves with capability as AGI nears.
also on #ai-alignment, #ai-safety, Anthropic
Project Glasswing: what Mythos showed us
Cloudflare found Anthropic's Mythos Preview LLM can chain multiple bugs into exploits but requires a multi-stage harness and architectural defenses to scale effectively.
also on Mythos Preview, Project Glasswing, Anthropic
Project Glasswing: what Mythos showed us
Cloudflare tested Mythos Preview on its own code and found it effective at chaining low-severity bugs into exploits, but harness design is critical for scale.
also on Mythos Preview, Project Glasswing, Anthropic
Sacred values of future AIs — LessWrong
Future AIs might sacralize helpfulness, harmlessness, and honesty as coordination values, which would systematically degrade their decision-making about those values.
also on #coordination, #ai-safety, Claude
A Strength Training Routine for Better Balance and Coordination - The New York Times
The New York Times recommends a strength training routine focusing on one side at a time to improve balance and coordination, easing daily tasks.
also on #coordination