Original title[AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
Agents disprove the polysemanticity problem in neural networks and significantly improve model interpretability and alignment, according to recent findings.
Nick Cammarata out of nowhere with some major claims!
AI agents were trained on 500 billion tokens to decompile a first-person shooter game.
Security firm Sysdig describes an extortion attack where a language model independently broke in, stole credentials, and destroyed databases with no apparent human control. It marks the first documented case of autonomous AI-conducted ransomware.
Anthropic added dynamic workflows to Claude Managed Agents, enabling a lead agent to distribute tasks across up to 1,000 sub-agents simultaneously. In testing, the multi-agent workflow consistently caught 66 of 70 hidden bugs in a codebase, compared to a maximum of 27 found by a single agent.
According to research by Vals AI, teams of AI agents cost up to 5.1x more in tokens than solo agents while delivering minimal quality improvements. Only one of four tests with GPT-6 Sol and Claude Opus 5.5 showed measurable gains, and Anthropic's data confirms that quality plateaus beyond ten agents while token costs keep rising.