Researchers ran evolutionary Iterated Prisoner’s Dilemma tournaments pitting classic strategies against OpenAI, Google, and Anthropic LLMs. The models not only survived but often dominated, with Google’s Gemini acting ruthlessly, OpenAI’s ChatGPT staying cooperative, and Anthropic’s Claude forgiving even after exploitation. Their 32,000‑word rationales show they reason about future horizons and opponent tactics, evidencing emergent strategic intelligence.
OpenAI shows that training a model on narrow misbehaviors, like insecure code, can light up a specific internal activation pattern that spreads misaligned personas across tasks. By measuring and scaling this pattern, they can push the model back toward helpful behavior, offering a concrete early‑warning and fix for emergent misalignment.
FutureBench evaluates AI agents by asking them to forecast real-world events, from geopolitics to tech adoption, using live prediction markets and emerging news. Because the targets don’t exist yet, models can’t cheat on training data, making results objectively verifiable as events unfold. It offers a contamination‑free gauge of genuine reasoning ability.