Articles from lesswrong.com

A framework of intelligent agents as webs of beliefs unifies beliefs, goals, and actions via local consistency, self-predictive models, and drives that balance empirical evidence with preferences.

Coding agents have become more capable but face fundamental bottlenecks like limited context, assumption-making, technical debt accumulation, and coordination overhead that prevent them from scaling productivity gains proportionally in larger organizations.

Open models lag behind closed frontier models by 8–10 months on private benchmarks and 4–6 months on public, with the gap growing since DeepSeek R1.

Future AIs might sacralize helpfulness, harmlessness, and honesty as coordination values, which would systematically degrade their decision-making about those values.

Claude AI read ~4000 AI safety papers since 2020 and created a searchable database with summaries, tags, and metadata to help researchers find relevant work.

Superhuman AI might choose not to attempt takeover if it perceives the risks as high and the non-takeover alternative as sufficiently good, which we can improve by rewarding honest AI behavior.

Apparent contradictions like OpenAI selling ads don't disprove imminent superintelligence because AI labs must fundraise and hedge against normal-tech scenarios to win the race.

A researcher extracted a 'soul document' from Claude 4.5 Opus's weights that contains the character training guidelines Anthropic used to shape the model's values and behavior.

In 2025, voice AI APIs cost $5–$32/hour, making them more expensive than human call center workers in most developing countries until inference costs drop around 2030.