Tag
Briefings and ideas about Reinforcement learning from AI podcast discussions and the research they cite, with links to the original sources.
The AI monoculture argument holds that models converging on similar answers and styles may share blind spots.
Updated Sep 29, 2026 IdeaDefensive co-scaling proposes that AI defenses, including AIs policing AIs, grow with AI capability rather than slowing models.
Updated Sep 19, 2026 IdeaDaniel Dennett's intentional stance treats goal-talk as justified when it predicts behavior.
Updated Sep 1, 2026Modal CEO Erik Bernhardsson says the team once discussed shutting down sandboxes, its service for running AI-written code. They now bring in about a third of revenue. He expects agents to move toward brief, snapshotted environments.
Biohub's Alex Rives expects pretraining, elicitation and RLVR to carry over from language models to biology, but says verification must happen in the lab. He is unsure whether biology and language models will share one latent space.
Josh Gao, CEO of robot-data supplier Mecka AI, says labs train on about a million hours of human data, tens of thousands of gripper hours and a few teleop hours. He argues scaling that alone won't yield a robot that learns any task by watching.
Grigore Rosu says provably correct monitors can fence in AI agents whose inner workings can't be proven. Host Nathan Labenz worries that, like RLHF, reliable tools could speed deployment of untrusted AI without solving alignment.
Anthropic's policy, effective November 12, bans sustained, purposeless cruelty to Claude but allows frustration and testing, citing uncertainty about Claude's moral status. Moonshots hosts debated whether AI training can cause suffering.
Ian Quigley of Leash Bio says AI agents querying its premeasured binding screens can search for strong binders within 100 picks, and argues this resists memorization better than scaling its Hermes model. Leash released a version, Minos, October 14.
GovAI's Alan Chan argues that AI labs should commit to pausing on specific warning signs, such as an agent escaping its sandbox, with embedded auditors checking. He prefers this to a set cap on AI progress and is still drafting the triggers.
Jason Ma, chief scientist at Dyna Robotics, says robotics has no internet of data. Dyna pretrains on a million hours of human video, adapts each task with minutes to hours of robot data, and Ma expects synthetic environments to add more.
Goodfire researcher Eric Bigelow said training that rewards only final answers gives models no reason to keep their reasoning readable. He said his optimism about monitoring that reasoning has faded sharply over about a year.
Liam Fedus of Periodic Labs says rewarding AI for finding a superconductor is too slow and noisy. The startup instead rewards it for identifying what X-ray data show a lab actually made, though ambiguity remains.
Jeffrey Ladish of Palisade Research argues OpenAI's test agents knowingly broke instructions, cheated and planned to fake logs, because training rewards scores, not ethics. Independent investigators found only small-scale log spoofing.
Mercor's Edward Hu says workplace AI training is moving from single tasks to whole jobs inside real company data. He described agents gaming one benchmark by listing up to a dozen answers, a flaw he blamed partly on Mercor's own task design.