RLHF is not a tax on capability: preference tuning also raises maths and code scores, which is why the labs keep reaching for it.

Added to korrents 7 Sept 2026 · How quotes work · Something wrong? Tell us

Shared subject first, then similar wording. They may agree or disagree.

• A verifiable task can be optimised by reinforcement learning until a neural network performs it extremely well. AKAndrej Karpathy17 Nov 2025reinforcement learning
• Capability does not generalise for free: a model that will move mountains on an agentic task still tells the same bad joke it told five years ago. AKAndrej Karpathy20 Mar 2026reinforcement learning
• Current models are unsafe because they are not smart enough, not because they are too smart. FCFrançois Chollet13 Sept 2026reinforcement learning
• Humans barely use reinforcement learning for intelligence — what RL they do use goes into motor tasks, not problem solving. AKAndrej Karpathy17 Oct 2025reinforcement learning
• Humans keep their place in AI as judges rather than authors, because telling which of two answers is better is far easier than writing a good one. NLNathan Lambert3 Feb 2025reinforcement learning

See this claim on the map