human-reinforcement learning
A training method where human judgments of which AI output is better (A vs B) are used to teach the model to produce preferred responses.
As defined in Why AI Is Good for Humans (with Reid Hoffman)
Where it comes up · 1
Most important claims · 2 of 2Search all claims for “human-reinforcement learning” →
- It is not true that 10x the data yields 10x the intelligence; what we have is an extremely good learning algorithm that captures the current basis of human inference and knowledge from data, but it is actually less data-efficient than humans, though it compensates by systematically learning and then sharing the result everywhere cheaply.Why AI Is Good for Humans (with Reid Hoffman)
- Benchmarks and tools like Chatbot Arena—which pits models head-to-head and lets users judge which answer is better, accumulating sports-like scores via the same human-reinforcement-learning A/B mechanism that trains the models—are useful and entertaining indicators of capability, but one can be overly rotated on them; they are not the substance of what AI will mean for our lives.Why AI Is Good for Humans (with Reid Hoffman)