HCH
In plain wordsA thought experiment: a person who can hand parts of a problem to copies of themselves, who can hand parts on again, endlessly — the ideal some AI training schemes try to approximate.
A recursively defined, idealized question-answering process: a human answers a question — in many later versions, within a short time limit — while able to consult copies of HCH on subquestions, each of which can consult further copies in turn. The name is a recursive acronym for 'Humans Consulting HCH'. HCH is the theoretical limit of iterated amplificationiterated amplification, n.: A proposed training scheme in which a person, helped by many copies of an AI, solves harder problems than either could alone; a new AI then learns to imitate that team, and the cycle repeats.: if amplification and distillation worked perfectly, the trained model would imitate HCH. It is used to reason about whether a tree of individually trustworthy, short-lived human thoughts could together solve very hard problems without the whole becoming misaligned — the factored cognitionfactored cognition, n.: The idea that a hard thinking task can be broken into many small, separately checkable pieces, each simple enough for a person or an AI to handle in a short time. question. Critics ask whether such a tree would be competent at tasks needing deep accumulated understanding, and whether small errors or manipulative subanswers could compound through the recursion.
Imagine HCH answering a research question: thousands of ten-minute humans, each delegating pieces to the next layer.
Origin
Introduced by Paul Christiano in the blog post 'Humans Consulting HCH' (January 2016) on his AI Alignment blog, as part of his amplification-based approach to alignment; he credited Eliezer Yudkowsky with suggesting the recursive acronym.
Related
- iterated amplificationiterated amplification, n.: A proposed training scheme in which a person, helped by many copies of an AI, solves harder problems than either could alone; a new AI then learns to imitate that team, and the cycle repeats.
- factored cognitionfactored cognition, n.: The idea that a hard thinking task can be broken into many small, separately checkable pieces, each simple enough for a person or an AI to handle in a short time.
- Paul ChristianoPaul Christiano, prop. n.: An American AI-alignment researcher who led the work that made training AI from human preferences practical, the basis of how chatbots are taught to be helpful.
- AI debateAI debate, n.: A proposed way to supervise very capable AI: two AI systems argue opposite answers to a question and a human judge decides who argued better, on the theory that lies are easier to expose than to defend.
- scalable oversightscalable oversight, n.: The problem of how people can check and correct AI systems whose work is too difficult, fast or plentiful for humans to judge reliably alone — usually by enlisting AI to help with the checking.
Sources
- Christiano, P. (2016). Humans Consulting HCH. AI Alignment blog (ai-alignment.com); reposted on the AI Alignment Forum, 2018. alignmentforum.org/posts/NXqs4nYXaq8q6dTTx/humans-consulting-hch
Mentioned in
factored cognitionfactored cognition, n.: The idea that a hard thinking task can be broken into many small, separately checkable pieces, each simple enough for a person or an AI to handle in a short time., iterated amplificationiterated amplification, n.: A proposed training scheme in which a person, helped by many copies of an AI, solves harder problems than either could alone; a new AI then learns to imitate that team, and the cycle repeats.