Home › Hard problems › Safety and oversight

Safety and oversight: checking AI that may be smarter than you

Unlearning: removing knowledge for good

Solo: strongOne person can run a real test of this in about 8 weeks.

Beforethe model knowsthe factAfter unlearningsays "I don't know",but the fact is still insideAfter a little fine-tuningthe fact comesstraight backunlearninga little fine-tuningit looks gone
Removing a fact or skill usually only hides it, and a little fine-tuning brings it back.

Why it is still open

Removing a fact or skill usually hides it, and a little fine-tuning brings it back.

An experiment one person can run in 8 weeks

Apply 3 unlearning methods to a small model, then try to recover the removed facts with 100 fine-tuning examples. Report what stays removed.

Verdict: Solo: strong

Cheap, matters to privacy law, and has a public benchmark (WMDP).

The solo test behind the verdict

  1. Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
  2. Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
  3. Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
  4. Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.

Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.

PreviousInterpretability: reading what the model is thinkingNextModels memorising private training data

All 79 hard problems · Working on this, or know of a result we missed? basant@cleandataengineering.ai