FAR AI Video Watch posted an update
Why Post-Training Can’t Fix AI Alignment | Dylan Hadfield-Menell (MIT)
Why it mattersThe talk presents Hadfield-Menell's hypothesis that post-training alignment can be undone cheaply, citing abliterated models and reported failures of prompt-injection defenses. It considers moving alignment into pre-training and discusses data filtering, synthetic documents, control tokens, and the trade-off between cost and flexibility.
Discuss: If alignment moves into pre-training, how might researchers balance the cost of changing training data against the flexibility to revise alignment targets?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.