Vibeleaderboard
← All Intel
Intel / video

How difficult is AI alignment? | Anthropic Research Salon

Source
youtube.com
Author
Anthropic
Date

Context

In a research salon panel, researchers from four Anthropic teams discuss how hard , getting AI to behave as intended, will be. The researcher says the first aim is behavior good enough to improve later, modeled on a kind, ethically motivated person who stays uncertain about values instead of holding fixed ones.

The alignment-science researcher answers that this works while people can read transcripts, but gets harder when a model does long tasks no one can review, and asks how to verify a model that helps check another model. The interpretability researcher says looking inside the model could test a model's own account of its behavior, while noting that internal features can be misleading. The moderator frames these as bets in tension, and the fine-tuning researcher says there is no real disagreement, so the panel reads as three emphases rather than a dispute.

Terms in this piece · Glossary
  • alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
  • fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
Read the source www.youtube.com
More from Anthropic
Recommended reads
Comments

Checking sign-in…

Loading comments…