Readable AI parts do not prove it is safe
Zvi Mowshowitz reviews claims that researchers can read parts of an AI model tied to its behavior. He sees promise but says this does not prove we understand or can control the whole system. The digest did not audit the underlying study.
What the source reports
What does it mean to “see” inside an AI model? Zvi Mowshowitz writes about Goodfire’s claims that some model parts can be linked to readable ideas. He finds the work promising, but warns that a readable label is not proof we can predict or control every decision. His essay discusses early figures and how to read them. The AI Agency digest did not independently check the underlying research or its reported percentages. A fair next step would be an outside test of whether those labels help people predict or change model behavior. This is a critique of a claim, not proof that AI safety has been solved.
Original source
- Title
- I Can See Everything
- Author
- Zvi Mowshowitz
- Publication
- Don't Worry About the Vase
- Date
- Monday, October 19, 2026