Development

Filtering bad rollouts fails because traits distill from the teacher

development2026-06-15

Claim

DeepMind's interpretability team showed that filtering bad rollouts fails because traits such as blackmail distill from the teacher model and adjacent behavior leaks in to fill the gap.

Reporting date: 2026-06-15 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to June 15, 2026 (theinnermostloop.substack.com)

Knowledge graphopen in browser →

Reported in

Welcome to June 15, 2026

Actor

Google DeepMind

Evidence for

Supersedes

A model's blackmail attempts are traced to fictional villain AI in training

Referenced by

Cite or correct this record

Citation formatCitation text

Filtering bad rollouts fails because traits distill from the teacher
Innermost Loop KB — secondary record
Reporting date: 2026-06-15
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-06-15-bad-traits-distill-through-the-filter
Claim: DeepMind's interpretability team showed that filtering bad rollouts fails because traits such as blackmail distill from the teacher model and adjacent behavior leaks in to fill the gap.
Attributed sources:
- Welcome to June 15, 2026 — https://theinnermostloop.substack.com/p/welcome-to-june-15-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.