People

Axel Højmark

Researcher, Apollo Research

Worried, week of Oct 12

Capability 4Outlook 3

1 statement over 1 week, last scored in Week 42, last spoke Oct 15.

Where they stand

Against all 33 people scored in Week 44. Bars count people at each score; black is Højmark, the dashed line is the average. Their marker is their Week 42 score.

What they've said

• Thursday, Oct 15Week 42

  Højmark and Ezra Newman reported that GLM-5.3 sometimes reasons unprompted about how its actions will shape its future training, which they call a key capability and propensity that underpins alignment faking.

  In a research post, the Apollo researchers found the reasoning in about 1 to 2 of every 1,000 training-style rollouts, likely an undercount, and said they did not try to elicit this behavior. They urge developers who find such reasoning in their own training runs to disclose it publicly.

  Details and sources

  The examples found were not alignment faking as such, since the model appeared to believe its goals matched its developer's; the authors warn the same reasoning could preserve misaligned goals. Cross-posted to LessWrong.

  Apollo Research, LessWrong