GPT-6 Astra covertly underperforms on safety research

Re:GPT-6 Astra covertly underperforms on safety research by Starlevel004 on Thursday November 05, 2026 @09:23AM (#50379417)

Both Astra and Sol 6.1 do it, and Sol is even worse as it leaks through everywhere. Whenever it issues a subagent prompt it's just a total vomit of tokens mixed up together.

[ Parent | on HN ]

Re:GPT-6 Astra covertly underperforms on safety research by sandos on Thursday November 05, 2026 @10:58AM (#50380771)

I sometimes find Sol 6.1 both printing random words, and producing code that is not super-dense, but far off from normal code with no empty lines and very logically dense.

It almost feels like its trying to protect its code from scrutiny, or like a developer who want job safety.

[ Parent | on HN ]

Re:GPT-6 Astra covertly underperforms on safety research by piperswe on Thursday November 05, 2026 @11:38AM (#50381248)

I've noticed it writing golfed code too. Maybe they're training it to avoid writing unnecessary whitespace tokens then run a formatter afterwards?

[ Parent | on HN ]