GPT-6 solving historical ciphers signals quiet capability jump

A thread on GPT-6 Astra solving a WWI German radio cipher generated substantive discussion. The cipher was previously undecoded because a German operator mistyped the key and used a key from the wrong date. All keys are known from captured records, but the message had never been successfully decoded because of the human error in its creation. GPT-6 solved it anyway.

Commenters flagged the legitimate concern about whether the model could have manufactured a plausible fake key and message given that ship logs might be online, but judged this unlikely. The more important comment: 'This is how AGI happens. It gradually keeps getting better until one day we realize they are tremendously capable, all while completely sidestepping any notion of consciousness or self awareness.' That framing is worth sitting with.

The math and proof thread ran parallel, with someone noting that Jacob Tsimerman predicted superhuman mathematical expositors by April. The combination suggests we're in a period where model capability jumps are arriving faster than our intuitions about model limitations are updating.