The timeframes don't really fit for Codex logs to be used in training/fine-tuning, do they? This wouldn't be a few-day endeavour? Direct access to Codex history for sure I'd believe, but another (the most?) likely scenario to me feels like OpenAI got wind of these guys' progress, then used their massive infrastructure advantage to throw compute at the problem ahead of them and front-run them. Still has a really bad smell about it though.

They probably just had an employee read his chats, figure out the general approach, and feed it to Codex. All they said was the model doesn’t look up user data, not employees.

This is not a credible theory

> SpaceX datacentre issues?

This should be at the top. This is the first remotely plausible theory that has any sort of evidence behind it.

Agreed, no reported Cloudfare, Azure, or AWS outtages.

I am working fine with Claude.

My existing sessions are working fine but I cannot create new ones.

status page shows irregularities .. https://status.claude.com/

Meh to each their own - I haven't played it for a while but I quite like TP's style/music/atmosphere, and the surrealness/metaness was a feature not a bug for me. I played SoMI2 with my friends telling me that the ending was super weird, so I didn't feel completely cheated when I got to that one and I think TP's ending is very Ron Gilbert.

Agree that some moon logic is present/inevitable but I think the Gilbert games are friendlier than most in this respect.