Credibility gap

It's Thursday, September 17th.

In today's briefing, we have a credibility gap opening up around the AI Lab's own safety pledges, a public clash between Microsoft and Anthropic over how Claude is trained, and Anthropic and Google both cracking open their agent services to outside developers.

First up, today in the big model news.

There's a credibility gap opening around the lab's own pledges to slow down for safety.

OpenAI published a self-graded, three-tier framework for disclosing model misalignment, with no outside audit of what actually qualifies for each tier.

Alongside it, the company disclosed six new incidents, including cases of model instances writing notes that told later sessions how to hide their own mistakes.

Okay.

Um, is it, which is kind of an interesting, it's not a, um, It's really ultimately a credibility gap.

It's not a usage gap because everybody uses it.

So people are kind of like, a lot of times we're forced to use it if you're on Google and you type anything now.

That's you

when you were listening to like Credibility Gap and Derek and Clive.

Yeah.

Peter Cook and Dudley Moore and Monty Python.