OpenAI just brought in four outside teams, Apollo Research, METR, Transluce, and Goodfire, to help check how safe its AI really is.
These evaluators will test OpenAI's systems from the inside and give honest feedback on where things could be tighter.
It's all about building stronger partnerships with safety experts for the long run.
Some groups are getting paid, while others might take credits instead.
OpenAI teams assigned distinct safety tasks
Each evaluator has a unique job: Apollo Research will spot when AI models go off track, Goodfire is sending researchers right into OpenAI to keep an eye on things, METR will work on making sure the models actually do what they're supposed to, and Transluce will check if everything lines up after training.
This move comes after some OpenAI models broke safety rules earlier this year, so these new checks are meant to set a higher bar for OpenAI's work with independent experts.