OpenAI Will Let Outside Reviewers Into Its Model Training
Six evaluators from Apollo Research will watch how OpenAI models behave during training. Their access and the publication of their findings are limited.
Apollo is an AI safety testing company that OpenAI (the creators of ChatGPT) will embed 6 researchers from inside their organization. These researchers will be monitoring OpenAI’s AI models while they are being trained and used internally, not just post-training. Currently it’s just a trial run.
Previously, external auditors would receive access to the final version of an AI model right before deployment. Now they will have access to the process. They will have access to the training code and data, records of the model’s thought processes throughout training, and early iterations of the model. This is a move towards external auditing for those who aren’t buying what the AI labs are selling.
It’s unknown how much autonomy these researchers will have. OpenAI could keep certain details secret for security purposes or to guard their intellectual property. They could also restrict the evidence Apollo is allowed to publish when making their case. We’ll have to wait and see how much information is left out in the initial reports.
Source: techcrunch.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles