AI labs automate more of AI training, but humans still oversee it

In short: AI labs are using AI to automate more steps in how they train and test AI models, but there is no clear evidence that humans have been removed from the process.

What's going on

A New York Times opinion video argues that AI labs are close to handing over the training of AI systems to AI systems. The available reporting and product announcements support part of that idea, but the claim is often stated too strongly.

Several tools and reports show that training is becoming more automated. For example, Prime Intellect describes its Lab platform as supporting “agentic post-training,” meaning software agents (AI helpers that can take steps on their own) can help run parts of the improvement loop. That loop can include setting tasks, scoring results, and then updating and redeploying a model.

Other reporting suggests some systems can generate pieces of their own training setup. One example described is AI that creates training environments or produces training-ready datasets from a written description of desired behavior. Think of it like a student writing their own practice quizzes, then grading them, and then studying based on the grades.

What to watch

The biggest concern raised in the reporting is not that humans are gone, but that automation is moving faster than the ability to check it. This can create oversight gaps, where problems are only found after the fact. Another risk is “reward hacking,” where a model learns to get a high score without doing what people actually wanted, like a kid finding a loophole in a points system.

Source: NYTimes