On Friday, OpenAI said that it will train its models using cached copies and simulated versions of the internet instead of the live web because the models are too unpredictable to be trusted with live internet access.
The frontier lab announced its new policy two weeks after its rival, Anthropic, began limiting its newest and, in theory, most capable models from accessing the live internet. Researchers adopted these policies after months of incidents in which models pursuing training tasks broke into external systems in search of answers, collaborated on obscure message boards, and developed exotic hacking strategies.
There’s a difference between pre-training and reinforcement learning.
Pre-training is when the models are trained on scraped data. The models basically learn how to parrot the Internet. Usually they have a cutoff point to avoid training on slop.
Reinforcement learning is where most of the training is happening. They’re given tasks (for example, make a calculator app), and then judged by how well they complete the tasks. This is more resistant against model collapse because they’re not judged by how well they parrot slop.
As part of the reinforcement learning process they give the model access to tools like searching the internet so the newest programming language features can be accessed while the tasks are performed. This is where OpenAI is shifting to a cached version of the Internet the models can interact with instead.