September 22, 2026
AI is already training the next AI: OpenAI gives agents tasks lasting weeks
On September 21, The Information reported that AI has already largely automated the training of OpenAI's experimental models. Models write programs to run and train models on GPUs, including kernels and their optimization.
Previously, OpenAI tested some ideas for several years. More server capacity has reduced that timeframe to 1 week. An engineer can show AI a single required optimization, after which it works on the implementation for weeks.
Agents used by different OpenAI employees already sometimes solve tasks together without involving people.
