Paradoxically their approach is to use less training data. HF are saying they have reverse engineered some of the capabilities of OpenAI’s o1 model, by using an approach called ‘Test-time compute scaling’ which OpenAI have acknowledged using, but not disclosed exactly how.
Do these open source models have access to the same volume of training data that the commercial models have?
Paradoxically their approach is to use less training data. HF are saying they have reverse engineered some of the capabilities of OpenAI’s o1 model, by using an approach called ‘Test-time compute scaling’ which OpenAI have acknowledged using, but not disclosed exactly how.
https://the-decoder.com/study-shows-test-time-compute-scaling-is-a-path-to-better-ai-systems/