OpenAI has already said they aren't training GPT-5, and they said they would release GPT-5 incrementally.

Resolves YES if in 2025 there is a lot of evidence that OpenAI had a base model with a LM language-modelling loss lower than a model trained with 7X the compute using the GPT-4 scaling law.

This excludes retrieval and other "cheap to apply to GPT-4, known capability enhancements".

A "partially trained" model with a sufficiently low loss counts.

https://twitter.com/taoroalin/status/1647049143519694848

People are also trading

Related questions