Part 3/10:
OpenAI's intent with this shift is to equip the model to handle complex problems in fields such as science, mathematics, coding, and biology—areas where straightforward answers are often insufficient. Since September 12th, GPT-1 Preview has been accessible via ChatGPT and an API as a preview version, with ongoing updates aimed at refining its capabilities.
Outstanding Performance in Complex Tasks
Early evaluations showcase impressive performance benchmarks:
- Science and Math: In a qualifying exam for the International Mathematical Olympiad (IMO), GPT-4 scored only 13.3% on difficult problems, while GPT-1 Preview achieved an 83% success rate, nearing the capabilities of PhD students.