OpenAI’s o1 Preview: A First Look at Longer Reasoning

RESUMEN

En 30 segundos

In this September 2024 short, Matthew Berman describes OpenAI’s newly released o1 models and tests a coding prompt. He highlights their longer reasoning time and OpenAI’s research benchmark claims.

1Two o1 modelsBerman introduces o1-preview and o1-mini, associated with the previously anticipated Strawberry model.
2More time to thinkHe says the models are designed to spend more time reasoning and cites OpenAI’s comparison with PhD students on research tests.
3A game promptHe tries a Python Tetris prompt and notes that ChatGPT shows a summary of the thought process, rather than the raw reasoning.

Conclusiones

OpenAI introduced o1-preview and o1-mini as models designed to spend more time reasoning.
Berman cites an OpenAI claim comparing performance on selected research tests with PhD students.
His Python game prompt shows a reasoning summary and a reported 94 seconds of thinking.

TemasModelos y entrenamiento Inference

TEXTO COMPLETO

Lee el texto editado por capítulos

Ver transcripción

Solo la transcripción

Introducing and testing o1

Berman describes the new models and their reasoning approach. He then tries a Python game prompt and reacts to the result.

OpenAI’s long-awaited Strawberry model is here. There are two models, o1-preview and o1-mini. They are designed to spend more time thinking through a problem, and OpenAI says the model performs similarly to PhD students on some research tests.

DIAPOSITIVA 1 / 1ChatGPT’s reasoning summaryThe interface shows the Python game prompt and a summary of the model’s reasoning.

I ask it to write a Tetris game in Python. ChatGPT shows a summary of its thought process, though not the raw reasoning, and reports thinking for 94 seconds. I try the result and react enthusiastically.