← Back to live feed1 story
OpenAI has met an internal benchmark for automating an entry-level AI researcher with Astra, its next model. The system can implement experimental ideas in the company's codebase, execute those experiments, and return the results to the user.
The model is designed to operate as a continuous agent that can run for weeks, with CEO Sam Altman targeting versions that run indefinitely through ChatGPT and the API. A new interface for these agents is coming soon, though the company noted that compute shortages limited revenue last year.
Sign in to suggest edits
Key sources
- SOURCE@alexeheath“I got to see OpenAI’s next big model, Astra”x.com
- SUPPORT@andrewcurran_“Astra is designed to run for weeks”x.com
- SUPPORT@himanshustwts“Given an exp idea, Astra can implement it in OpenAI’s codebase, run the experiment, and return the results”x.com
- SOURCE@synthwavedd“targeting next Thurs (the 3rd) to the end of the following week”x.com
- SUPPORT@testingcatalog“Internally named "mozaik-alpha-fdm"”x.com
- SUPPORT@iruletheworldmo“starting to see lots of fable 5.1 and astra outputs getting leaked”x.com
- SUPPORT@kimmonismus“Presumably also following the upcoming Fable 5.1 release, likely due next week”x.com
- SUPPORT@hesamation“There will also be a version that can run forever which is the other word for agent employees”x.com