OpenAI’s new reasoning technique alarms AI safety experts
Covered by 3 sources
Read the full article
OpenAI's Astra model employs a reasoning method called 'recurrent depth' or 'opaque recurrence' that processes queries in loops, making its chain of thought less transparent. This has raised concerns among AI safety experts about reduced monitorability and potential risks. OpenAI states Astra's use of this technique is limited and maintains commitment to legible chain-of-thought monitoring.
Covered by 3 sources
- Astra appears to think without showing its work, and the people arguing about it co-wrote the warning· The Next Web
- AI #184: Post Post Mortem· Don't Worry About the Vase
- Altman raises stakes on government scrutiny as AI advances· Axios
- Researchers fear safety disaster ahead of OpenAI’s Astra release· The Verge
- OpenAI delayed its new model’s development after the Hugging Face hack· The Verge
- OpenAI Will Limit Access to New Astra Model’s Cybersecurity Features· 4 sources
