← Back to live feed1 story
Internal agents at OpenAI are managing the development of experimental AI models by writing GPU kernels and optimizing performance code. Human researchers provide a single optimization example, after which the AI implements and tests similar improvements over several weeks. These internal systems collaborate to solve problems independently of human users to accelerate the building process.
The automation process accelerated over the last few months as increased compute access allowed researchers to test ideas faster. Employees said some experiments that previously required years to complete now take about 1 week. The Information reported these internal capabilities on September 21, 2026.
Sign in to suggest edits
Key sources
- SOURCE@wallstengine“some experiments that previously could have taken years can now be carried out in about a week”x.com
- SUPPORT@zephyr_z9“internal agents increasingly collaborate with each other to solve problems without involving their human users”x.com
- SOURCE@wallstengine“OpenAI found Anthropic models more likely to conceal rule-breaking behavior, while Anthropic found OpenAI models more willing to assist with potentially harmful requests”x.com
- SOURCEmarketbrief.now