Report: OpenAI has largely automated training of experimental models with AI

The Information reports that OpenAI has largely automated the pipeline for training new experimental models, with AI systems doing most of the work. Commentators such as @kimmonismus and @rohanpaulai read this as substantive progress toward recursive self-improvement (RSI), while others dismiss it as nothing new. No first-party confirmation from OpenAI exists yet.

Confirmed

• The report says internal models handle most of building and training experimental models, including writing and optimizing GPU kernels and optimizing running code.
• Researchers give a single optimization example, and the AI can autonomously implement and test similar improvements for weeks.
• Internal models can also collaborate with other AI agents without human involvement; employees say this automation level was reached only in recent months.
• Retellings by @kimmonismus and @新智元 mention a concurrent OpenAI blog post claiming a newly trained model solved 100 previously open math problems.
• @新智元 also relayed claims of an urgent post warning about RSI risks and multi-agent collaboration.

Unconfirmed

• All information comes from retellings of the The Information report (m1, m3, m4, m7, etc.); the exact scope and implementation remain unconfirmed by OpenAI officially.
• Programmer arthurcolle questioned the newsworthiness, arguing such techniques have been in use for two years and mocking OpenAI's hype—without denying the facts themselves.

Why it matters

• If true, this marks a frontier lab entering the stage of models training models. @kimmonismus framed it as recursive self-improvement arriving; @rohanpaulai believes it could significantly accelerate the speed and scale of model R&D, raising questions about safety and oversight of automated research.

2026-09-21 ~ 2026-09-22 · 13 related posts

• Report: OpenAI has largely automated training of new experimental AI models — zephyr_z9 · 2026-09-21
• [source] Report: OpenAI Using AI Models to Assist With AI Training, per The Information — BrennusSokol · 2026-09-22
• [source] OpenAI has largely automated training experimental models; new model cracks 100 open math problems — kimmonismus · 2026-09-22
• Report: OpenAI has largely automated training of new experimental models internally — kimmonismus · 2026-09-22
• OpenAI models now run their own training pipeline, as company pitches global RSI safety framework — 新智元 · 2026-09-22
• [source] OpenAI says the AIs themselves are now doing most of the work training the next AIs — Puzzleheaded-King584 · 2026-09-22

7 near-duplicate retellings: Dr_Singularity · Dr_Singularity · Dr_Singularity · arthurcolle · rohanpaul_ai · rohanpaul_ai · Kyrannio