# AI Automates the Labs' Own R&D

Frontier labs will increasingly disclose that AI systems perform a rising share of their own research and engineering work, compressing the loop between one model generation and the next, and that disclosed share will become the anchor for governance proposals keyed to automated R&D rather than to training compute.

- Conviction: 25 / 100 (weakening)
- 7-day move: -14
- Horizon: Emerging (watchlist)
- Tracking since: 2026-09-29T00:00:00.000Z
- Last updated: 2026-10-12T14:07:13.413Z
- Canonical: https://polylog.news/ai/trends/ai-automates-ai-research
- Publisher: Polylog
- Affected regions: Global

## Recent score history

- 2026-10-11: 27
- 2026-10-12: 25

## Recent evidence

- [confirms] OpenAI Ships an Ultrafast Serving Mode for GPT-6 Astra, Claiming Up to 8x Faster Token Generation (2026-10-02): OpenAI said the inference software behind the up-to-8x Astra serving mode was optimized using its own internal models on Nvidia Blackwell hardware. This is a disclosed case of the labs' models doing the engineering that improves the labs' serving economics, tightening the loop the thesis claims will be increasingly disclosed and increasingly load-bearing.
- [confirms] A 27B Open-Weight Agent From BAAI Posts 81.8 on MLE-bench Lite by Learning to Iterate (2026-10-01): AREX-2 was trained on synthesized long-horizon improvement trajectories and scores 81.8 on MLE-bench Lite, a benchmark of machine-learning engineering tasks — evidence that ML research-engineering automation is now reachable with 27B open weights, not just inside frontier labs.

1 more evidence entry, the full score history, the conviction-driver timeline, and affected assets are for subscribers: https://polylog.news/pricing