2026-09-07 · MondayPublished daily at 08:00
AI Daily
- 1
- 2Fortune reports that OpenAI repeatedly changed GPT-6 Astra benchmark data, with some scores shifting substantiallyIT Home · Artificial Intelligence
- 3
01Tips and perspectives
OpenAI releases an internal research acceleration report: it has achieved its goal of an automated research intern and is advancing toward an automated AI researcher by March 2028
OpenAI published an update on its progress in automated research, announcing that it had met the goal set last fall of having an automated research intern by September this year—one capable of completing well-defined research tasks that take days under human guidance—and plans to build an automated AI researcher by March 2028.
Original textFortune reports that OpenAI repeatedly changed GPT-6 Astra benchmark data, with some scores shifting substantially
According to Fortune, OpenAI has revised its benchmark data multiple times since announcing GPT-6 Astra on September 3: Astra’s hallucination rate was once reduced from 4.2% to 2%, then changed back.
Original textOpenAI explains the challenges of alignment and monitoring in a lengthy article, saying that chain-of-thought monitoring capabilities are weakening
OpenAI published a long-form article, “An Alien Mind,” revisiting the beginnings of the 2023 RLSlow project, which confirmed that reasoning models could be trained at scale, and systematically explaining the distinction between goal alignment and value alignment. The article notes that chain-of-thought monitoring is becoming less effective as model capabilities improve. GPT‑6 Astra is significantly better aligned than GPT‑5.6 Sol. The authors expect progress may continue toward recursive self-improvement (RSI), and call for voluntary limits on scaling, third-party safety thresholds, and stronger international coordination.
Original text