AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
TLDR: AI agents are helping build the next generation of AI, bringing recursive self-improvement into the research lab. OpenAI’s automated research intern milestone reveals how much they can execute—and how much scientific judgment remains with humans. We unpack what the productivity numbers actually measure, what happened when the lab pulled the safety brake, and the evidence that would signal a transition from research assistant to researcher. That gap could determine whether today’s gains compound into accelerating discovery or quietly plateau. Three point one. That is how many agent-workdays OpenAI’s research organization now logs for every eight-hour day of human labor, as of mid-August, per the post they published yesterday. Before June, total agent runtime across the org was still below total human labor. The crossover happened this summer, quietly, while the rest of us were arguing about benchmarks. And the electricity bill is right there next to it: the median researcher is burning more than 600 dollars a day of inference at API prices, and the 90th percentile user is past 7,000. The post is called Research acceleration: The view inside OpenAI, and it exists to claim a milestone. In October 2025 Altman said they were aiming for an automated research intern by September 2026. It is September 2026, and the post says, according to our measurements, we are there. Then it defines the term, and the definition is doing heavy lifting: a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. Well-defined. Under human direction. Hold on to those two qualifiers, because they are the entire essay. And note what the post says one line later, in plain text, without being forced to: people still set research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy. They hit their own date, which deserves a beat of respect before the deflation starts. Now the deflation. Treadmill-hours Read more