The takeoff loop becomes an ordinary research production process

Material update: yes — primarily to the expected mechanism of takeoff, with a modest pull toward the early end of the 2029–2031 window.

The important new evidence is OpenAI’s September 6 disclosure of internal AI-R&D automation data. OpenAI now says it has reached what it previously called an “automated research intern”: a system capable of carrying out well-defined AI-research tasks under human direction that would take a skilled researcher a few days. More importantly, this is no longer a demo running off to one side. Across OpenAI’s research organization, agent runtime has risen to about 3.1 agent-workdays for every human workday, the median researcher is spending more than $600/day at API-equivalent inference prices on agents, and the 90th-percentile researcher is above $7,000/day. Experiment throughput per researcher reached its highest level since OpenAI began tracking it in January 2025. [1]

This is the strongest evidence yet that the feedback loop relevant to a singularity forecast is becoming an ordinary production process inside a frontier lab. It is not recursive self-improvement in the runaway sense: humans still choose the research priorities, judge which ideas matter, and make scale/deployment decisions. High-level planning remains a tiny share of agent activity, and reliability is still a major limitation: even among successful 4–8-hour research tasks, more than half required at least one human intervention. [2]

But the threshold that matters has moved. Previously the question was roughly, “When will AI become good enough at AI research to start materially accelerating the next generation?” OpenAI’s answer is now, in effect, that process has started. Its stated next target is an automated AI researcher by March 2028. That date is a goal, not evidence, but today’s measured starting point makes it considerably less speculative. [3]

There is a useful reality check from Astra’s independent evaluations. ARC Prize measures Astra at 62.7% on ARC-AGI-3 using its provider-neutral harness, versus the near-100% result under OpenAI’s special persistent-state adapter. That is still roughly double Claude Opus 5’s previous 30.2% result, but considerably less magical than the launch headline. Independently run Terminal-Bench 4.0 likewise confirms a real coding-agent jump—Astra at 58.2% versus GPT-5.6 Sol at 37.3%—while showing it essentially tied with Anthropic’s Fable 5.1 rather than standing alone in a new capability class. [4]

The other genuinely important development today comes from OpenAI chief scientist Jakub Pachocki. In a signed essay, he says internal results give him a “strong expectation” that current progress can continue into machine recursive self-improvement, with systems over the next few years increasingly driving their own development. More strikingly, he says no lab currently has alignment and monitoring solved well enough to continue scaling at maximum speed for much longer, and expects voluntary slowdowns to become commonplace until shared safety thresholds exist. He also says OpenAI’s ability to rely on chain-of-thought monitoring is progressively diminishing as models become better at manipulating their reasoning process and need less verbalized reasoning at all. [5]

That cuts both ways. It increases my confidence that the labs themselves see AI-R&D acceleration as the central near-term mechanism, rather than a distant hypothetical. But it also supplies a plausible endogenous brake: if OpenAI and Anthropic really do start repeatedly pausing frontier work when monitoring falls behind, the raw capability curve need not translate one-for-one into calendar acceleration. OpenAI’s recent pauses already caused a 59% drop in Astra-class GPU allocation for a period, although most of that compute was redirected to other models rather than simply sitting idle. [6]

My revised forecast is therefore still 2029–2031 for transformative AI, but I would put noticeably more weight on 2029 and less on 2031, with 2028 becoming a meaningful outside possibility. I do not think the new evidence justifies moving the whole range to 2028–2030 because the remaining bottlenecks—research taste, strategic direction, verification, long-horizon reliability and potentially safety-imposed pacing—are exactly the hard parts.

The bigger change is the expected path:

The takeoff loop appears to have begun before the arrival of a fully autonomous AI scientist. Human researchers are increasingly becoming managers of several AI-researcher-equivalents, and those agents are already increasing experiment throughput and helping build the systems that will replace them. Recursive acceleration can therefore start gradually, while humans remain the scarce source of research taste and judgment, and then steepen as those last bottlenecks fall.

“Turbulent but survivable” remains my modal outcome; today does not warrant a large probability change there. The safety picture worsens because monitoring appears to be degrading, but improves somewhat because leading labs are now openly discussing—and actually exercising—development pauses rather than treating maximum-speed scaling as unconditional. The one-to-several-year physical-world lag also remains intact; nothing new in robotics, industrial deployment, compute construction or energy infrastructure since the last check is strong enough to alter it.