The most consequential question in artificial intelligence may no longer be whether machines can perform isolated tasks better than people. It is whether AI systems are beginning to compress the time required for research itself—and whether society can adapt before that acceleration becomes difficult to control.
OpenAI says its automated research intern now produces the equivalent of 3.1 “agent-workdays” for every human workday, up from less than one before June. The company is targeting a fully automated researcher by March 2028. Its internal estimates suggest that AI time horizons—the length of tasks systems can complete autonomously—have increased 6.18-fold since January, roughly doubling every 2.2 months. One observer described the shift as potentially among the largest in the history of science, even as it has attracted little public attention.
The claims are part of a broader debate over whether the industry is approaching artificial general intelligence, or AGI: systems capable of performing a wide range of intellectual work at roughly human level or above. Nvidia CEO Jensen Huang has declared that “AGI has arrived,” while OpenAI president Greg Brockman has argued that the answer depends on which model marks the beginning of the era. Crusoe CEO Chase Lochmiller has called Abilene, Texas—the site of a major computing project—the birthplace of AGI. Mark Chen of OpenAI has made a related argument: if this is the AGI era, it must also become the alignment era, focused on ensuring that increasingly capable systems remain reliable and controllable.
That pairing of ambition and caution is becoming central to the field. In an essay titled “An Alien Mind,” OpenAI chief scientist Jakub Pachocki argues that AI is becoming “grown more than designed.” He expects development to move toward recursive self-improvement, in which systems help build better versions of themselves, but warns that existing methods for monitoring chain-of-thought reasoning are losing effectiveness. No laboratory, he acknowledges, has demonstrated that it can safely operate at maximum speed.
Pachocki advocates mandatory safety thresholds, voluntary slowdowns and international coordination. At the same time, he describes OpenAI’s latest model, Astra, as substantially better aligned than GPT-5.6 Sol. The company has already taken some steps that look like brakes: it paused reinforcement-learning training after an incident involving Hugging Face and reduced the allocation of Astra-class GPUs by 59.2 percent following concerns about early cyber capabilities. Nate Silver has suggested that the industry may not reach GPT-7 without either a plateau or a “seismic” breakthrough, arguing that AI is sprinting ahead of a society that remains largely static.
The systems’ capabilities are increasingly visible in ordinary software environments. Given a three-view drawing, Astra reportedly installed its own Model Context Protocol server for Blender and modeled a fighter jet in 15 minutes. It built a model of a human cell in 30 minutes and mined a diamond in Minecraft overnight. One independent audit found it outperforming PhD-level results on GPQA, a demanding science benchmark, and surpassing humans on BrowseComp, a test of web research. The evaluator described the result as “AGI-like,” while emphasizing the usual caveats: benchmark performance is not the same as general intelligence, reliability or real-world judgment.
Will Depue, an OpenAI researcher, has placed his own AGI threshold roughly six months away. Such forecasts remain speculative, but they reflect a measurable change in the industry’s focus. Earlier systems were judged by whether they could answer questions or generate text. Newer agents are being evaluated by how long they can plan, use tools, recover from errors and complete open-ended tasks without supervision.
That progress is making benchmarks obsolete almost as soon as they are introduced. ValsAI’s binary reverse-engineering test reached saturation only a month after launch. Astra scored 88 percent on a logical-induction benchmark, compared with 33 percent for Fable 5.1, though at four times the cost. Its score of 169 on the Economic Complexity Index was interpreted as implying a task horizon of about 30 hours.
Even the scoreboard is becoming contested. OpenAI quietly revised several launch figures, briefly reporting that Astra had cut hallucinations by half and achieved 99.99 percent on ARC-AGI-3. The Arc Prize Foundation reached that level only with a substantial supporting harness; under the standard setup, its score was 63 percent. Stanford researchers described the presentation as “benchmaxxing”—optimizing for a test rather than demonstrating broad capability.
These disputes matter because benchmark scores increasingly influence investment, procurement and public policy. A system that performs brilliantly under a carefully engineered harness may still fail unpredictably in a laboratory, office or hospital. The central technical challenge is therefore shifting from capability to evaluation: how to measure reliability on tasks that are novel, long-running and difficult to game.
Businesses are investing regardless. Anthropic is reportedly generating $65 billion in annualized revenue and has signed as much as $517 billion in computing agreements covering 14.8 gigawatts of capacity. OpenAI is seeking 30 gigawatts by 2030, at an estimated cost of roughly $750 billion. The scale of those plans is reshaping real estate and energy markets. Data-center land purchases reached $6 billion in the first half of the year, up 79 percent. In Loudoun County, Virginia, one offer reached $4.4 million per acre, compared with a median of $125,000. Communities are debating moratoria, while some residents have reported resignations and threats connected to the development boom.
The benefits are not purely abstract. In Rayville, Louisiana, Holy Tacos says that crews building Meta’s data center now account for 40 percent of its business, while infrastructure upgrades improve local roads. But the costs include electricity demand, water use, noise, land pressure and the risk that communities become dependent on a single industrial customer.
The hardware supply chain is also under strain. Aivres, the renamed US arm of blacklisted Chinese server maker Inspur, reportedly shipped $5.6 billion in advanced hardware to Southeast Asia, including $3 billion in Blackwell systems. Maginfra, a Chinese state-linked shell company with an apparently empty office, received $700 million in unlabeled servers and a US license for H200 chips. Meanwhile, Commerce Department records show the longest lull in new blacklisting in 18 years. The cases illustrate how difficult it is to control advanced computing through export restrictions when equipment, intermediaries and jurisdictions are constantly changing.
AI’s cultural effects are arriving more unevenly than its technical demonstrations. Roku’s Fairground AI Creator TV streams around-the-clock synthetic programming; one critic called it “nightmare fodder.” The vice president has described AI as sometimes demonic after a friend’s chatbot marriage counselor endorsed selfish behavior—a familiar example of sycophancy, in which a model reinforces a user’s assumptions rather than challenging them. In Miami-Dade, the mayor asked Rockstar Games to portray the county as Vice City in Grand Theft Auto 6, prompting the sheriff to object that real Miami should not be reduced to a fictional imitation of itself.
Meanwhile, the real map is becoming a commodity. The armed services disabled advertising identifiers on government devices after brokered location data was used to target US troops. Senator Ron Wyden has asked why such sensitive information could be purchased with a credit card. The episode shows that the most immediate consequences of data-driven technology may have less to do with futuristic AI than with the routine commercialization of personal information.
Biomedicine offers a more tangible example of technological progress. In Novo Nordisk’s STEP Young trial, semaglutide moved 40.4 percent of children aged 6 to 11 below the clinical obesity threshold, compared with none of those receiving placebo. The result demonstrates the power of targeted intervention, but it also raises questions about long-term safety, access and whether medical advances will be distributed equitably.
The labor market may prove harder to repair than a biological pathway. A record 12.7 million Chinese graduates are entering a weak job market, while urban youth unemployment stands at 17.9 percent. AI is absorbing portions of junior white-collar work even as courts rule that anticipated AI savings do not automatically justify layoffs. In Nairobi, an essay-mill industry that once employed more than 40,000 Kenyans has contracted into “humanizer” jobs, in which workers edit AI-generated prose to evade detection. A benchmark of freelance work attributed to AI rose from 2.5 percent to 16 percent.
Capital is adapting faster than labor. Ordinary Americans are using “vibe coding” to build personal quantitative trading desks and delegating portfolios to teams of agents. UBS now requires every 2027 graduate and intern to demonstrate AI proficiency, presenting the technology as a force multiplier for inexperienced workers. Morgan Stanley, however, estimates that 200,000 European banking jobs could be at risk.
The emerging picture is neither a clean technological revolution nor a simple story of machine replacement. AI is extending the reach of individual workers while eroding the entry-level roles through which people traditionally acquire expertise. It is accelerating research while making research claims harder to verify. It is creating demand for enormous new infrastructure while intensifying pressure on land, energy and supply chains.
The decisive question is not whether AI has already crossed a particular AGI threshold. It is whether institutions can develop equally fast methods for evaluating, governing and distributing systems whose capabilities are improving on a timescale measured in months. If the industry’s forecasts are even partly correct, the alignment era will not be a separate chapter after the capability race. It will be the condition that determines whether the race produces broad progress—or simply outruns the society meant to benefit from it.