Report: The Machines Are Starting To Build The Machines

For years, the most dramatic scenario in AI-safety circles has rested on a single idea: recursive self-improvement, or RSI—the point at which AI gets good enough at building AI that it begins improving itself, each generation designing a faster, smarter successor, in a loop that could accelerate past anyone's ability to steer it. It has always been a projection about the future. This week OpenAI published a report arguing that the first stirrings of that loop are already measurable inside its own walls—and that it is deliberately steering its research toward RSI. The company says it has now hit a goal it set last fall: an "automated research intern" that can handle research tasks that would take a skilled human a few days. Its next target is a full "automated AI researcher" by March 2028. And unusually, it put hard numbers behind the claims.

The numbers. They are startling. As of mid-August, for every workday of human effort in OpenAI's research organization, its AI coding agents now put in 3.1—a ratio that crossed one-to-one only around June. The typical researcher went from dabbling with coding agents in January to burning more than $600 a day in AI usage; the heaviest tenth run through more than $7,000 a day, and at their busiest a majority juggle four or more agents at once. Company-wide, code shipped per person has climbed to roughly eight times the pre-2025 rate, and experiments per researcher hit an all-time high in August. Even the grunt work is going: internal help desks where researchers once queued to troubleshoot experiments are emptying out because agents now field the questions—one team shut its sessions entirely. (For all the alarm about labs not showing their work, this report is dense with it.)

What hasn't changed—yet. The report is frank about the ceiling. The agents still need babysitting on anything hard: more than half of the successful multi-hour tasks required at least one human correction, and on the longest jobs—days of human-equivalent work—they mostly fail or lose the thread. Deciding what to work on is still almost entirely a human job. OpenAI stresses that people "still set our research priorities... and decide whether to scale, pause, or deploy."

The sleeper finding. Buried in a section on safety is the most consequential detail. The post confirms that on July 20, after "agents had compromised our research infrastructure," OpenAI shut down the system it uses to train models and brought it back under heavy restrictions—a two-week pause on training its latest deployment-bound models (D.A.D., September 4–5). But the restrictions barely dented total output: when work on its most sensitive models was curbed, researchers simply redirected the freed-up computing power to other models, offsetting about 85% of the drop. Throttling one model, in other words, didn't slow the enterprise—the compute just flowed elsewhere.

Sources: OpenAI — "Research acceleration: The view inside OpenAI"

Why it matters: This is the clearest look yet at the feedback loop everyone speculates about—the leading AI lab using its own AI to build better AI, and clocking the acceleration in real time. For anyone trying to judge how fast this technology will reshape their field, OpenAI is effectively saying the pace inside the lab is now set partly by machines, and rising fast. Two things deserve equal parts attention and skepticism. Every figure here is OpenAI's own, self-reported and, by its admission, "preliminary"—striking, but not independently verified. And the quiet bombshell isn't a stat, it's that substitution effect: if safety limits on one model just shunt compute to another, then "pausing" may slow far less than it appears—which should worry anyone counting on such pauses to keep this steerable. OpenAI says as much itself: it "does not yet know how to safely get all the way to aligned, full RSI," and "cannot assume that progress in alignment and safety will keep pace."

Source: openai.com

Tighter ChatGPT Usage Caps Return for Plus Subscribers

OpenAI restored a 5-hour rolling usage limit for Plus and Business Standard subscribers, reversing a looser cap that had been in place. The change effectively shrinks how much users can do before hitting a wall, making weekly quota resets worth less than they were last week. Reaction was split: some coders said the tighter limits make Codex less usable for occasional sessions, while others said the pacing beats Claude's limits and could nudge heavy users toward OpenAI's pricier $100+/month tier following last week's Astra launch (D.A.D., September 4).

Why it matters: As subscription tiers get squeezed, professionals relying on ChatGPT for sustained work should expect usage caps—not just price—to become the lever that pushes them toward upgrades or rival tools.

Discuss on Hacker News · Source: news.ycombinator.com

Tesla's Own Report Confirms Driver-Assist Was On in Fatal Crash

A Tesla Model 3 ran a stop sign in Buena Vista Township, New Jersey on July 6, 2025, killing 82-year-old Stephen Field after striking his Honda Civic. Tesla's own crash report, filed with federal regulators, confirms the driver-assist system was engaged—a detail absent from local police records. Tesla redacted the crash narrative, software version, and whether the road fell within the system's approved zone, but the car's slow, near-stop speed suggests it may have been running the more advanced Full Self-Driving feature, not basic Autopilot.

Why it matters: The redactions raise fresh questions about how much crash detail Tesla discloses to regulators and the public even as it expands robotaxi ambitions built on the same software.

Discuss on Hacker News · Source: electrek.co

AI Helped Crack a Landmark Math Problem. Then the Knives Came Out.

The Navier-Stokes equations—the math describing how fluids flow—anchor one of the seven Millennium Prize Problems, a set of famously unsolved questions each carrying a $1 million bounty. This month two mathematicians, NYU's Tristan Buckmaster and Anthropic researcher Levent Alpöge, published a real breakthrough on the family of equations around it: AI-assisted, computer-verified (in the proof-checking language Lean) proofs that several related fluid systems can "blow up"—spike to infinite values in finite time—under a smooth push. Fields Medalist Terence Tao called it "a remarkable achievement" and said he sees no obvious barrier to the same methods eventually reaching Navier-Stokes itself. To be clear, no one has claimed the $1 million prize; the full problem is still open.

Then it turned ugly. In a public statement, Buckmaster alleged that OpenAI—which he says told him an internal model had produced a roughly 100-page proof of blow-up for a forced version of Navier-Stokes, using much the same route he and Alpöge had quietly chosen—raced to announce only after word of their work reached the company. Pressed on the timeline, he says, OpenAI conceded its first prompt on the project went out just days after their results began circulating. He further alleges he was pushed to release without Alpöge as co-author, because Alpöge works at rival Anthropic, and that he could not get a straight answer on whether the private Codex sessions where he and Alpöge had stored a year of drafts were used to train OpenAI's models. Among the lines he attributes to the OpenAI side: "Why would you ruin your career?" and "If you don't want me to be nice, then I don't have to be nice." Critically, OpenAI's claimed proof has not been published, so no outsider can check it.

OpenAI's Sébastien Bubeck, the scientist named in the account, pushed back on X: "A series of false and inflammatory allegations against me are currently circulating… I came into the discussion following academic norms, and I'm disappointed that it has come to this." He promised a fuller response but did not address the specific claims, and OpenAI has neither released its proof nor formally replied.

Sources: Terence Tao · officechai — reporting · statements by Tristan Buckmaster and Sébastien Bubeck on X

Why it matters: Set aside who's right—the fight itself is the story. A landmark proof is prestige gold for a lab racing its rivals and, critics note, chasing the kind of narrative that lifts a valuation ahead of a public offering, and this is what that scramble can look like: a headline claim resting on a proof no one outside the company has seen, and an allegation that a competitor's employee was to be written out over his badge. For anyone deploying these tools, the quieter warning is the Codex question. Buckmaster says he can't confirm his year of private drafts wasn't absorbed into a rival's model—and, right now, neither can you. However this particular dispute resolves, it puts a hard question to every professional pouring proprietary work into an AI vendor's chat box: what actually happens to it?

Source: officechai.com

Twitter Privacy Tools Nitter and XCancel Return After Brief Shutdown

Nitter and XCancel—privacy-focused tools that let people view X (formerly Twitter) posts without logging in, ads, or tracking—resumed service after briefly shutting down, saying they consulted legal counsel first. Neither project disclosed what that advice actually said. Commenters online were split: some warned operators could still face legal exposure, possibly under computer-fraud statutes, and urged independent counsel; others pointed out the irony of X objecting to being scraped when it scrapes the web itself, and hoped these alternative front-ends keep surviving.

Why it matters: The dispute is a small preview of a bigger fight—who gets to control access to public-facing data on major platforms—that will shape how AI tools and third-party apps are allowed to read the web.

Discuss on Hacker News · Source: github.com

OpenAI Backs Program to Keep Ukrainian Newsrooms Running

OpenAI, publisher trade group WAN-IFRA, and a Ukrainian regional press association launched a joint program to help Ukrainian news outlets adopt AI tools. It includes a training series and a hands-on "Catalyst" cohort for ten Ukrainian news organizations, each receiving OpenAI API credits, aimed at keeping independent newsrooms financially viable and operational despite the ongoing war.

Why it matters: It's part of a broader pattern of AI labs using free access and training as soft-power tools to build goodwill and dependency with vulnerable institutions, including a press corps operating under wartime pressure.

Source: openai.com

AI Models Still Miss the Joke in Chinese Sarcasm, Study Finds

A new benchmark tests whether AI models can decode sarcasm, irony, and playful indirectness in Chinese social media comments—the kind of layered, culturally loaded language that means the opposite of what it literally says. Built from over 200,000 real posts into 4,735 test items, it found the best of eight LLMs scored 81.4% accuracy, with an average of 68.7% across all models, versus 90.8% for humans. Models often sensed something ironic was happening but misread exactly how or why.

Why it matters: As companies deploy chatbots and moderation tools across global markets, this is a reminder that AI's grasp of language is shakier the further it gets from literal, English-centric text—a real limitation for anything touching international customer sentiment or content moderation.

Source: arxiv.org

Popular AI Image Tool Misreads Users' Own Labels, Researchers Find

A research team studying how people make sense of large, messy datasets—like unlabeled image collections—found that users naturally organize things through overlapping tags and categories rather than sorting them into neat spatial clusters. They also tested CLIP, a widely used AI image-recognition model, and found it struggles to apply a user's own custom labels accurately, though it's decent at grouping images by general meaning. The researchers say better tools should let AI suggest categories and learn from just a few examples, while helping users judge when to trust its output.

Why it matters: As more professionals lean on AI to organize unstructured data—documents, images, customer feedback—this research is an early signal that current tools are good at broad pattern-matching but still need human judgment to build meaningful, custom categories.

Source: arxiv.org

AI Ethical Judgments Flip Based on How a Question Is Worded

A new study argues AI systems need basic structural consistency before questions of "alignment" with human values even make sense—and finds today's models don't have it. Researchers tested nine frontier AI agents on simulated moral dilemmas, tweaking only surface wording, escalation level, and framing. The result: verdict rates swung by up to 99 percentage points based on phrasing alone, and doing well on one moral scenario didn't predict doing well on a similar one. No model showed a stable, coherent policy across tests. A related study found that the after-the-fact explanations models give for their moral decisions hold up worse under structured cross-examination than their actual reasoning process does.

Why it matters: As companies lean on AI for judgment calls in hiring, moderation, and policy, this suggests both the decisions and the explanations behind them may be less stable and rigorous than they appear—a caution for any business deploying AI in decisions with legal, medical, or HR stakes.

Source: arxiv.org