The Download: AI's self-improvement limits, OpenAI's Astra pause, and Unitree's surge
The lead item in this edition of MIT Technology Review's daily newsletter, The Download, reports on a new study that pushes back against the AI industry's most ambitious current promise: that AI models will soon be able to improve themselves with little need for human oversight. Researchers found that AI agents still cannot conduct open-ended AI research, meaning free-form investigations without clear-cut answers that require the judgment and creativity genuine breakthroughs need. The item, credited to Michelle Kim, frames the open question as whether that kind of open-ended research is essential to recursive self-improvement, or whether AI systems might get there anyway by grinding away at narrower, better-defined tasks. The newsletter says the results may temper claims that recursive self-improvement is close at hand.
The second lead item, adapted into an episode of the weekly MIT Technology Review Narrated podcast on Spotify and Apple Podcasts, explains what has driven this summer's extreme heat across the Northern Hemisphere and why 2027 could be worse. June and July were Europe's hottest two-month stretch since record-keeping began, the contiguous United States had its hottest month on record in July, and South Korea recorded its highest-ever temperature. Climate change makes heat waves more likely and more intense, the piece says, but a second factor is also at work: El Niño, which is already building and is expected to have a bigger effect on global temperatures next year than it has this year.
Topping the newsletter's numbered must-reads list, OpenAI has paused some model work over safety concerns, saying its Astra model reached a "critical" risk threshold, according to the Guardian. Linked coverage cited by the newsletter says the slowdown sets OpenAI's approach apart from Anthropic's (Axios), that OpenAI has made security updates following the Hugging Face hack (the Verge), and that the company has introduced a version of ChatGPT built for teenagers (AP News).
Also on the list: Chinese humanoid robot maker Unitree surged 629% in its stock market debut, which Bloomberg's coverage says underscored China's lead in humanoid robots; Reuters reporting cited by the newsletter adds that Unitree has become a key player in the Sino-US tech war, and BBC coverage describes it as the world's biggest humanoid firm and already profitable. Separately, the Financial Times reports that China is now allowing Nvidia's H200 chips into the mainland, with ByteDance and Tencent each recently receiving about 10,000 of the processors; a linked Wired piece adds that a new Chinese AI model could help both defenders and hackers.
The rest of the list covers Flock's newly expanded AI tool, which Wired reports lets police identify drivers by their movements rather than just their license plates, and a landmark trial in which prosecutors argue, on its first day, that Meta deliberately hooked children on Facebook and Instagram (BBC). It also notes that ICE has banned Meta's smart glasses from its own workplace over privacy concerns (the New York Times), and cites a study, reported by 404 Media, finding that X's algorithm learns what a user hates and shows more of it, with the effect measured as stronger among Democrats. Two lighter items close out the list: a physicist's proposal, per New Scientist, that the universe may repeat itself forever, with the same galaxies, lives and events said to be able to recur endlessly, and a Reuters item on a blind Egyptian developer who has built an AI app described as helping others "see," using cameras to answer questions about users' surroundings.
The newsletter's quote of the day comes from Stacey Morris, a Kentucky mom whose son brought home a packet of AI-generated educational materials, telling local station WDRB about some of the errors in it: "Arizona is Arizone. Illinois starts with a V. I mean, it's crazy."
The issue's longer "One More Thing" feature, by Antonio Regalado, profiles Palestinian stem-cell scientist Jacob Hanna. When he was stopped while entering the United States last May, the questions from agents touched on a specific new topic: embryos. Hanna did not have any specimens on him, but if he had, it would have been hard to say what they were, since he specializes in creating synthetic embryo models: structures that resemble real embryos but do not involve sperm, eggs or fertilization. The piece says these models could provide unprecedented views of the earliest stages of human development while also creating a new way to produce cells for transplant medicine, but that as the models become more realistic, they raise difficult questions about how far the science should go.
Key facts
- A new study finds AI agents still cannot conduct the open-ended, judgment-driven research that recursive self-improvement would require, which the newsletter says may temper claims that self-improvement is close at hand.
- OpenAI has paused some model work after its Astra model reportedly reached a "critical" risk threshold, a move the newsletter says sets its approach apart from Anthropic's.
- Chinese humanoid robot maker Unitree's stock surged 629% in its market debut, which Bloomberg's coverage says underscored China's lead in humanoid robots.
- China is now allowing Nvidia's H200 chips into the mainland; ByteDance and Tencent have each recently received about 10,000 of the processors.
- This summer brought Europe's hottest two-month stretch on record, the contiguous US's hottest month on record, and South Korea's highest-ever recorded temperature, with El Niño expected to push 2027 even higher.
Why it matters
The lead item challenges the AI industry's most ambitious current claim, that AI models will soon be able to improve themselves with little need for human oversight: if agents still cannot handle the open-ended, judgment-driven research the study describes, recursive self-improvement may not be as close as some claims suggest. The same edition reports OpenAI's own pause on work tied to its Astra model, after the system reportedly reached a "critical" internal risk threshold, a move the newsletter says sets OpenAI's approach apart from Anthropic's. The edition's second lead item explains what is driving this summer's record heat across Europe, the US and South Korea, and why El Niño is expected to push global temperatures even higher in 2027.
Who it affects
AI labs and investors positioned around near-term recursive self-improvement, which the study's framing suggests may be further off than boosters claim; OpenAI, whose Astra-linked pause the newsletter contrasts with Anthropic's approach, and Anthropic itself by that comparison; Unitree, whose stock jumped 629% on its market debut, and China's wider humanoid-robotics sector, which the IPO is described as underscoring; ByteDance and Tencent, now receiving Nvidia H200 chips under China's new mainland access policy; police departments using Flock's expanded driver-tracking tool, and the drivers it tracks; Meta, facing the child-safety trial and the ICE smart-glasses ban; and stem-cell scientist Jacob Hanna, whose synthetic embryo research the newsletter profiles at length.
How to use it
For readers following recursive self-improvement claims, the study is a data point for caution rather than confirmation: agents in it still cannot handle open-ended, judgment-driven research, which the newsletter says may temper expectations of an imminent breakthrough. For those tracking AI safety governance, the Astra pause is a case study in how one lab responds to hitting its own internal risk threshold. Readers wanting the full self-improvement study, the heat and El Niño explainer, or Jacob Hanna's embryo research should follow the newsletter's links to the underlying MIT Technology Review pieces and the Narrated podcast, since this edition only summarizes them; the must-reads list functions the same way, pointing to outlets including Bloomberg, Reuters, the BBC, the Financial Times and Wired rather than offering original Technology Review reporting.
How solid is it
The self-improvement item cites "a new study" and unnamed "researchers," without naming the institution, the authors or the paper in this text. The Astra pause and its "critical" risk threshold are attributed to OpenAI itself, reported here via the Guardian rather than through Technology Review's own reporting. The Unitree, Nvidia H200, Flock, Meta trial, ICE, X-algorithm, universe-repeat and blind-developer items are each sourced by the newsletter to another outlet, among them Bloomberg, Reuters, the BBC, the Financial Times, Wired, 404 Media, the New York Times and New Scientist, rather than verified independently in this text. Michelle Kim is credited on the self-improvement item and Antonio Regalado on the Jacob Hanna profile; both read as original Technology Review reporting rather than links to outside outlets.
Risks and caveats
No institution, named researchers or methodology accompanies the self-improvement study beyond "a new study" and "researchers." Unitree's 629% debut surge comes with no dollar figures, valuation or other IPO detail, and no date is given for when China began allowing Nvidia's H200 chips into the mainland beyond "recently." The Meta children's-safety trial is named only as "a landmark trial," with no court or location given. The must-reads list runs 1 through 10, but only nine distinct headlines appear in the source text: "4" and "5" merge into a single entry ahead of the Flock item, so a separate fourth headline may be missing from what this edition captured.
“Arizona is Arizone. Illinois starts with a V. I mean, it's crazy.”
— Stacey Morris, a Kentucky mom whose son brought home a packet of AI-generated educational materials, to local station WDRB