01Barret Zoph rejoined OpenAI in January to lead enterprise sales. Five months later he's out.
Barret Zoph ran enterprise AI sales at OpenAI. He came back in mid-January, The Verge reported, after a stretch as co-founder and CTO of Thinking Machines Lab. Five months into the second tour, he has left again.
Thinking Machines Lab competes with OpenAI. Mira Murati, OpenAI's former CTO, founded it and pulled talent out the door when she went. Zoph was one of those people. His January return reversed that exit, and his second stint did not outlast the spring. The Verge's report does not state why he departed.
Traffic at OpenAI runs in both directions. As Zoph walked out, Noam Shazeer announced he is joining the company, leaving Google after years there. "It was a difficult decision to move on," he wrote, calling the move away from his Google team an honor and a pleasure. He praised what they had built before signing off.
So the headcount stays roughly level while the names rotate: one senior leader back out within months, one prominent hire walking in from a rival. The faces change faster than the org chart does.
Even the attempt to dramatize OpenAI's instability has stalled. Luca Guadagnino's "Artificial," a film about CEO Sam Altman, has reportedly been dropped by Amazon MGM, according to The Verge. Andrew Garfield stars. The project had been in development for about a year. It covers the five days in 2023 that spanned Altman's firing as CEO and his reinstatement, the same boardroom rupture that scattered executives across the industry in the first place.
That 2023 turmoil is what put people like Zoph and Murati in motion. The film meant to capture it spent roughly a year in the works before being shelved. Its subject keeps generating new exits and entrances faster than any production can lock a final cut.
02Two teams hit the compute ceiling from opposite ends, days apart
Subquadratic, a Miami-based startup, came out of stealth last month claiming it had solved a mathematical bottleneck that constrained large language models for nearly a decade. According to MIT Technology Review, the details were thin and many onlookers stayed unconvinced. The company has since begun sharing data to back the claim. It remains a claim: a young startup asserting it cracked a structural limit on how much compute a model needs.
That same window produced a quieter version of the identical bet. A research group released Moebius, an image-inpainting framework of 0.2 billion parameters that, the authors report, matches the output of 10-billion-parameter industrial models. The paper frames the problem the way deployers feel it: 10B-class foundation models have pushed inpainting quality forward, but their compute costs block practical use.
So one group attacks the cost ceiling through architecture math, and the other through aggressive compression of a specific task. Both target the same wall. Running frontier-grade capability is too expensive to deploy widely, and each team says it found a way under that price.
The two routes carry different risks. Subquadratic's pitch is general and unproven; a subquadratic architecture, if real, would lower inference cost across many workloads, which is also why skeptics want more than early receipts. Moebius is narrow and measurable. Compressing 10B-level behavior into 0.2B triggers what the authors call a severe representation bottleneck, and their fix is a rebuilt diffusion backbone tuned for one job: filling in missing pixels.
For anyone shipping models, the distinction is the whole story. A specialist that runs 50 times smaller is something a team can test against its own inpainting pipeline this quarter. A general architecture breakthrough is worth far more and proportionally harder to trust before independent replication. The Moebius weights and method are published for inspection. Subquadratic's evidence is still arriving in pieces, on the company's own schedule.
What separates them next is verification, not ambition. One team's claim can be checked by downloading it. The other's depends on whether outside researchers reproduce results the startup has only started to release.
03The pitch says AI makes engineering easier. The people shipping it say it demands more discipline
The default position through most of 2025 held that AI-generated code was slop and might always be slop. A widely circulated essay argues that question got settled last November, and that the moment now resembles the earlier move from handcrafted "server pets" to immutable infrastructure. That transition did not reward less rigor. It punished it.
The post is a reply to readers who took an earlier piece as permission to skip code review and push unread code straight to production. That was never the claim, the author writes. The technical argument runs the other way: as generation gets cheaper, the surrounding engineering gets more demanding. Review, testing, and the controls that keep output from running unchecked become the actual work.
The essay frames this as one front in a longer list it still wants to cover, including AI mandates, communication norms, and code review itself. Its prior piece carried the line that AI enthusiasts race against time while skeptics race against entropy.
The cost shows up elsewhere too. OpenAI released new spend controls and usage analytics for ChatGPT Enterprise, which it says help organizations manage costs as they scale AI use. Stripped of the framing, large buyers needed a way to see where the money goes and to cap it. The model's capability was not the bottleneck. Accounting for what it consumes is.
Two stories collide here. One says AI strips work out of building software. The other, from the people running it in production, says it relocates the work. Generating code stops being the hard part. Reviewing it, testing it, and paying for it become the new floor.
That reframes who carries the load. The first to feel it are engineering managers and finance teams now expected to instrument per-user usage, set budgets, and defend the bill. The next signal to watch is whether spend dashboards become a standard line item in enterprise AI contracts, the way cloud cost tooling did after compute went elastic.
ASML disputes US claim that its top EUV tool sits in China The US government said ASML's most advanced chip-making machine may have reached a Chinese customer. ASML denied it, noting the export license it would forfeit outweighs any sale. techcrunch.com
Reliance plans to push AI into telecom services for 500 million users Mukesh Ambani wants AI embedded in calls, apps, and home services across Reliance's customer base. The rollout targets India's largest mobile and broadband network. techcrunch.com
Bernie Sanders proposes a $7 trillion fund to give Americans a stake in AI The senator unveiled a plan to channel AI-industry wealth into a public fund. The largest AI firms would face new levies to finance it. arstechnica.com
OpenAI ships GPT-5.5 Instant to handle health questions in ChatGPT OpenAI updated ChatGPT's health responses with stronger reasoning and physician-reviewed evaluations. The company says answers now weigh user context before replying. openai.com
Midjourney unveils a full-body ultrasound scanner as its first hardware CEO David Holz showed The Midjourney Scanner, a ring of sensors that captures vertical ultrasound images. He also disclosed plans for a San Francisco spa. theverge.com
Elastic agrees to acquire Deductive AI for up to $85 million Elastic is buying the three-year-old startup, which uses AI to find and fix software bugs. CRV had backed the company. techcrunch.com
Taiwan expands drone production for its own defense and US military buyers Taiwan is raising drone spending amid pressure from China. Officials see the buildout as a path to export sales abroad. arstechnica.com
Amazon engineers say the company is firing them for backing data center limits Three software engineers testified at Seattle City Council in favor of data center caps. They now accuse Amazon of retaliating, citing a city law against political-speech discrimination. theverge.com
UK will use facial age-estimation on asylum-seekers despite flawed tests The Home Office's own trials found the technology produces life-altering errors. The agency is deploying it for age checks anyway. wired.com
Tech-worker PAC Guardrails enters AI politics with a $5 million fund The group runs on small donations from AI-industry workers. It faces Big Tech political spending that reaches roughly $100 million. techcrunch.com
Researchers release Kairos, a world-model stack for physical AI systems Kairos uses a native pre-training method drawing on cross-embodiment data to maintain persistent state over long horizons. The team targets real deployment constraints rather than passive video generation. huggingface.co