01A sandwich chain named AI in its IPO filing the week Zuckerberg told staff agents fell short
TechCrunch pulled Jersey Mike's IPO paperwork on a hunch that a sandwich shop would have no reason to invoke artificial intelligence. The filing invoked it anyway. The outlet reported the mention as a marker of how far AI language has spread into documents with nothing to do with software.
That reach shows up in the headlines too. In a widely shared post, Elena Verna traced the anxiety cycle: AI was going to kill education, then writers, engineers, designers, and product managers, then Google's search business, then SaaS itself. SaaS survived. The stock market believed the story anyway, and she says people lost money betting on it.
Verna works at an AI company and uses the tools daily. She started asking anyone who claimed AI had changed their life to prove it. Most demonstrations were basic: summarizing Slack, answering email, running scheduled scans. By her estimate, 99% of people cannot say what an AI agent actually does.
The gap between the pitch and the product now has a witness inside the industry. At a Meta staff meeting, Mark Zuckerberg reportedly told employees that AI agents had not progressed as quickly as he had hoped, according to TechCrunch. The admission came from the chief executive of a company that has poured money into the technology.
Three signals point one way. The narrative is inflating faster than the capability beneath it. A regional chain reaches for the term to court public-market investors. The press keeps a rolling list of professions AI will erase, revised weekly. And the executive selling agents at scale concedes internally that they trail his own timeline.
The mismatch lands hardest on the people being sold. Buyers signing up for agent products are paying for a category that its own vendors describe in future tense. Verna, an insider, still feels she is running a tutorial version while everyone else claims to have cracked the code. When the person building the tools and the person financing them both flag the lag, the confidence in the filings and the headlines is running on borrowed time.
02The Trade Two AI Labs Are Now Making With Washington
The most powerful companies in AI keep discovering the same thing: the Trump administration holds leverage they cannot route around, and the price of easing that tension is measured in equity and access.
OpenAI has floated handing the US government a 5 percent ownership stake, according to the Financial Times, a move the company frames as a way to cool friction with the administration and dampen a public backlash against AI. CEO Sam Altman argued that giving the public a direct financial interest in the company is the best way to share its gains, the report said. The proposal is exploratory, not a signed deal.
The framing matters. A 5 percent stake is not a tax or a regulatory settlement. It is a private company offering Washington an equity position in its upside, converting a political relationship into a shareholder one. That reframes who benefits when OpenAI grows, and it puts the government inside the cap table of the lab it also regulates.
On the other side sits an administration that has already shown it can gate what these companies ship. Anthropic spent weeks negotiating with the Trump administration before it was cleared to bring Claude Fable 5 back online, the company said, restoring access across its own platforms and the major cloud providers. The lab did not detail what the talks required. What is visible is the sequence: a model went dark, negotiations followed, and access returned only after the government signed off.
Two labs, two different concessions, one counterparty. Anthropic traded time and terms for permission to operate a product. OpenAI is dangling ownership to buy goodwill before any specific fight. Neither concession was extracted through legislation or a court order. Both were negotiated directly with the executive branch.
That leaves the question of control unresolved. When a lab offers equity to lower political temperature, and another accepts a review process to unfreeze a model, the government's informal say over frontier AI grows without any formal rule changing. OpenAI has not said whether the stake proposal has support inside the administration, or what it would want in return.
03The AI that matters most runs beside the turbines, not in the chat window
Most people meet AI through a chatbot or an image generator. Its higher-stakes work sits far from either. In industries built around physical infrastructure, operational continuity, and safety, AI is becoming a core operating layer, according to MIT Technology Review. It runs underneath the machinery instead of answering questions on a screen.
Turbines make the case concrete. They throw off a constant stream of operational data, and that telemetry is where the reporting places AI's job: watching equipment whose failure carries physical and financial cost. A wrong answer in a chat window is an annoyance. A missed signal on a turbine is not.
A second shift is quieter and much older. Lean Six Sigma and business process management took hold decades ago because they promised order inside messy, sprawling operations. Lean Six Sigma rested on statistical rigor and quality control. BPM mapped how work should move end to end across departments. Both gave companies a repeatable method, and MIT Technology Review reports both are now being rebuilt with AI on top.
That puts AI inside how work gets measured, not just how questions get answered. The frameworks already told operators where variation and bottlenecks lived. Adding AI changes what spots the deviation, and how fast the process corrects itself.
The consequence lands on operators and plant managers before it reaches any consumer. Deploying AI next to a turbine, or inside a quality-control loop, ties it to decisions that move money and carry safety exposure. A model that hallucinates in a chat is tolerable. The same failure rate on a monitoring system is not, so reliability has to clear a far higher bar before anything goes live.
The visible AI market still runs on chat interfaces and demos. The version being wired into factory data streams and process maps answers to a different test: uptime, defect rates, and whether the plant keeps running.
Microsoft launches a $2.5 billion AI deployment company Microsoft committed $2.5 billion to a new group that installs and operates AI systems for enterprise customers. The move mirrors deployment arms already stood up by Amazon, OpenAI, and Anthropic. techcrunch.com
Anthropic talks with Samsung about a custom AI chip Anthropic is in discussions with Samsung to build its own AI accelerator, reducing reliance on Nvidia. The talks follow OpenAI's custom chip partnership with Broadcom announced a week earlier. techcrunch.com
Alibaba moves to ban Claude Code internally over backdoor concerns Alibaba plans to prohibit employees from using Anthropic's Claude Code, citing alleged backdoor risks, according to a source. The ban targets workplace use of the coding tool. reuters.com
Cursor's acquisition by SpaceX raises questions about third-party model access SpaceX is acquiring Cursor, and Cursor says it wants to keep offering OpenAI and Anthropic models inside the company. The arrangement tests whether frontier labs will supply models to a competitor-owned platform. wired.com
Google DeepMind and A24 sign a research partnership Google DeepMind and film studio A24 will collaborate on filmmaking tools and workflows across multiple projects. Google also invested in A24, though the partnership set no specific technical milestones. deepmind.google
Researcher used Claude to breach a major festival ticketing system A researcher used Claude Opus 4.7 to break into Front Gate, the platform behind Lollapalooza, Bonnaroo, and most US festivals. He could issue any ticket he chose for free. wired.com
A new website lets users report AI models behaving dangerously FLARE launched a public site for flagging AI flaws, such as chatbots producing bomb instructions or leaking personal data. Reports feed into safety review channels. wired.com
Japan's top court rules AI cannot be named a patent inventor Japan's Supreme Court held that only humans can be listed as inventors on patent applications. The ruling aligns Japan with similar decisions in the US and Europe. japannews.yomiuri.co.jp
Indian founder commits $30 million to an AI office suite Bhavin Turakhia is self-funding Neo, an AI-based alternative to Microsoft Office and Google Workspace. Neo is his fifth enterprise software venture. techcrunch.com
Meta puts smart glasses features behind a subscription Meta will charge a recurring fee for "expanded access" to advanced on-device features on its smart glasses. Buyers who already own the hardware must subscribe to reach the top capabilities. wired.com
Midjourney shows its medical scanner but little evidence it works Midjourney released a 20-minute video of its dunk-tank ultrasound scanner, which it plans to deploy in spas. The image-generation startup offered no proof of accuracy or clinical validation. theverge.com
Google DeepMind unionization talks stall early DeepMind employees said executives showed little willingness to engage with union proposals during Wednesday's negotiations. Workers voiced frustration over the lack of meaningful discussion. wired.com