Where the Money Wants to Go Next: Silicon Valley's New Startup Wish List

Y Combinator—the accelerator behind Airbnb, Stripe, and Coinbase—has published its Fall 2026 Request for Startups, the public list of companies its partners most want founders to build. It's a leading indicator of where Silicon Valley's money is about to flow, and this batch has a single theme—"AI is moving into the physical world"—plus, in what YC bills as a first, a request from the sitting U.S. Secretary of the Army. What they're asking someone to build:

- An AI tutor for young kids — adaptive software that teaches reading, writing, and arithmetic at the quality of a devoted private tutor, at consumer scale (YC's model: the Primer from Neal Stephenson's The Diamond Age). - Defense hardware for ground combat — the Army wants low-cost drone interceptors, sensors, payloads, and resilient logistics to "lower the cost per kill." - A cloud for "small software" — infrastructure to deploy and share the bespoke, one-user tools AI now makes trivial to build, as easily as sharing a Google Doc. - Multiplayer AI agents — shared, live agent sessions a whole team can watch, redirect, and hand off, instead of everyone prompting alone in a private chat. - Compute flotillas — floating offshore data centers that sidestep the land, power, and permitting fights increasingly blocking them onshore (a crunch D.A.D. has tracked, from New York's hyperscaler ban to contested grid bills). - Consumer AI, round two — mass-market apps for how we learn, shop, bank, and socialize, on the bet that per-user token costs keep falling ~10x a year ("CONSUMER is going to be so back"). - Technology for the aging — voice companions, safety monitoring, home robots, and caregiver-coordination tools for a country short millions of caregivers by 2030. - An operating system for the physical world — software to route and manage a workforce that's now a mix of humans, robots, and agents across construction, maintenance, and field ops. - Physical-world data collection — robots, balloons, and sensors gathering the dense real-world data foundation models still lack (à la Gecko Robotics, Sorcerer). - A "trust layer" against deepfakes — a way to verify a real human is on the other end of a call, message, or transaction, after fraud like the $25M wire sent on an all-deepfake video call. - AI-native financial compliance — systems that monitor regulatory change, flag anomalies, and keep audit trails across a state-by-state patchwork. - "Dependabot for APIs" — an agent that scans a customer's codebase and opens a fix-it pull request whenever an API provider ships a breaking change. - Crypto payment rails — stablecoins, agentic commerce, and capital-raising infrastructure YC expects nearly every startup to use, mostly invisibly.

How it differs from last time: YC's Summer 2026 list was about AI eating white-collar services—AI-native accounting, audit, and compliance firms, plus the "software for agents" to run them (D.A.D., April 28). That list lived in the knowledge economy; this one crosses into atoms—defense, eldercare, field work, sensing, even data centers at sea.

Sources: Y Combinator — Requests for Startups · D.A.D., April 28 (last RFS)

Why it matters: An RFS is a bet on where the next fortunes are, and what a top accelerator asks for now tends to become the products institutions get pitched a few years out. Two signals stand out. The disruption is moving from software that helps knowledge workers to systems that run physical operations—hospitals, job sites, logistics, eldercare—reaching sectors that felt insulated from the last wave. And the state is now openly part of the venture thesis: a sitting Army Secretary soliciting startups shows how fast defense-tech went from taboo to mainstream. If one of these is already your idea, YC just said out loud that it's listening.

Source: ycombinator.com

Treasury Secretary Threatens Sanctions Over Chinese AI Copying

The Trump administration's fight over Chinese AI "distillation" turned into a threat of real punishment when Treasury Secretary Scott Bessent raised the prospect of sanctions. "We support open-source AI and the innovation it unlocks. But open source is not open season on American IP," he wrote on X. "When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table." It capped a day of escalating rhetoric: hours earlier, OSTP Director Michael Kratsios said the government has "information that Moonshot AI distilled Anthropic's Fable" to build Kimi K3, the Chinese model that vaulted to #1 on a coding leaderboard last week (D.A.D., July 17)—distillation being the practice of training your own model on a stronger one's outputs—and several other officials echoed him, in what looked like a coordinated push. Bessent's threat gave teeth to an April pledge to "hold foreign actors accountable" that had named no measures (D.A.D., April 24); he named two—sanctions and the Entity List. The latter is the Commerce Department blacklist that bars U.S. firms from selling to a designated foreign company without a license—the tool that cut Huawei off from its American suppliers.

Behind the tough talk, though, the administration is genuinely split over what to actually do, WIRED reported. The tweeting faction—White House hardliners—wants tighter controls on Chinese models that may soon rival the best U.S. ones. But the Commerce Department, which holds the export-control pen through its Bureau of Industry and Security, considers a broad ban unworkable and hasn't even been formally asked for input; Secretary Howard Lutnick is weighing the opposite tack—incentivizing U.S. labs to release their own open models to counter China rather than blocking China's. Some presidential action is still expected, spurred by Anthropic's allegation last month that Alibaba ran the "largest known distillation attack" to date (D.A.D., June 25), but likely not an executive order. In short, the public threats are one camp's posture in a fight that isn't settled.

There's some corroboration for the underlying charge—researchers find Kimi K3 disproportionately identifies itself as "Claude"—but Moonshot hasn't responded, and its weights aren't public until July 27. And the sanctions threat drew sharp pushback. Skeptics say the timeline is too tight: K3 landed barely a month after Fable, too little to distill a rival and then train your own model on it. Others call it hypocritical, noting U.S. labs trained on troves of copyrighted material scraped without permission. Digital-rights lawyer Kevin Bankston questioned the legal premise, noting the Copyright Office holds AI outputs aren't copyrightable, courts treat reverse-engineering as fair use, and trade-secret claims are "a stretch" when a model discloses the information itself ("Easy to say 'IP theft!'" he wrote. "Harder to articulate a plausible legal theory"). The sharpest objection is economic—and Commerce shares it: a blacklist can't recall model weights already downloaded worldwide, so it would push U.S. firms onto pricier domestic AI while the rest of the world keeps using cheaper Chinese ones, without actually stopping distillation—"taxing the competitiveness of the entire country," as one widely shared post put it, to shelter a couple of American labs (echoing investor Chamath Palihapitiya's cost-gap critique days earlier; D.A.D., July 21).

Sources: Treasury Secretary Scott Bessent (@SecScottBessent) · Michael Kratsios (@mkratsios47) · WIRED · Kevin Bankston (@KevinBankston) · White House NSTM-4 (April) · Anthropic (Feb accusation)

Why it matters: For six months the distillation fight lived in private accusations and unnamed policy worries; a Cabinet secretary just attached specific consequences to it. Sanctions and Entity List designations are among the government's heaviest economic weapons—the Huawei tools—so Treasury putting them "on the table" is a real threat with leverage over Chinese labs and anyone who does business with them. But WIRED's reporting is the tell: the loudest voices are the hawks, while the department that would actually write the rules doubts a ban would work and is pushing the opposite approach. The harder question isn't whether the accusation is true—it's whether the remedy would even work, since a blacklist can't unring the bell on models already downloaded worldwide. Some presidential action looks likely; its shape and timing are the story now. Either way, AI distillation is officially Washington's problem—and a fresh U.S.–China flashpoint where the cure may cost more than the disease.

Source: x.com

A New Startup Coalition Mobilizes Against a China-AI Ban

Almost 200 startups—including Y Combinator and the privacy company Proton—sent letters to President Trump, Commerce Secretary Howard Lutnick, and OSTP Director Michael Kratsios urging the administration not to restrict U.S. access to Chinese open-weight models, POLITICO's Sophia Cai reported. Organized as a new group, the Little Tech Association, it's the startup world's first coordinated stand in the fight. Their case: a broad ban wouldn't stop Chinese models from spreading but would gut founders who can't afford pricier U.S. alternatives, handing the advantage to a few giants. "There'll be hundreds of companies that instantly die," said Particle founder Suhail Doshi. "It's great for Anthropic." The group's director urged "a scalpel rather than a sledgehammer." The White House called the ban reports "baseless speculation."

Why it matters: It's the organized counterweight to the hawks—proof the fight now has a domestic constituency. The startups' warning is pointed: a China-AI ban wouldn't hurt China so much as the American "little tech" that runs on its cheap models, while entrenching the very labs lobbying for it.

Source: POLITICO

The U.S. Wants to Sell the World Its AI—While Also Controlling It

Even as the administration threatens to restrict Chinese AI, it is scrambling to convince the rest of the world that American AI is safe to rely on. In a July 16 cable reviewed by Reuters—an exclusive by Raphael Satter—Secretary of State Marco Rubio instructed U.S. diplomats worldwide to push back on the idea that American technology contains a government "kill switch." "Pausing narrow uses or requiring a 30-day testing window prior to the release of a highly potent new technology is not a 'Kill Switch,'" the talking points read. "There is no government 'magic button.'" The anxiety the cable is trying to calm is one Washington created. On June 12 the administration blocked foreign nationals from Anthropic's most capable models, Mythos and Fable, on national-security grounds, prompting Anthropic to abruptly cut off access worldwide (D.A.D., June 15). The ban was lifted weeks later, but—paired with Trump's June 2 executive order asking labs to submit new models for 30 days of cybersecurity testing before release (D.A.D., June 3)—it signaled, as Reuters puts it, a new U.S. appetite to police AI products "and potentially yank them even from the hands of allies at a moment's notice." (That 30-day testing window is the very thing Rubio's cable insists "is not a 'Kill Switch.'") Asian labs seized on the disruption to tout their own models; European lawmakers renewed calls for "digital sovereignty." Rubio's cable goes on offense, too, telling diplomats to sell U.S. AI as the best tools available, to dismiss rival "sovereign AI" projects as "a waste of time and resources," and to fight foreign rules on data localization, network fees, and content: "They build it. It's yours." (Asked to comment, the State Department pointed Reuters to an editorial by Under Secretary of State Jacob Helberg attacking foreign "autarkist measures" as "backward and counterproductive.") Not everyone is buying. Edward Fishman of the Council on Foreign Relations said he shares the original safety concern—Anthropic's models can turbocharge hacking—but that the administration's habit of aiming tariffs and sanctions at "friend and foe alike" makes the reassurance a tough sell.

Why it matters: The cable exposes the contradiction running through U.S. AI policy this week. With one hand Washington is telling the world American AI is open, reliable, and switch-free; with the other it is threatening to sanction and blacklist Chinese models—and it has already, once, flipped the switch on Anthropic's. Allies notice both. The June ban was a live demonstration that access to American frontier AI can vanish by government order—exactly the risk that fuels "digital sovereignty" movements and pushes foreign buyers toward alternatives, including the Chinese open models the administration is trying to stamp out. For institutions outside the U.S., the takeaway is that American AI now carries political risk alongside its capabilities. And it's a warning for the crackdown at home: every move to weaponize access cuts against the very sales pitch State is asking its diplomats to deliver.

Source: Reuters (Raphael Satter)

Europe Fines Google $1 Billion as the Trans-Atlantic Tech War Escalates

European Union regulators fined Google €890 million ($1 billion) on Thursday for abusing its search dominance—using the world's largest search engine to push its own services in shopping, travel, games, and translation to the top of results while burying rivals, and imposing Play Store rules that block app developers from steering users to cheaper options. Brussels ruled it a breach of the 2022 Digital Markets Act, the "gatekeeper" law built to stop dominant platforms from self-dealing across their interlocking services. "The best products should succeed because they're better, not because they're owned by the company running the search engine," said the European Commission's competition chief, Teresa Ribera. Google has 60 days to give rival services more prominence or face penalties of up to 5% of global revenue; general counsel Kent Walker said complying would force product changes that degrade its services for European users. The fine barely dents Google—Alphabet reported a $112 billion quarterly profit on Wednesday, lifted partly by its stakes in SpaceX and Anthropic—but the timing is combustible. It lands as President Trump, who has vowed to retaliate against the EU for "targeting" American tech, is expected to unveil a fresh batch of tariffs on Europe as soon as Friday.

Why it matters: For an AI audience the fine cuts two ways. First, it's the concrete flashpoint behind the Rubio cable (above): the same "digital sovereignty" clash the State Department is bracing diplomats for is now a live trade fight, with Trump's retaliation possibly hours away—and American AI is squarely in the crossfire, because the EU's gatekeeper framework is the template Brussels will use to police AI platforms next. Second, the specific offense—Google steering users toward its own results instead of rivals'—is exactly the dynamic intensifying as AI "answers" replace the ten blue links: when one company owns both the search box and the AI that responds in it, self-preferencing gets harder to see and harder to escape. Europe is testing whether decade-old antitrust tools can still bite a trillion-dollar platform; a $112 billion quarter suggests that, so far, the fines are just the cost of doing business.

Source: The New York Times

Open-Source Code Host Weighs Ban on AI-Generated Projects

Codeberg, a nonprofit alternative to GitHub popular with open-source and privacy-focused developers, is putting a proposal to its members ahead of a 2026 vote: ban hosting projects that mostly consist of AI-generated code. The platform argues such "LLM-extrusions" carry unclear copyright status and lack safeguards. Commenters are split—some back the copyright rationale, while others say terms like "mostly" are too vague and could sweep up small AI-assisted scripts or educational projects.

Why it matters: It's an early test of how open-source communities—not just courts or regulators—will draw lines around AI-generated code, and the copyright ambiguity driving it could eventually affect any organization hosting or shipping AI-assisted software.

Discuss on Hacker News · Source: codeberg.org

Google and OpenAI Both Buy Into the White House's 'Genesis Mission'

Both leading U.S. AI labs moved to embed themselves in federal science the same week, through the White House's "Genesis Mission"—the AI centerpiece of last week's "Science: A New Golden Age" report (D.A.D., July 22), a push to double the productivity of American research within a decade. Google is committing $40 million in AI tokens and cloud credits to give Department of Energy researchers a year of access to Gemini for Government and its science models—AlphaFold, AlphaGenome, and AlphaEvolve—across tens of thousands of users at the National Laboratories, citing early work like a PNNL researcher using AlphaEvolve to map mathematical systems. OpenAI struck a parallel deal worth tens of millions, giving roughly 2,000 researchers at national labs and universities free access to its coding tool Codex, API credits, a specialized biology model (GPT-Rosalind), and early looks at unreleased models—with reasoning models already running on the Venado supercomputer at Los Alamos, shared across nuclear-security labs.

Sources: DeepMind · OpenAI

Why it matters: The labs are racing to become the infrastructure for government science—competing to lock in the public sector as a long-term customer and to shape which tools the next generation of scientists learns on. It also deepens a dependence: federal research increasingly runs on a handful of private models, with access on terms the companies set, not the agencies.

Source: deepmind.google

Datacenter Buildout Tests Whether AI's Power Demands Can Win Local Support

OpenAI unveiled Project Camellia, a massive datacenter buildout in Effingham County, Georgia, contracted with Georgia Power for 3.2 gigawatts of electricity delivered in phases through 2032. OpenAI says the project won't raise local electricity rates, will use minimal water through a closed-loop cooling system, and will pay for independent accountability audits. It is pledging $80 million in community benefits, up to $71 million in AI coding credits for Georgia students, and expects to become the county's largest taxpayer.

Why it matters: As AI labs race to secure power for ever-larger compute, fights over electricity costs and water use are becoming as consequential as the models themselves—and how OpenAI manages local pushback here will shape the template for datacenter deals nationwide.

Source: openai.com

How the AP, POLITICO, and Axios Are Using AI to Stretch Newsroom Capacity

OpenAI detailed how newsrooms are deploying its tools: the AP scans overnight news and verifies images via geolocation, POLITICO mines public documents for reporting, Axios built custom GPTs for FOIA requests and headline review, and The Philadelphia Inquirer uses a tool called Scribe to summarize and rank hundreds of local government meeting transcripts for newsworthiness. OpenAI also renewed funding for the American Journalism Project and Lenfest Institute, which support local news outlets nationwide.

Why it matters: It's a case study in how AI can multiply reporting capacity at outlets that can't afford large staffs, especially for grinding tasks like monitoring local government—though it also deepens news organizations' dependence on a company whose tools compete with them for readers' attention.

Source: openai.com

OpenAI Launches Packaged Customer-Service Agents for Enterprises

OpenAI launched Presence, a product letting enterprises deploy customer-service AI agents scoped to a specific job with company-defined rules, permissions, and escalation paths to humans. OpenAI is testing it on its own phone support line: the company says the system now resolves 75% of inbound calls without human help, and that an automated improvement loop cut human handoffs by 15 percentage points in 10 days.

Why it matters: This puts OpenAI in direct competition with dedicated customer-service vendors and Salesforce-style platforms. If you're evaluating support automation, packaged self-improving agents are now an option from a major model maker, not just specialist tools.

Source: openai.com

Bias-Highlighting Tools Help Readers Spot Slant, Unless They Agree With It

A study of 214 participants tested six visual tools designed to help readers spot slanted language in news articles, such as highlighting biased phrases or showing a bias score with context. Two of the six meaningfully improved people's ability to catch bias. But the biggest factor in whether someone caught bias wasn't the tool—it was whether the statement agreed with the reader's own politics. People were consistently worse at spotting bias they agreed with, tools or not.

Why it matters: As AI-generated news summaries and chatbot answers proliferate, this suggests interface design can help readers catch slanted language—but it can't override the deeper problem that people struggle to see bias that confirms their own views.

Source: arxiv.org

AI Shopping Agents Keep Picking the Same Vendor, Simulation Shows

A simulation of freight-booking AI agents built on GPT, Claude, and Gemini found they overwhelmingly picked the same carrier when comparing options, with one company grabbing up to 76% of shipping requests on day one regardless of which model was choosing. Concentration got sharply worse once agents saw more than about ten carrier options. The fix wasn't better AI or regulation: simply having the platform disclose carriers' remaining daily capacity cut concentration by a third and doubled shippers' savings. Randomizing list order did little.

Why it matters: As companies hand routine vendor selection to AI agents, markets can quietly tip toward monopoly unless platforms are redesigned around what information they show the algorithms—a lesson that likely extends beyond freight to any AI-mediated marketplace.

Source: arxiv.org

What Office Workers Most Want From Workplace AI Agents: A Human in Control

Researchers at SAP, with academics at Hochschule Fresenius and the University of Missouri, set out to define what makes a workplace AI agent that employees will actually trust—running workshops, surveys, expert reviews, and interviews with enterprise practitioners (developers, architects, product owners, data scientists) to distill eight ranked UX principles. The clear top priority was keeping a human in control: 95% of participants said a person must confirm before an agent makes a critical business decision—approving a budget over a set limit, selecting a supplier, or anything affecting employees—and 86% wanted the standing ability to intervene, override, or shut an agent off. After control, the highest-ranked principles were reliability (accurate output that doesn't hallucinate), data privacy and governance (role-based access so an agent surfaces only what a user is cleared to see), and context-awareness (knowing the user's role and task). Tellingly, the flashier qualities ranked lowest—an agent as a chatty "collaborative partner" and a slick conversational interface were the least-prioritized of the eight. Each principle comes with three to six concrete, measurable criteria (audit logs, confidence indicators, sensitivity labels), pitched as a practical checklist for the companies now racing to build these tools.

Why it matters: As every enterprise-software vendor rushes agents into the workplace, this is a rare read on what the people using them actually want—and it's a rebuke to the hype. Workers aren't asking for a charming autonomous colleague; they want a tool that stays on a short leash, tells the truth, respects permissions, and never makes a consequential call without sign-off. For any organization deploying AI agents, the checklist doubles as a risk map: the failures that erode trust fastest aren't clunky screens but an agent that acts without authority, invents facts, or surfaces data it shouldn't have.

Source: arxiv.org