> daily_signal(2026_07_24)

The FAA cleared over 1,000 police agencies to fly drones out of sight, on a form no city council sees. Abbott turned on his own data centers, ChatGPT Business cannot export, DOE named 278 projects.

PickBits Daily Signal · Friday, July 24, 2026

The FAA has cleared more than 1,000 police agencies to fly beyond the operator's line of sight, more in ten months than in the previous seven years, on a federal form with no public-comment step and no local vote. Meanwhile the data-center fight we picked up on May 13, when one Texas county tried to stop a project, has turned into Greg Abbott promising to strip the tax breaks off the $40 billion Google buildout he announced himself, with Ohio collecting signatures to ban construction outright. The ChatGPT one surprised me most: back in June, when a frontier export ban lasted eight days, we said leaning on a single AI vendor was a supply risk because the tool can go dark on somebody else's decision, and it turns out the tiers companies pay the most for cannot get their conversations out at all, while a $20 personal account can. Then there is the good one. DOE named 278 Genesis Mission awards with 168 of them going to universities, and NIH opened its pediatric-cancer data ecosystem to DOE supercomputers. That is public money buying compute for university labs and pooling cancer data no company had a reason to pool, a day after Anthropic's rare-disease grants did a smaller version of it.

1. The FAA cleared more than 1,000 police agencies to fly drones beyond the operator's line of sight, and no city council had to vote on it.

The FAA cleared more than 1,000 police agencies to fly beyond the operator's line of sight. No city council voted on any of them. The approval is an administrative waiver. A federal form. No public-comment period, no local hearing, no requirement that a city council ever hear about it. A drone answering a 911 call is useful, and almost nobody argues otherwise. The model is called Drone as First Responder: a department mounts drones on fixed rooftops around a city, a 911 call launches one automatically, and the aircraft is overhead before a patrol car leaves the lot. One department running a program says it can reach anywhere inside its drone's range in under two minutes. The hardware was never the constraint. The constraint was the FAA rule requiring an operator to keep the aircraft in visual line of sight, which meant one human standing outside per drone. In April 2025, the FAA streamlined the Part 91 waiver process for Beyond Visual Line of Sight operations, and the approvals came fast. EFF's July 23 analysis puts the count at more than 1,000 public-safety agencies holding waivers as of February 2026. The seven years before that, 2018 through April 2025, produced 976 in total.

A department that would never get a helicopter through a budget meeting now has drones on rooftops for a fraction of the price. The record of what they capture on the way to a call, over back yards and protests and everyone who is not the subject of it, is governed by whatever retention policy the department wrote for itself. This is not new ground for us. On July 19, we covered Flock Safety pulling back an AI distress-detection feature after public pressure, and on July 22 the DHS plan to grow its AI surveillance towers from 830 to 2,300 by 2034. Both became arguments in public. This one has no public step to argue in. Nothing happened to a specific person this week, which is exactly the problem, because when people go looking for a victim and cannot find one they stop paying attention. What happened is that a thousand departments got permission and nobody outside them signed off.

The departments running these programs say the camera sits at the horizon on the way to a call and only launches on a 911 dispatch. I don't doubt they wrote that policy in good faith. They also wrote it about themselves, and nobody outside the department ever voted on it. EFF's Atlas of Surveillance is free and takes about a minute to search; look up your own agency, because with a thousand waivers out there, "mine probably doesn't have one" isn't a safe assumption anymore. A listed program leaves paper: the waiver itself, a standard operating procedure that says when a drone launches, and a flight log covering the past twelve months. All three can be requested by name. The clause I'd actually fight for at a council meeting is automatic deletion of footage that never became evidence, on a fixed schedule. That single line decides whether a year of video of your street still exists.

Sources: eff.org, July 23 · atlasofsurveillance.org

2. The governor who announced Google's $40 billion Texas data centers now wants their tax breaks gone.

On May 13, we covered the first Texas county to try stopping a data center. The fight is a statewide electoral one now, and the clearest evidence is Greg Abbott. Abbott championed the buildout and personally announced Google's $40 billion Texas investment. Texas is projected to pass northern Virginia as the world's largest data-center market by 2030, and Google is building four data centers in northwest Texas, including a facility near Amarillo billed as the world's largest. Abbott has now reversed. He ordered regulators to ensure Texans do not pay higher electricity bills because of data centers, promised a legislative agenda to regulate them, said he would eliminate their state tax breaks, and proposed blocking construction in "rural Texas neighborhoods." His Democratic opponent, Gina Hinojosa, calls that last phrase "doublespeak that makes no sense" and has made the issue central: "They are owned by the richest men in the world. We're all footing the bill. There are no rules. It is the wild west of data centers." In Ohio, Sherrod Brown is running television ads calling Sen. John Husted "the face of data centers in Ohio" while citizens gather signatures for a statewide referendum to ban data-center construction. In Wisconsin, gubernatorial candidate Francesca Hong pledged to "stop AI data centers." In Pennsylvania, Gov. Josh Shapiro is taking fire from his own side over farmland.

The opposition is not partisan, either. Texas Agriculture Commissioner Sid Miller, a Republican, opposes the projects. The Lubbock-area protest organizers, Stephen Sanders and Hallie Bertrand, are Republicans who say they may not vote for Abbott and are not sure they can vote for Hinojosa either. Giles Dalby, a Republican county commissioner and cattle rancher, worries that data centers will get priority in water rationing in a region where shortages already hurt the sugar-cane industry. Clayton Tucker, the Democratic nominee for state agriculture commissioner, calls it "the most cross-partisan issue I've ever seen." The jobs and the tax revenue are real, and these projects land in towns that have watched their main streets empty out for twenty years. My own read is still that the ratepayers win this one, because a jobs announcement lands once and a utility bill arrives every month.

What used to gate a project was power and chips; now it's whether the state still wants you there, and in four states both parties are answering no. The bill for the new transmission and generation gets split somewhere, too, and it's not the zoning hearing. It's the state utility regulator's cost-allocation docket, a proceeding that's open, barely attended, and exactly what Abbott just told Texas regulators to look at. In Ohio, the parallel lever is the referendum signature drive. Anyone siting capacity right now wants three answers in writing before committing: whether the tax abatement in the model is still politically alive, where the facility sits when water gets rationed, and whether the county has actually voted. Agreeing terms with a town before it organizes against you is a lot cheaper than after.

Sources: clickorlando.com, July 23 · bostonglobe.com, July 23

3. The ChatGPT tiers companies pay the most for are the only ones that cannot export their own chats.

A $20 ChatGPT account can export every conversation it's ever had. A ChatGPT Enterprise workspace, the kind a company pays for by the seat, can't. That's the product OpenAI is shipping right now: chat export for Free, Plus, Pro, and some Edu accounts, no export for Business or Enterprise. What the paid workspaces get instead is the Compliance Platform, an audit and eDiscovery surface whose retention stops at 30 days. The Register asked OpenAI why. It got no answer.

So a freelance journalist named Conrad Quilty-Harper built a workaround and put it on GitHub this week. Scrapemychats drives a logged-in browser session, walks the conversation list thread by thread, and saves everything with its attachments as offline HTML. It runs slowly, because OpenAI throttles the requests. He isn't pretending it's sanctioned: "It's probably a breach of their terms of service. But hey, their entire business model is based around scraping copyrighted material."

Every other business tool I've bought works the other way around: the company gets the bulk export and the admin controls, and the individual gets whatever the product decides to give them. Here it's backwards. The individual walks out with the archive and the company can't. Companies have been running two or three years of working memory through these workspaces without checking whether any of it can come back out. Years of an engineering team's design decisions, sitting in a workspace whose only sanctioned exit expires after 30 days. Nobody finds this during procurement. They find it during a migration, or a litigation hold. Back on June 20, when a frontier AI export ban lasted eight days before being rescinded, we wrote that depending on a single AI vendor makes the supply risk explicit, because the tool can go dark on somebody else's decision. Same risk, other half. Enterprise is the ground rivals are fighting hardest for right now, alongside coding, and what keeps an enterprise customer in place is that its data doesn't leave with it.

Before the next renewal, I'd put one question in writing: can we export our workspace data, and in what format? Then check the answer against your own tenant, not the docs. "Customer owns its data" has never once produced a file; a clause that produces one spells out what gets delivered, in what format, and how many days after cancellation you can still ask. Decide in writing whether these workspaces are a place for real records or a scratchpad. And keep the scraper away from a corporate tenant. Its own author says it probably breaks the terms, so that decision belongs to counsel, not IT.

Source: theregister.com, July 23

4. DOE named 278 Genesis Mission projects, and NIH opened its pediatric-cancer data to DOE supercomputers.

Then the good one. Most rare childhood cancers have too few patients anywhere to support a study; no single hospital sees enough cases of any one subtype to run one. The cases exist. They're scattered across hundreds of hospitals, and this week the federal government started pooling them. NIH launched its Bio Genesis Mission, carrying more than $1.2 billion in obligated FY2026 and planned FY2027 funding, and the concrete move is a data move: HHS is opening its integrated pediatric-cancer data ecosystem and its national network of cancer centers so DOE supercomputers can train models across hundreds of rare cancer subtypes at once. Nobody announced a model. They moved the data, which is the part that was actually missing.

The scale comes from the Energy Department. On July 22, Energy Secretary Chris Wright named the first 278 projects under DOE's Genesis Mission funding call: $293 million across 342 institutions, and DOE called it the largest response to a funding opportunity in its history. The number I keep coming back to is the leadership split. Of the 278 awards, 168 are university-led. Eighty-seven are led by national labs. Federal AI money doesn't usually flow that way; normally the labs and the big contractors take most of it. Awardees also get access to the Genesis Mission Platform, which means agent frameworks, industry-supplied models, and DOE supercomputer time, and for a university group that could never buy that compute on its own, the access matters more than the check.

The caveat belongs up top, not buried at the bottom. "Selected for award negotiations" isn't funded and running. The White House OSTP science blueprint released the same week drew a mixed reaction from researchers for weighting AI over life-sciences funding, so FY2027 appropriations decide whether this grows alongside biology money or at its expense. What I want to see next: do those 168 university awards get real frontier compute, or a queue position? For the pooled pediatric data, which consent framework covers it, and do findings travel back to the treating hospitals? NIH has published the program scope and its six challenges, so both are checkable today. Anthropic closed yesterday's edition with private credits for rare-disease researchers. A company can fund the researchers. Opening a national hospital network's cancer data is a thing only a government could do.

Sources: energy.gov, July 22 · energy.gov, funding · nih.gov · whitehouse.gov

» What I'm watching

Ohio first: does the construction-ban referendum qualify for the ballot? That's town-hall anger trying to become a statewide vote, and I'd put it at the top of any siting team's watchlist. Texas has the quieter test, because the legislature still has to repeal the abatements Abbott promised to end, and a campaign promise isn't a repeal passed in a session. On the drone side, I'm waiting for the first city ordinance written specifically for these programs; the tell will be whether it names dispatch criteria, deletion schedules, and an annual flight report. The FAA's proposed limits on drone flight over critical infrastructure drew a First Amendment objection from the National Press Photographers Association on July 20, and those rules could end up regulating journalists more tightly than police. Watch whether Scrapemychats stays up; the fastest way to make an export button appear is a tool that embarrasses the absence of one. And the Genesis awards have to move from selection to signature before any of the cancer work touches a patient.

Tomorrow's signal lands here.