Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and we've got a fresh menu of AI morsels today, plus a couple of side dishes I couldn't leave off the plate. Grab your coffee, pull up a chair, and let's have at it.
Here's what caught my eye first. There's a story out of TechCrunch about a company called Thrive Holdings, and once I started pulling on the thread, I realized it's not really a funding story. It's a story about a new kind of business that's quietly becoming the most interesting play in this whole AI economy — not building the models, not building the chips, but going out and buying boring, old-fashioned companies and jamming AI into their guts. Accounting firms. IT shops. That's the deep dive today, and we'll get there. But let me lay out the table first, because there's a lot to chew on.
Let me start with a story that has nothing to do with data centers or agents, and everything to do with the strange world we're living in now. Hundreds of Homeland Security drones — the ones that normally sit on the southern border watching for people crossing — got retasked this summer to hunt flesh-eating parasites. I'm not making that up. Now, I want to be careful here: this is a story that surfaced fresh but the underlying operation goes back to June, so treat it as an older story catching new attention rather than something that broke this morning.
Here's the deal. There's a bug called the New World Screwworm. It's a fly larva that hatches in open wounds on warm-blooded animals — cattle, sheep, dogs, and yeah, humans too. Untreated, it can kill. The flies had been pushed way down past Panama for decades, but they breached the barrier back in 2022 and started marching north, and the first cases in South Texas showed up in June. So the USDA and Customs and Border Protection took at least 200 drones — the Skydio X10 quadcopters with thermal sensors — and started flying them over ranches at sunrise, looking for animals running a fever or wandering off alone, which is a sign of infestation. More than a thousand flights, twenty-two thousand animals surveyed.
And here's the piece that matters for you, the builder. The USDA says it's now researching how to use AI to analyze that drone imagery — detect wounds, identify individual animals, prioritize which ones the inspectors go look at. So you've got this collision of thermal drones, computer vision, and a nineteenth-century eradication trick called releasing sterile flies by the hundred million. What I want you to notice is the pattern: agencies with a fleet of expensive surveillance hardware sitting around will always find a second job for it. The border drones become livestock drones. The tool outlives its original mission. Keep that in your pocket, because we're going to see that same logic show up again with satellites in a few minutes.
Now, speaking of tools doing a second job, let's talk about a clever little piece of defensive engineering that made me smile. A pair of designers — Isaque Seneda and Gabriel Abrucio — put out a thing called ShieldFont. And the idea is beautifully sneaky.
You know how AI companies scrape the whole public web to feed their training machines. Publishers hate it, lawsuits are flying, everybody's looking for a way to opt out. ShieldFont's answer: a font that shows humans one thing and shows the scrapers something completely different. It uses ligatures — normally that's the typography trick where two letters get smushed into one prettier shape — but here they weaponized it to swap whole words when the page draws on your screen. So a human reads a perfectly normal sentence. But the raw HTML that a scraper pulls down? That says the cowboy was riding a potato.
And they were smart about it. They didn't just use synonyms, which a smart scraper could reverse. They didn't use gibberish, which a filter would catch. They swap words for the same part of speech but a totally wrong meaning. Horse becomes potato. The sentence still reads as grammatically fine — it just asserts nothing true. They built a dictionary of nearly twelve thousand words, refined it over three months. In testing on six scraper pipelines, over ninety percent of pages that would've been accepted got rejected by the quality filter. And the ones that slip through are poisoned — real English, correctly spelled, saying nothing real. As the authors put it: dropped means they didn't get your work, kept means they got something wrong.
Now here's where the builder in me perks up. This isn't a magic shield. Screen readers choke on it. Copy-paste breaks. Search engines get confused. And any scraper willing to render the full page and run OCR on a picture of it defeats the whole thing. But — and this is the whole point — that rendering costs five to thirteen times more than just yanking down the raw HTML. So it's not about making scraping impossible. It's about making it expensive at scale. That's the real lesson for anybody thinking about content defense: you're not building a wall, you're raising the toll. Whether that's a business you'd bet on, I don't know, but it's the right mental model.
Let me shift from defending your content to somebody's content getting scooped up whether they like it or not. Amazon told Twitch streamers this week that their content will now train Amazon's generative AI models — and here's the part that lit the community on fire — it's opt-in by default. You have to go dig through your channel settings to opt out. And Twitch's own Chief Product Officer, Mike Minton, said the quiet part into a live mic. Somebody in chat asked why it's not opt-in, and he said, and I'm quoting, "If this was opt-in, nobody would opt in. That's honestly the answer."
Give the man credit for candor, at least. That's the whole game right there. These platforms know their creators don't want this, so they bury it in a default nobody flips. And when a streamer asked whether their videos had already been fed to Amazon, Minton said he genuinely didn't know. Think about that. The Chief Product Officer of the platform cannot tell you whether your years of livestreams are already inside a model.
Now, for you as a founder, there are two lessons stacked here. One is about the data itself: thousands of hours of human voices and faces talking naturally is gold for training, and the platforms sitting on it are going to grab it. The second is about trust, and it's the more important one. Every time a company does the opt-out-by-default two-step, they burn a little more goodwill, and that goodwill doesn't come back cheap. If you're building anything that touches user-generated content, watch this backlash closely, because the regulatory and reputational bill for this behavior is going to come due.
Alright, let me connect that to a bigger conversation, because there was a panel last week at the Ai4 conference in Las Vegas that gets at the philosophical fight underneath all of this. Three heavyweights — Geoffrey Hinton, Fei-Fei Li, and Andrew Ng — got up and made the case for keeping AI open. And what I liked is they didn't agree with each other, which made it an actual conversation instead of a press release.
Ng's worry was gatekeepers. He said, quote, "I don't want there to be gatekeepers. That limits how all of us can access AI." His fear is a world where a handful of best-capitalized firms control the platform the way Apple and Google control mobile, and everybody else builds on their terms.
Hinton drew a line I think is worth holding onto. He said open source software is great — you show people the code, and somebody spots the bug. But open weights, he said, is different. You train a giant expensive model and just hand out the parameters, and now anybody can cheaply fine-tune it to do bad things. He was against it. But — and this is the honest part — he admitted the battle's over. His words: "I think that battle's been lost. We now have open-weight models... It's too late."
And Fei-Fei Li did the thing I most respect, which is refuse the false choice. She said, quote, "It's very dangerous to make this a dichotomy between complete openness all the way to complete closedness." She used nuclear physics — the papers get published openly, the uranium gets regulated, the lab work sits somewhere in between. Different layers, different levels of openness. That's the grown-up answer, and it's the one that doesn't fit in a headline, which is exactly why it usually loses.
I'll note the continuity here for you regulars — this open-weights debate has been circling our show for a couple weeks now, from Meta's Muse Glimmer release to Yann LeCun cheering Meta turning back toward open source. What's fresh today isn't the topic, it's these three specific voices staking out three genuinely different positions. And the reason I care as a builder is that the outcome of this fight decides whether your startup rents its intelligence from three landlords or picks it up off the open shelf for the cost of the hardware. That's not abstract. That's your margin.
Now let's get into that deep dive I promised, because it's the most consequential thing on today's menu for anybody building a company.
Thrive Holdings raised two billion dollars this week at a twelve-billion-dollar valuation. SoftBank, D1 Capital, Altimeter. And on the surface that's just another big AI check. But what Thrive actually is — that's the story. Thrive Holdings is essentially a private equity firm for the AI era. But instead of buying companies, loading them with debt, and flipping them, it buys boring traditional businesses — accounting firms, IT shops — and rebuilds their workflows around AI.
And the key to the whole thing is its relationship with OpenAI. Now, I want to be precise, because the ownership piece here goes back to December of 2025, so that's not new news — that's the background. Thrive Holdings is a spinout of Thrive Capital, which is one of OpenAI's major investors. And back in December, OpenAI took an ownership stake in Thrive Holdings, and part of that deal had OpenAI literally sending its own employees to go work inside Thrive's companies and accelerate the AI rollout. Sit with that for a second. The model lab isn't just selling you API access. It's embedding its people inside the acquired accounting firm to make sure the tech actually lands.
And here's why investors are throwing money at it — the numbers Thrive is putting up. They've got more than seventy businesses on their platforms now. Two pillars so far. There's Current, the accounting arm — more than fifty firms, over two thousand professionals. And they built something called TaxAI, self-improving tax agents that they say processed more than seven thousand tax returns at ninety-eight percent accuracy and cut tax-prep times by over thirty percent. The other pillar is Shield, the IT arm, around twenty companies, and they claim their AI sped up help-desk resolution times by thirty-six times. Thirty-six x. And they doubled the number of custom agents deployed in a single month.
Now, and I always say this, those are the company's own numbers, so salt to taste. But even if you cut them in half, you see why the check got written. And here's the part of the raise that tells you where this is going. They're launching a third platform, and it's not accounting or IT — it's regulatory services for the built environment. A founding member named Anuj Mehndiratta laid it out, and I'll quote him, because it's the thesis: "The U.S. needs to build and modernize more critical infrastructure, but projects are often constrained by local, technical, and regulatory complexity. This applies across data centers, manufacturing, healthcare, power, water, transportation."
You hear that? Data centers. Power. Water. Permits. The exact bottleneck that's strangling the whole AI infrastructure boom — the paperwork, the inspections, the compliance tracking, the permit prep. Thrive wants to buy the firms that do that grunt work and put agents on it. Another founder, Kareem Zaki, said the goal is to compress the regulatory bottlenecks, keep safety standards high, but do it cheaper and faster.
Here's why I'm spending real time on this, and why it's the deep dive instead of some flashier model release. For years the AI startup playbook was: build a shiny tool, sell it as software, hope enterprises adopt it. And what we keep learning — including from OpenAI's own research this week on how enterprises actually put AI to work — is that adoption is the hard part. The tool isn't the bottleneck. Getting a two-thousand-person accounting workforce to actually change how they do the work — that's the bottleneck.
Thrive's answer is radical and honestly kind of brilliant: don't sell the software to the business, buy the business. Own the workflow, own the P&L, own the humans, and then you control whether the AI actually gets used. And you're not competing on being the best model — you're competing on owning the customer so completely that adoption isn't optional. OpenAI and Anthropic both see this, by the way. OpenAI's got a thing called The Deployment Company, Anthropic's got one called Ode — both billion-dollar ventures putting elite engineers inside enterprises to force the integration.
So if you're a founder listening to this, here's what I want you to take away. The money right now is betting that the value isn't in the intelligence — the intelligence is getting commoditized, we just talked about that with open weights. The value is in the last mile. It's in owning the messy, regulated, human-heavy business where the AI has to actually work, and where somebody's got to sign their name at the bottom of the tax return. That's a moat you build with acquisitions and embedded engineers, not with a clever prompt. Whether roll-up-and-automate turns out to be a great business or a debt-laden mess in three years, I genuinely don't know. Private equity roll-ups have a spotty history. But the strategic bet is clear, and it's worth understanding whether you love it or hate it.
Let me come up for air and hit a few faster ones, because there's good stuff on the plate.
Remember I told you to hold onto that idea about tools outliving their mission? Here's the satellite version. Northrop Grumman's got what's basically a robot space mechanic. This surfaced fresh but the events go back to July, so, background not breaking. A spacecraft called a Mission Extension Vehicle had been bolted to the back of an Australian communications satellite for over a year, keeping it in position. This week it unplugged to make room for its replacement — a new fleet with two robotic arms, developed with DARPA, that'll go around servicing satellites, attaching little propulsion pods to keep old birds alive.
The economics here are the story. Satellites usually don't die because the computers fail — they die because they run out of fuel to hold their position. That Australian satellite launched in 2009, was built for fifteen years, and now might fly six more and keep printing revenue. Cheaper launches and cheaper parts suddenly make repair-in-orbit pencil out. But here's the wrinkle I can't skip: the Space Force has previously called a Chinese satellite with robotic arms a weapon, because arms that can grab and fix a satellite can also grab and disable one. Same hardware, and now we're back to that theme — the tool always has a second job, and sometimes the second job points a gun.
Quick one on the money-flying-around front. Cognition — the folks who make the Devin coding agent — are reportedly already talking to investors about a raise that could hit a forty-billion-dollar valuation. And I say "already" because they just raised at twenty-six billion back in May. Three months. What's interesting isn't the number, it's what founder Scott Wu said about how Devin's actually used — not as a human replacement, but for the grunt work programmers hate: dragging old software up to date, moving apps from one platform to another. That's the same theme as Thrive. The value's in the boring long tail, not the flashy demo. Customers include Mercedes-Benz, NASA, Goldman.
And on the healthcare side, Google put out research on AMIE, their medical AI system, now doing real-time video consultations. Built on Gemini, it interprets what it sees and hears — a cough, a wince — guides a virtual physical exam, and reasons through a diagnosis live. In a randomized study with patient actors and real primary-care physicians, the evaluators rated it favorably on thoroughness, diagnostic accuracy, communication. And the patient actors actually preferred the video experience over text chat. Now, Google's careful to say it's still a research system, needs a lot more work before anybody deploys it on real sick people. Fair. But the direction's clear, and if you're building in health, the bar for what "AI in the loop" looks like just moved.
Let me get out of AI for a minute, because there's world happening.
Over in the Black Sea, Romanian army divers blew up two Russian-made drones found floating near the country's biggest natural gas project, a site called Neptun Deep. This is a current story. The drones were the Gerbera type — Russia uses them for one-way strikes and as decoys. Romania checked with Ukraine, who confirmed they weren't theirs. Now, this matters beyond the headline because Neptun Deep is set to make Romania the biggest gas producer in the European Union, supplying Germany and Moldova — exactly the alternative energy Europe's been scrambling to build to get off Russian supply. So whether these drones drifted there by accident or on purpose, you've got Russian hardware showing up next to the infrastructure that's supposed to replace Russian gas. And it comes a week after an explosive drone tried to hit a Ukrainian cargo plane at an airport in Germany. The drone war is leaking across borders, and it's landing on energy infrastructure. If you're thinking about geopolitical risk, that's the map.
And a security note that hits close to home for any of you running a company. Uber Freight — the logistics arm of Uber — is investigating a data breach claimed by a gang called Helix. And I want you to hear how this crew operates, because it's not some genius zero-day. Google, which tracks them, says they use voice phishing. They call the IT help desk, pretend to be an employee, and ask to reset a password. That's it. Crude, rudimentary, and according to Google, these guys have pulled in at least ten-point-six million dollars in ransom between January and May. The lesson for every founder: your fanciest security stack doesn't matter if somebody can sweet-talk your help desk into a password reset. The soft target is always the human on the phone. Go tighten that today.
Couple more before we wrap. On the policy front — and I'll frame this as background since the nomination goes back a few days — the FCC is headed for a three-to-one Republican majority. Trump nominated a third Republican but is not filling the empty Democratic seat, breaking a long tradition where the president pairs a nominee from the other party. There's actually a law against one party holding an FCC supermajority, which is the technical reason he hasn't just fired the lone Democrat, Anna Gomez. Why should you care? Because this FCC, under Chairman Brendan Carr, has been leaning on broadcasters over coverage the administration doesn't like. A more lopsided commission means fewer checks on that, and if you're in media, telecom, or anything the FCC touches, the referee just got more partisan.
And one for the science-fiction lovers, because I can't resist. Physicists in China, at the Beijing Spectrometer III, say they've found the strongest evidence yet for something called a glueball. Now stay with me. Everything around you is made of quarks glued together by particles called gluons. A glueball is the wild idea of a particle made of just the glue — no quarks at all, held together by pure binding force. The Standard Model says they should exist, and they've been hunting for decades. This particle, catchily named X-2370, matched all three predicted properties — a physicist not even involved called it "an experimental triumph" and "the strongest evidence yet." Still needs independent confirmation, and there's only one machine on Earth built to look for these things, so that could take a while. No business angle here, kiddos. Sometimes a thing is just beautiful, and after a day of drones and data breaches, I figure you earned a particle made of pure glue.
So let me tie the bow on it. What connects most of today isn't a single headline, it's a stance. Thrive buying the accounting firm instead of selling it software. Northrop keeping old satellites alive instead of launching new ones. Cognition's Devin grinding through the migrations nobody wants. Over and over, the money's betting that the win isn't the shiny new thing — it's owning the boring, hard, human middle where the technology actually has to work. The intelligence is getting cheap. The last mile is not. If you're building, that's the ground worth standing on.
That's the menu for today. Take the Thrive story and really sit with it — I think it's a preview of the next two years. Watch that Twitch backlash, because the opt-out-by-default trick is going to get expensive for somebody. And go call your own IT help desk and ask them how hard it'd be to talk them into a password reset. You might not like the answer.
I'm Tony DeLuca, this has been Barely Possible, and I'll be right here tomorrow with another plate. Be good to each other out there.