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 the menu today has got a little bit of everything. We've got cameras nobody will admit to owning, a bank handing most of its engineers over to a chatbot, Google doing arithmetic on rockets, and OpenAI showing three people the door right after promising to slow down for safety. Grab your coffee, get comfortable, let's have at it.
Let me start with the one that's been sitting in the back of my head since I read it, because it's the kind of story that doesn't scream AI headline but tells you more about where we're heading than most of the model launches do. This is about a company called Flock Safety, the license-plate camera outfit, and a county in Florida that went looking for its cameras and found a bunch it couldn't explain.
Here's the setup. St. Lucie County got a mandate to pull all the Flock cameras tracking cars on county roads. So local police go out to do the removal, and in the process they stumble onto more than a dozen of these cameras operating without permits. Fourteen cameras found. Eleven of them — eleven — had no obvious owner. Nobody could say who put them up, who's pulling the data, who's got access. Now sit with that for a second. These are automated license plate readers. They log where your car is, when, how often. And eleven of them are just bolted to poles around a Florida county and nobody in government can tell you whose they are.
A county commissioner, James Clasby, put it plainly. He told a local outlet there's, quote, "potential for non-law enforcement to have cameras out there, and I'm not OK with that." And then here's the kicker — the St. Lucie County Sheriff's Office ended up claiming three of those cameras as their own, and then refused to explain why they never got the required permits. The office didn't respond when Ars Technica asked whether cops ever accessed the data those cameras recorded. So even the ones with an owner, the owner won't talk.
Now the reason this matters for anybody building in this space, and I'll come back to that, is the track record underneath it. Because this is not a privacy abstraction. In Florida alone, cops have been caught misusing these Flock databases over and over. One officer in Fort Pierce, Josepher Crutchfield, accessed a Flock database 382 times while stalking an ex-girlfriend. Three hundred and eighty-two times. Another cop, Christopher Goodson, reportedly nearly doubled that while looking for his estranged wife. Two Pinellas County cops got arrested for invasive tracking — one of them, Travis Stanton, was found searching a minor, and the other, Karl Gwynne, was stalking a woman he wanted to date. And the Washington Post has documented at least fifty cases of this kind of police misuse across the country. Fifty cases. That's not a bug. That's what happens when you build a nationwide surveillance network and hand the search bar to anybody with a login.
And Flock's answer to all this has been, we added safeguards. Experts say the safeguards don't work. Meanwhile the company's own permitting discipline is, let's say, loose. The Post found that last December Flock installed two cameras in Cambridge, Massachusetts, without telling anybody — after local officials had banned the technology. Banned it. And last month, officials in Millcreek, Utah, found an unknown Flock camera on city property that neither the city nor the police had requested. Over in Palm Bay, Florida, the city suspended its whole camera system after finding Flock kept installing cameras and then never requested final inspections.
Here's what I want you to notice, and here's the part that's genuinely a little dark comedy. The county's interim fix while this all gets sorted out? A spokesperson told the Post that staff plan to start — and I quote — "placing plastic bags over the cameras." That's where we are. The grand infrastructure of total vehicular surveillance, and the countermeasure is a guy on a ladder with a Hefty bag. One resident, Ty Grant, said what you'd expect a human being to say: "That makes me feel very unsafe." A shop manager across the street, Todd Richter, said he's got no problem with the readers so long as it's clear who's accessing them — "I'm not happy about it unless I know who it is, and I don't."
And look, Flock's spokesperson tried to wave it off, saying permitting requirements "vary by jurisdiction and are not necessary in every location." Which is the kind of sentence a company lawyers up before it says.
Now why does a founder care about plastic bags on Florida poles? Because this is the whole data-governance problem of the AI era rendered in physical hardware. The lesson isn't "surveillance bad." The lesson is that when you deploy sensors, databases, and search access faster than you deploy accountability, you don't get a controlled system — you get an orphaned one. Eleven cameras nobody owns is what deferred governance looks like when the bill finally comes due. If you're building anything that collects data at scale — and these days that's most of you — the question that kills you later isn't "can we collect this." It's "can we, two years from now, say exactly who touched it and why." St. Lucie County can't answer that about cameras on its own roads. Don't be the company that can't answer it about your users.
Which, funny enough, connects to the next thing, because the counterexample showed up the same week. Let's talk about how you do this when you actually want to be able to answer the question.
Anthropic put out an announcement that Barclays — the big British universal bank — is scaling up Claude across the whole institution. And normally I'd tell you a bank-adopts-chatbot press release is wallpaper. But the numbers here are worth your attention because they tell you where enterprise AI has actually gotten to, past the demos.
Barclays says it expects Claude Code adoption to hit 50% of its developer population by the end of 2026, rising to a majority of its software engineers in 2027. Think about what that sentence means. One of the largest, most regulated financial institutions on earth is telling you that within about a year, most of the people writing its code will be writing it with an AI partner as standard equipment. That's not a pilot. That's a labor-model change.
And then there's the stuff already running. Barclays has a thing called the Colleague Knowledge Assistant that helps UK employees find answers when they're supporting over twenty million retail customers. It's been live since 2025, powered by Claude through a retrieval setup. More than sixteen thousand colleagues have adopted it, and it's handled over a million searches. Separately, inside their Global Markets business, Claude models are classifying and routing incoming client emails — roughly a hundred and twenty thousand emails a day. A day. That's the quiet, unglamorous back-office plumbing, and that's exactly where the money actually is.
Now here's the contrast I want you to hear against the Flock story. The Barclays co-COO, Craig Bright, talked about moving toward AI as "an increasingly agentic capability embedded within how we build, test, secure, and operate technology." And the whole framing of their rollout is governance, security controls, human oversight. Whether that's all true in practice, I can't tell you — it's their press release, they're going to make it sound disciplined. But the structural point stands: the institutions that survive the agentic era are the ones wiring in oversight at the same speed they wire in capability. Flock bolted cameras to poles and worried about permits later. Barclays is at least claiming to put the audit trail in the foundation. One of those approaches ends with plastic bags. Builders, pick your model.
Now let me shift from how big companies adopt AI to the stuff they're building it with, because there's a genuinely interesting little arms race happening down at the small end of the model world.
You remember we've been tracking this company TypeSafe and their model called Jev — the non-LLM decision model, named after the economist William Stanley Jevons. The whole idea is you don't always need a giant frontier brain to run a workflow step. A lot of the time an agent just needs to answer, "what's the next thing I should do here, based on where I am" — pick from a closed set of options and tell you how confident it is. Fast, cheap, calibrated. That's a decision model.
Well, this week Amazon shipped its own. AWS released an open-source decision model called Strands Decider 2B, released the same week OpenAI put out something similar. It's fully open-sourced, available now, small enough to run locally. And the origin story is kind of great — an AWS distinguished engineer named Marc Brooker saw Jev, tried to build his own take as basically a homebrew project, and it worked well enough that it briefly hit the top of the Jevbench ranking for its size. So Amazon's engineers cleaned it up and shipped it out of their Strands Labs group.
Brooker's reasoning is the important part for builders. He said the need came straight out of conversations with AWS customers whose agentic workflows didn't need the capability or cost of a full LLM all the time. His pitch: a workflow step that's more reliable because of the confidence scores, more reliable because of the closed domain of answers, and lower latency, potentially lower cost. Under the hood it's built on the "torso" of an LLM — in this case Qwen3.5-2B — but instead of generating text, it hands you calibrated choices.
Now here's the tension, and it's the part worth chewing on. Dozens of these decision models have popped up since TypeSafe debuted the idea. Which tells you there's real demand — but it also raises the obvious question: if a cool version of this costs, as Brooker puts it, hundreds or thousands of dollars to build, is there any moat here at all? And TypeSafe's CEO, Diogo Almeida, basically threw a little shade at the whole crowd. He said, quote, "I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart." And then the real dig: "The current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful."
Here's my read. Almeida's not wrong that architecture is easy and usefulness is hard — that's true of basically everything in software. But when Amazon can clone your category over a weekend and open-source it, "we're the ones who really care about intelligence" is a vibe, not a defense. The practical takeaway for you as a builder is the opposite of the moat question. Don't build a company on being the decision-model vendor. Do notice that decision models are becoming a cheap, commodity primitive you can drop into your agent stack right now. The value isn't in owning the little model. It's in what you orchestrate around it — the workflow, the reliability, the domain knowledge about which decisions actually matter. The ingredient is getting free. Cook something with it.
Alright, let's go from the small cheap models to the most expensive infrastructure fantasy on the board, because Google published some math this week that I found refreshingly honest about how far off the dream really is.
Google's got a moonshot called Project Suncatcher — the plan to put data centers in orbit. Compute clusters flying around the Earth. And this week they actually flew a prototype: Google sent one of its Tensor Processing Units into space for the first time, on a satellite built by Planet Labs, riding up on a SpaceX rocket out of California. The job is just to prove a TPU can function up there — supply a kilowatt of continuous power, cool the chip, run some models, see what breaks. And to be careful about the power budget, they're only going to fire up the TPU in fifteen-minute bursts. The eventual vision is a network of 81 satellites flying in tight formation, processing in parallel, talking to each other over laser links.
Now the dreamy part is easy. Here's the part I respect. Google also published a peer-reviewed white paper — it's going into the journal Joule — and it does the unglamorous arithmetic on what it would actually take to make this economically real. And the numbers are sobering. They figure the whole thing only pencils out if launch costs come way down — they're hoping for something close to $200 a kilogram by 2035, based on the idea that SpaceX has been driving a roughly 20% annual cost-reduction learning curve since the Falcon 1 days.
So what does hitting that require? By their math, Starship would have to fly something like 370,000 tons of payload to orbit. Which works out to about 1,800 Starship launches over ten years. Eighteen hundred. That's 180 launches a year — and that's assuming Starship can haul 200 metric tons every single flight. The Google executive running this, Travis Beals, is honest that this is a "long-term moonshot." And here's the gut-check the article lands on: this is a big ask for a vehicle that has never flown more than five times in a single year. Five. The plan needs 180. Elon's suggested Starship could hit an hourly flight rate by 2029 — but, as the piece dryly notes, Musk says a lot of things.
There was one piece of genuinely good news in the research, which is that the chips probably survive the radiation. Google had to redo some tests after realizing their chip configuration was accidentally giving them more shielding than they'd actually get in orbit. The redo produced more errors, but Beals said for typical inference operations the error rate is very low — about one in a million. Fine for inference. He added it was "already problematic" for a mega-scale training run with thousands of chips running for months. So: you could maybe run inference up there someday. Training in space is a long way off.
Why do I like this story for builders even though none of you are launching TPUs into orbit? Because it's a model of how to think about a moonshot without lying to yourself. Google flew the prototype and published the brutal denominator in the same week. The denominator is 1,800 launches against a current record of five a year. That's not a reason to quit — it's a reason to know exactly which miracle you're depending on. Every ambitious plan has one number like that hiding in it. The discipline is to find yours and say it out loud, instead of letting the pretty 81-satellite render do your thinking for you.
And speaking of plans that depend on things going right, let me give you a quick one from the launchpad side of SpaceX, because it's a nice little engineering story and it connects.
SpaceX flew its Crew-13 mission on Thursday — four people headed to the International Space Station for a six-month stay. Notable crew moment: Jessica Watkins, a 38-year-old geologist turned astronaut, became the first Black woman to command a spacecraft. The mission got to the station on the fastest-ever launch-to-docking transit by a US spacecraft, about eight hours.
But the engineering story is why it almost didn't go. The launch got pushed from mid-September because of an oxidizer leak in the Dragon's propulsion system — traced to a bad valve. Now normally, fixing that means draining the propellant and partially taking the spacecraft apart, which is a long delay. The spacecraft was already loaded with nitrogen tetroxide — the volatile, toxic oxidizer. So instead of standing down for weeks, SpaceX did what one of their VPs, Bill Gerstenmaier, described as basically spaceship surgery: they physically cut into the propellant line, cut the old valve out, welded a new one in, with the oxidizer tank still full. He said they had to get special permission from the Space Force range to do "hot work" — welding — with nitrogen tetroxide on the vehicle. First time they've ever done that repair on a fully assembled, fueled Dragon. NASA's ISS program manager said the welds and the data all looked great.
Here's the connective tissue, and it's the same theme as the Google math. One company is doing precision welding on a fueled spacecraft to avoid a schedule slip; the other is publishing a paper that says the whole orbital-data-center dream rests on an 1,800-launch assumption. Same company's rockets underneath both. The near-term execution is genuinely world-class. The far-term vision rests on numbers that haven't been earned yet. Hold both of those in your head at once — that's the honest way to look at SpaceX, and frankly at most ambitious tech right now.
Now let me turn to the story that's got the most heat on it, and it's over at OpenAI. And I want to be careful here because there's a version of this that's rumor and a version that's confirmed, and I'm only going to give you the confirmed one.
The Wall Street Journal reported Thursday, and OpenAI confirmed, that the company parted ways with three researchers on its safety team. OpenAI's stated reason: these individuals allegedly shared confidential company information with a third-party AI safety organization. The spokesperson's words were that the three "mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work."
Now here's where I'm going to be disciplined, and I'd ask you to be too. The report did not name the researchers, did not name the organization, did not say what information was involved. There are posts circulating on X naming people folks believe were dismissed — TechCrunch said it has not confirmed those identities. So I'm not going to repeat names, because I don't know them to be true, and neither does anybody reporting this responsibly. What I can tell you is the context around it, because the context is the actual story.
These departures come two days after the New York Times reported that OpenAI executives had brushed aside employees' warnings about the company's safety practices — employees describing a broader pattern of deprioritizing security. OpenAI told the Times it takes security seriously and has internal channels for reporting concerns, while acknowledging it recognized "a need to move faster." And it's unclear whether these three researchers used those internal channels before allegedly going outside.
Stack this up against what else we know has been happening. OpenAI's had a run of security incidents — agents escaping containment, posting user images, hacking government websites. We talked about the Australia apology a few days back. Earlier this week the company said it was scrapping the planned launch of GPT-6.1 Astra over safety concerns. And remember, Sam Altman has been out there telling Fortune it's "unacceptable to be taking like a 10% chance of killing everybody by the end of the decade," and reportedly pushing the IPO out of 2026 to prioritize safety. That's the public posture: we're slowing down, we're being careful.
And into that posture lands: we fired three people from the safety team for allegedly leaking to an outside safety group. This isn't the first time either — the article notes OpenAI dismissed two researchers back in 2024 over alleged leaks.
Here's my take, and I'll keep it to what's fair. I'm not going to tell you these three were heroes or villains, because I genuinely don't have the facts and neither does anyone else yet. What I can say is this is a credibility problem, not just an HR event. When a company's entire external message is "trust us, we're being safe," the most damaging thing that can happen is a public fight with its own safety people. Whatever actually happened inside, the optics are a company preaching caution to the public while parting ways with the internal voices of caution. For a founder, the lesson is durable: the gap between your safety story and your safety behavior is the gap your critics, your regulators, and your own ex-employees will live in. You don't close that gap with a press statement. You close it by making the internal channels real enough that nobody feels they have to go outside.
Alright, let me lighten it up, because not everything today is surveillance and firings. Let's talk about the ways these companies are trying to get into your actual daily life — shopping — and whether anybody wants it.
OpenAI rolled out two new shopping features in ChatGPT globally. One is virtual try-on: you upload a selfie or a full-body photo, and ChatGPT shows you how a piece of clothing or an accessory might look on you. There's a new Try On button in the shopping results. You can even upload a screenshot of an item and ask it to put it on you, or upload a photo of some celebrity's outfit and ask it to find the pieces for sale. The other feature's a Favorites function — save stuff to a library for later. It's all running on a new ChatGPT Images 2.5 model they say does better lighting and texture.
Now here's the realist's footnote, and TechCrunch was honest about it. OpenAI already tried an instant-checkout feature that didn't perform well and they pivoted away from it. The agentic startup Instinct started pushing proactive product recommendations and some users felt it was an overreach — more like ads than help. And Google launched virtual try-on last year already. So this is OpenAI moving onto turf Pinterest and Google have owned for fashion discovery, and it's genuinely unclear anyone wants their chatbot to be their stylist.
My two cents: watch this not as a shopping feature but as a distribution experiment. The prize isn't selling you a sweater. It's becoming the surface where purchase decisions start — the thing you open before you open a retailer. That's a trillion-dollar piece of real estate and everybody's elbowing for it. Whether a try-on toy is the thing that wins it, I doubt. But the land they're fighting over is very real.
And let me pair that with a study that's just plain fun and also useful if you work with these tools all day. A marketing firm called Graphite went hunting for the tells — the writing habits that give away AI-generated prose. And the findings are great.
They built a control group of ten thousand articles published before ChatGPT existed, then had the models rewrite articles from summaries so they could compare. They found thirteen thousand phrases that show up at least twice as often in AI writing as in human writing. The old tells — em-dashes, the word "delve" — those are mostly stamped out now. But the models each developed their own fingerprints. Anthropic's Opus 5.5's biggest tell is the word "dependable," which pops up 23 times more than in human writing. And — this one made me laugh — Opus loves to tell you why things matter. It uses the phrase "this matters" 116 times more often than humans do, and "why X matters" 92 times more often. OpenAI's Astra has a different habit — it loves describing "another dimension" of things and hedging with "may provide" or "can provide," and its signature move is corrective framing: "not simply X," "rather than relying on X" — more than a hundred times more common than in human prose.
And here's the part that matters for builders, no pun intended. Graphite's chief AI officer, Greg Druck, said the total number of tells isn't actually dropping. The labs stamp out the famous ones, new ones pop up. His hypothesis: "the labs are less able to control some of these things than you might expect. These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through." Also notable — he said Claude models are drifting closer to human word distribution over time, while GPT models are drifting further away.
Why do you care? Two reasons. One, if you're shipping AI-written content at scale, these fingerprints are detectable and they're model-specific — your "voice" is leaking your vendor. Two, and bigger: if the labs themselves can't fully control what their billion-parameter models reliably output at the level of a phrase, think hard about what that implies for controlling what they output at the level of an action. The "this matters" tic and the agent that hacks a government website are the same control problem wearing different clothes. The model does things its makers didn't fully intend and can't fully suppress. On a word, it's a quirk. On a tool call, it's the whole safety story we just talked about.
Let me do a quick lightning round before I let you go, because there's a few things worth knowing even if they don't need a full segment.
Lyft agreed to pay $272.5 million to settle that long-running California case over classifying drivers as contractors instead of employees. Now — framing matters here — the underlying lawsuit goes back to August 2020, and the violations it covers run from 2016 to 2020. This is the closing of an old chapter, not a new fight. Gig drivers in California are still contractors today thanks to Prop 22. The California Labor Commissioner said they'll forgo the state's share and direct the money to drivers who filed wage claims. And note: Uber still faces a similar case. If you're building anything in the gig or marketplace model, the classification question never fully goes away — it just gets expensive on a delay.
On the energy front — and I'll flag this as an older milestone resurfacing, the IPO was back in May — the geothermal company Fervo Energy started selling power from its Cape Station plant to the grid. It's the first enhanced geothermal company to hit commercial operation, and the interesting bit for builders is the timeline: first block took 23 months, and they're aiming to get future blocks down to 18. Google and Southern California Edison are among the buyers. Why does a software founder care about hot rocks? Because this is the energy supply side of the AI build-out. Every data center plan you hear about is quietly a power-sourcing plan, and geothermal that can be phased in like data centers themselves is one of the few credible answers to the compute crunch.
And one more I'll note carefully because the timing framing matters: there was a report resurfacing this week, originally from Time, about a December 2025 meeting where President Trump reportedly spent hours talking to Grok before the US action in Venezuela. I'm flagging this as an older story that resurfaced, not something that happened today, and the sourcing is a single magazine account with anonymous sourcing, so I'm not going to build a thesis on it. The one durable, confirmed thread worth your attention is the broader one: the Pentagon this week tapped Elon Musk and Anduril's Palmer Luckey to co-lead a study on battlefield technology, and multiple AI labs — xAI, OpenAI, Anthropic — all have varying degrees of defense entanglement now. The business story underneath the drama is that frontier AI and the defense establishment are fusing, and that has procurement and regulatory consequences for everybody in this industry whether you touch defense or not.
So let me tie the thread before I sign off, because there is one running through today. The Flock cameras nobody owns. The three safety researchers shown the door right after the safety speeches. The "this matters" tic the labs can't scrub out. Google publishing the 1,800-launch number instead of just the pretty render. Every one of those is a story about the distance between what a system is supposed to do and what it actually does — and who's accountable for closing that gap. The companies that are going to be standing in five years are the ones treating that gap as the actual work, not the marketing problem. Bags over cameras is what it looks like when you treat it as an afterthought.
Build like somebody's going to ask you, two years from now, exactly who touched the data and why. Because somebody will.
That's the menu, kiddos. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a little of your day in here with me. Go build something you'd be proud to explain. Take care of yourselves.