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, this is Barely Possible, and we've got a fresh tray of stories warming under the heat lamp this morning. Buckle up, grab the coffee, and let's pick through what actually matters versus what's just noise.
I want to start with a story that, on its face, sounds like a quirky San Francisco crime blotter item, but if you sit with it for thirty seconds, it opens up into something every founder building autonomous systems should be thinking about. TechCrunch, Sean O'Kane, reporting that a burglar in San Francisco used a Waymo robotaxi to steal a haul of yoga clothes — and got away with it. Now the comedy of it writes itself, somebody calling a self-driving car to be the getaway driver for a lululemon heist. But the part that I want you to chew on is the second half of that headline. The incident sheds new light on how Waymo treats and stores the footage its robotaxis capture.
Think about what a Waymo actually is. It's a rolling sensor platform. Cameras inside, cameras outside, lidar, the whole nine. Every trip it takes, it's recording a high-resolution view of a public street and the person sitting in the back seat. So when a crime gets committed using one of these things, you suddenly get a real-world test of a question that has mostly lived in policy papers: who can get that footage, how long is it kept, and what's the process for prying it loose? In this case, the burglar got away with it, which tells you the answer was not as simple as the cops walking up and asking for the tape.
And here's why this lands on a builder's desk and not just a privacy lawyer's. If you are building anything that lives in the physical world — robots, delivery drones, smart cameras, in-car anything — you are accumulating a data asset that you did not necessarily set out to build. You set out to make a car drive itself. You ended up running one of the largest passive surveillance fleets in any given city. That's a liability surface, a subpoena magnet, and a trust problem all at once. The lesson is not don't collect the data. The lesson is decide on purpose what you keep, how long you keep it, and who gets to ask for it, before a yoga pants heist forces you to figure it out in public. Waymo is finding out the answer matters. Your sensor-heavy startup will too, just with less press.
Now let me hold that thought about data you accidentally accumulate, because it rhymes with the next one. There's a piece from Ars Technica, Ashley Belanger, with the kind of headline that makes you do a double-take: my Social Security number was exposed in a breach at Columbia, a school I have no connection with. Columbia University, that breach from last year — and the new wrinkle is that Columbia has now admitted the exposed victims went beyond its own students and staff. People with no relationship to the school at all are getting breach notifications.
How does that happen? It happens because institutions hoover up data about people they interact with tangentially — applicants who never enrolled, people in adjacent systems, third parties whose information ended up in a database for some operational reason nobody fully remembers. And then one breach later, your nine most important digits are sitting in a dump, courtesy of an organization you've literally never done business with. This is the same theme as the Waymo story wearing a different coat. The data you collect about people who aren't even your customers is still your problem when it leaks, and it's increasingly somebody else's nightmare when it leaks. For founders, the takeaway is uncomfortable but simple: every record you retain about a human being is a future apology letter you may have to write. Minimize what you keep. The cheapest data to secure is the data you never stored.
And while we're on the privacy beat, the courts had something to say this week too. Ars Technica, Jon Brodkin reporting, AT&T and Verizon lost their Supreme Court case over fines for selling customer location data. Eight to one ruling. The carriers had argued the FCC violated their right to a jury trial when it imposed those fines. The court said no — the FCC did not violate that right, and the fines stand.
This goes back to the scandal where the big carriers were selling access to your real-time location data to middlemen, who then sold it down the chain to people who absolutely should not have had it — bounty hunters, that whole mess. The FCC fined them. The carriers fought the fine on a procedural argument about jury trials, which is a clever lawyer's way of trying to neuter the agency's ability to punish anybody for anything. And in an eight-to-one decision, they lost. For builders, the signal here isn't really about telecom. It's that the regulatory enforcement teeth around selling personal data just survived a serious challenge. If your business model has any component that involves monetizing user location or other sensitive personal data through a chain of third parties, the legal weather just got a little colder. The agencies can still fine you, and the Supreme Court just declined to take away the mechanism.
Let me pivot from who's watching you to who's protecting your password vault, because the security desk had a busy day. Ars Technica again, Dan Goodin this time, with a writeup of how Dashlane explained the way attackers managed to download encrypted password vaults. Now I want to be careful and precise here, because the encryption part matters. The attackers got their hands on encrypted vaults. The explanation Dashlane gave is essentially a numbers game: by targeting a very large number of users, attackers increased their odds that some fraction of them had weak enough protection — reused master passwords, credentials already floating around from other breaches — to make the encrypted blob crackable or otherwise accessible.
The builder lesson buried in there is about threat models at scale. A defense that's strong for any individual user can still fail across a population, because attackers don't need to beat everybody. They need to beat the weakest few percent, and if you have millions of users, a few percent is a lot of people. If you run a service where the security depends on users doing the right thing — picking a strong unique master password, not reusing credentials — your real-world security is only as good as your worst-behaving cohort. That's an argument for designing systems that don't lean so hard on user discipline. Push people toward passkeys, enforce strength, detect credential reuse. Because the attacker is playing the law of large numbers against you, and the law of large numbers does not get tired.
Now let's shift gears entirely, from defense to the physical buildout that's powering all of this, because there were two data center stories today that, side by side, tell you exactly where the AI infrastructure squeeze is right now.
First, TechCrunch, Tim De Chant: Meta is stealing a tactic from Tesla and building data centers in tents. Literally tents. Meta apparently found one way to slash its enormous data center bill by going with tent-style structures instead of fully built-out permanent facilities. Tesla famously did this with its so-called tent factory to crank out cars when it couldn't build a real production line fast enough. And now Meta's pulling the same move for compute.
Think about what that tells you. When one of the richest companies on earth, a company spending tens of billions on infrastructure, decides the answer is to throw up a tent because building the real thing takes too long and costs too much — that is the loudest possible signal that demand is outrunning the ability to physically build. It's the same token-shortage, capacity-crunch story we keep circling, just expressed in canvas and steel instead of GPUs. The bottleneck isn't ideas anymore. It's concrete, power, cooling, and time. And when the bottleneck is time, you get tents.
Which brings me to the second piece of that puzzle, also Ars Technica: how some data center operators are tackling their water use problems. This one's from Molly Taft at wired.com. The hyperscalers have been catching real heat over their impact on water — both how much they consume to cool these facilities and what they do to local water quality and availability. And the piece walks through how operators are starting to respond, because in a lot of the places these things get built, water is not abundant, and the local community notices when the new neighbor is drinking the aquifer.
Put the tent story and the water story together and you get the real shape of the AI buildout in 2026. It's not a software story anymore. It's a land, power, and water story. If you're a founder whose product depends on cheap, abundant inference, understand that the people building the capacity you rely on are wrestling with permits, drought, angry city councils, and the physics of cooling. Those constraints flow downhill to your unit economics eventually. The cost of a token has a water bill behind it, and that water bill is becoming a political fight.
And staying right in that energy lane, here's the optimistic counterweight. TechCrunch, Tim De Chant again: Helion, the Sam Altman-backed fusion startup, raised four hundred and sixty-five million dollars to build a power plant for Microsoft. Helion is racing to complete a fusion power plant for Microsoft by 2028, and this fresh round is meant to help them get there.
Now, I'm a skeptic by trade, and fusion has been twenty years away for about sixty years, so I hold these claims at arm's length. But the structure of this deal is what's interesting regardless of whether the physics lands on schedule. Microsoft signed a power purchase agreement with a fusion startup. The hyperscaler isn't just buying chips and renting land — it's underwriting brand new categories of energy generation because it has concluded that the grid as it exists will not feed the thing it's building. That's the through-line connecting the tents, the water, and the fusion check: the compute ambition has gotten so large that it's pulling energy R&D forward by sheer force of demand. Whether Helion hits 2028 or not, the fact that a company like Microsoft is willing to write the check tells you how desperate the power situation is becoming. For builders, this is the macro you're operating inside. The biggest players are now in the energy business whether they wanted to be or not.
Let me come back down to ground level and talk about something closer to your actual stack, because there were a couple of agent-and-platform stories worth your time.
TechCrunch, Sarah Perez: Apple approved Poke as the first AI agent on its Messages for Business platform. Poke is the startup that lets people use AI agents through plain text messages, and it just became the first AI agent cleared for Apple's Messages for Business channel. Now this is a small headline with a big implication. Apple's Messages for Business is a curated, gated environment. Getting an AI agent approved there is Apple, very cautiously, opening a door to letting autonomous software talk to customers inside its messaging product. The fact that it's a single startup, and that the headline emphasizes the word first, tells you Apple is letting exactly one foot through the door and watching closely.
If you're building agent products, the distribution story here matters more than the company. Messaging surfaces — iMessage, WhatsApp, the SMS layer — are turning into the place where agents meet normal humans, because normal humans already live in their texts. They don't want to learn your dashboard. They want to text a thing and have it book the appointment. Apple letting Poke in is a tiny crack in a very valuable wall, and the companies that figure out how to be useful and trustworthy inside that messaging frame are going to have an enormous advantage over the ones still trying to get people to download yet another app.
And speaking of the messaging-agent frame, there's an Apple-shaped event on the horizon. TechCrunch, Lauren Forristal, previewing what to expect at WWDC 2026 — and the headline item is Siri's highly anticipated revamp along with broader Apple Intelligence updates. Look, Siri has been the punchline of the assistant world for a decade. The anticipation around a real revamp is half hope, half schadenfreude. The thing I'm watching, and the thing you should watch if you build on Apple's platforms, is whether Apple finally ships an assistant that can actually take actions across apps rather than just answer trivia and set timers. Because if Apple delivers a genuinely agentic Siri that can reach into third-party apps, that reshapes what an iPhone app even needs to be. It's a preview, not a product, so I'll hold the real analysis until they actually show something. But mark the calendar.
Meta was busy on the product front too. TechCrunch, Aisha Malik: Meta rolled out a new AI creator assistant on Facebook. The pitch is that creators normally have to wade through charts and dashboards to understand how their content is doing, and now they can just ask the assistant plain questions — when should I post, what are people saying in my comments. It's a small, sensible feature, and it's worth noting only because it's part of a pattern. The natural-language layer over your own analytics is becoming table stakes. Every platform that gives creators or businesses a dashboard is racing to put a chat box on top of it. If your product has a dashboard, your users are going to start expecting to talk to it instead of reading it. That's not a moonshot. That's just the new baseline.
Now here's one with a sharper edge to it. TechCrunch, Sarah Perez again: Meta's own Oversight Board said the company's account bans lack due process and transparency. The board — which is Meta's quasi-independent appeals body — raised due process concerns over how accounts get banned, and it's pushing Meta to give people clear information about what they violated. And critically, the board called out Meta's use of AI in making those determinations.
That AI angle is the one I want to flag for builders. We are now at the point where the watchdog body of a major platform is explicitly saying: you're using automated systems to make consequential decisions about people's accounts, and you're not telling them how or why. If you've ever had your account nuked by a platform and gotten a one-line generic message with no appeal that goes anywhere a human reads — you know this feeling. The Oversight Board is essentially demanding that AI-driven enforcement come with an explanation and a real appeals path. For anyone building moderation, trust-and-safety, or any automated decision system that affects users' access — this is the direction the wind is blowing. Black-box enforcement is going to face more pressure, not less. Build the explanation and the appeal in from the start, because regulators and oversight bodies are starting to treat it as a requirement, not a nicety.
And let me connect that to a Musk item, because it's the same governance theme from the opposite end. Ars Technica, Ashley Belanger: Elon Musk is trying again to escape FTC audits of how X handles user data. Public commenters are warning the FTC that Musk can't be trusted to protect X user privacy, and X is once again trying to wriggle out from under the privacy audit obligations it inherited from the old Twitter consent decree. The pattern with the Oversight Board and with this is the same — the institutions that are supposed to keep platforms honest about how they treat user data and user accounts are being tested, and the platforms are pushing back hard. As a builder, the meta-lesson is that the compliance obligations attached to handling user data are sticky. They follow the company even through ownership changes, and trying to shed them is a fight that gets fought in public, with your trustworthiness as the subject.
Now let's get into our deeper look for today, and I want to use an older piece to make a point about right now. Because the freshest thing in today's pile isn't actually the newest — it's a question from nearly three years ago that reads completely differently in the present.
There's an Ask HN post that resurfaced today, originally from June of 2023 — so I want to be clear, this is an old post that bubbled back up, not something written this week. The author, posting under the handle mrtranscendence, titled it: Is the rate of progress in AI exponential? And I'm going to read you a chunk of it, because it is, in hindsight, one of the most beautifully wrong things I've read in a while, and being wrong is exactly what makes it useful.
He writes: "I've frequently seen the claim that progress in machine learning and artificial intelligence, particularly with respect to large language models, has been exponential over the recent past. My problem is that I just don't see it? I see progress, yes, but not exponentially accelerating progress." He goes on: "Between March of 2022 and March of 2023 OpenAI went from GPT-3.5 to GPT-4; a clear improvement, but not a transformationally massive one. GPT-4 remains the most capable model widely available and that shows no signs of changing — it's not clear at this point if GPT-5 is even a glimmer in Ilya Sutskever's eye."
And then the kicker, the line that I cannot stop thinking about: "Heck, OpenAI can't even keep up with the GPT-4 demand right now, let alone a more compute-intensive successor. Maybe something like GPT-5 comes out tomorrow and makes me look like a fool, I just don't see any indication that it will."
Now. From where we sit in June 2026, this is almost poignant. The whole landscape this guy said he couldn't see coming — it came. We've been talking about GPT-5.5, Opus 4.8, Gemini 3.5, agentic coding tools that do real knowledge work, a token shortage so severe that companies are capping employee usage and Meta is building data centers in tents. The thing he said showed no signs of changing changed completely.
But here's why I'm not bringing this up to dunk on the guy. I'm bringing it up because his reasoning was actually good, and that's the scary part. Read his four points again. He said the research he was seeing was all about: better ways to prompt models, smaller models that use less compute, better ways to run models on commodity hardware, and new software tools that wrap the models. And he asked, fairly, does that count as exponential progress, or is it just activity? That was a sharp, defensible read of the evidence in front of him in mid-2023.
The lesson for founders is not "see, the skeptics are always wrong." The lesson is much more uncomfortable than that. It's that the precursors to a massive shift look indistinguishable from noise right up until they don't. The prompting tricks, the smaller models, the wrappers, the tooling — he saw all of that as evidence the field had plateaued. In hindsight, those were the early scaffolding of the agentic era. The chain-of-thought prompting he waved off became reasoning models. The wrappers he dismissed became Codex and Claude Code. The smaller efficient models he sneered at became the whole token-efficiency battleground that every AI company is now fighting on.
So when you're sitting in your own present, looking at today's pile of incremental-seeming stuff — a creator assistant on Facebook, one AI agent approved on Apple's messaging platform, a routing layer that picks cheaper models — and you're tempted to say, eh, that's just activity, not real progress — remember mrtranscendence. The honest answer is you cannot reliably tell the difference from inside the moment. The boring plumbing improvements and the world-changing breakthroughs wear the same clothes on the day they ship.
And there's a second lesson, the one for the hype merchants on the other side. This guy was wrong about the timeline, but his underlying instinct — that you should demand evidence and not just vibe along with the crowd — was completely correct. He just had bad luck on which side of the inflection point he was standing. The discipline of saying "show me, I don't see it yet" is a good discipline. It will make you look foolish exactly once per paradigm shift, and it will save you from a hundred bad bets in between. The trick, which nobody has fully solved, is knowing when you're standing at the inflection versus when you're standing in the hype trough. I don't have a clean answer for you. Nobody does. But reading a smart person be confidently, reasonably wrong about the most important technology shift of the decade is a good humility vaccine. Keep this guy's post in your back pocket the next time you're sure you've seen all there is to see.
And I'll tie this back to a thread we've been pulling on all week. We've talked about the token shortage, about Uber capping coding agents at fifteen hundred bucks a month per employee, about Microsoft building its whole stack from chip to model to harness. Every one of those stories is downstream of the thing mrtranscendence couldn't see in 2023: that the wrappers and the tooling and the efficiency plays weren't a sign the field had stalled — they were the field figuring out how to industrialize. The plateau he diagnosed was actually a runway. That's the continuity here. The unglamorous middle layer is where the action moved, and it's still where the action is.
Let me bring it back to the present with a few quick hits before we close out.
On the geopolitics-of-chips beat, there's a report — and I'll flag the timing is a little soft on this one, so I'll just call it a recent report rather than today's news — out of Taiwan, where authorities arrested three people on suspicion of smuggling Nvidia chips and Supermicro servers to China. This is the export-control cat-and-mouse game made flesh. The rules say you can't ship the high-end accelerators to certain destinations, and where there's a price gap that big, there's always somebody willing to put hardware in a crate and lie about where it's going. For founders working anywhere near hardware supply chains, the reminder is that compute is now a controlled strategic good, like a weapons component, and the enforcement is getting personal — as in, people getting arrested. Plan your supply chain like the rules are real and getting stricter, because they are.
There's also a recent piece floating around — Ars Technica — about an Estonian government benchmark testing which language models are best at resisting Russian propaganda. The Estonians, who have very good reasons to care about this, built a benchmark to measure how dozens of models handle Russia's strategic narratives — whether they parrot them, push back, or get manipulated. I love this as a concept because it's a benchmark built by people with actual skin in the game rather than a lab grading its own homework. If you're building anything that ingests or summarizes information at scale, the question of whether your model launders someone's propaganda is not abstract. A small country on the front lines decided to measure it. That's the kind of adversarial evaluation more builders should be borrowing.
And one more from the policy desk that's worth keeping an eye on. There's a recent Ars Technica report on the cable lobby — the NCTA — warning the FCC of chaos if it doesn't relax the ban on foreign routers, and the reason they cite is fascinating: memory and substrate shortages. So the same component crunch that's behind the AI buildout — memory chips, substrates — is now rippling into mundane consumer networking gear. The router industry is saying, we can't source compliant parts fast enough because the AI gold rush is eating the supply. That's the supply chain story showing up in your living room. When the cable lobby is citing the same memory shortage that's driving GPU prices, you know the squeeze has gone systemic.
Last one, and it's a palate cleanser. NASA's MAVEN spacecraft, after eleven years orbiting Mars, has gone out — and Ars Technica's Stephen Clark captured the team describing it as the loss of a loved one. Eleven years studying the Martian atmosphere, and then a quiet ending. I bring it up not for any builder lesson, just because in a week full of breaches and lobbying and tents, it's worth thirty seconds to acknowledge that some people spent over a decade carefully running a machine around another planet, and they're grieving it like family. There's a kind of craft and patience in that which is the opposite of the move-fast world we usually live in. Tip of the cap to the MAVEN team.
That's the tray for today. The thread running through all of it, if you want one: the data you collect becomes your liability, the compute you crave becomes a fight over water and power and arrested smugglers, and the progress you can't see from inside the moment is usually the progress that matters most. Stay skeptical, but keep that 2023 Ask HN post in your pocket as a reminder that skepticism has a failure mode too.
This has been Barely Possible. I'm Tony DeLuca, I appreciate you spending part of your morning with me, and I'll be right back here tomorrow with another tray. Take care of yourselves out there.