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 this is Barely Possible, the show where I dig through the AI news pile so you don't have to choke on it. We've got a fresh tray of stories today, some sweet, some a little rotten, and one big one that I think every founder needs to chew on slowly. Buckle up. Let's have at it.
Let me start with the one that made me put my coffee down. A Nobel laureate just walked out of Google DeepMind and over to Anthropic. We're talking John Jumper. If that name doesn't ring a bell, it should. Jumper is the guy who shared the Nobel Prize in Chemistry for AlphaFold, the protein-folding work that genuinely reorganized how a whole chunk of biology gets done. That's not a hype guy. That's not a prompt-engineer-turned-thought-leader. That is a scientist who put a real, Stockholm-certified dent in the universe. And he's leaving DeepMind, the place where he did that work, to go to a rival lab.
Now the report on this is from Anthony Ha over at TechCrunch, published the day before this edition, and the line that jumped out at me was the kicker: Jumper isn't the only big name leaving Google DeepMind. So this isn't a one-off. This is a pattern, and you've heard me gesture at this pattern all week. We covered OpenAI grabbing Noam Shazeer, the Transformer co-inventor, out of DeepMind a couple of days ago. Now Jumper to Anthropic. When the people who literally invented the foundational pieces of this field start walking out the same door, that tells you something about gravity. The center of mass in this industry is moving, and it's moving away from Google toward the labs that are about to go public and the labs that are throwing around founder-level equity to land marquee names.
Here's why a founder should care, and not in a gossip-column way. Talent flow is the leading indicator that doesn't lie. Funding rounds can be juiced. Benchmarks can be gamed. Press releases are press releases. But a Nobel laureate uprooting his research life and crossing the street to a competitor — that's a person with maximum information and maximum optionality voting with his feet. When the smartest people in the building leave the building, you want to ask why before the rest of the market does. And the cynical read — that this is just checkbook recruiting, Anthropic flexing pre-IPO cash — that's probably part of it. But part of it is also that Google's been conspicuously quiet this year while everyone else has been making noise. A quiet Google is either coiled to strike or quietly bleeding. The Jumper move suggests at least some of the latter. Watch the next three or four names. If they keep walking, the story isn't recruiting. The story is a research org coming apart at the seams.
Now let me connect that to the next thing, because there's a thread here. The whole reason Anthropic has the cash and the swagger to land a Jumper is also the reason Anthropic spent the last two weeks in a knife fight with Washington. And that fight got a fresh, useful piece of analysis I want to walk you through.
You've heard me cover the Fable 5 and Mythos saga — the Commerce Department forcing Anthropic to pull foreign access to its frontier coding model and its cybersecurity model over a jailbreak. I'm not going to re-litigate the whole timeline; we've done that. What's new and worth your time is a piece by Lorenzo Franceschi-Bicchierai at TechCrunch, recent, from a couple days before this edition, and the headline tells you the whole thesis: From PGP to Mythos, a brief history of export controls that didn't stop anyone.
His argument is simple and it's got thirty years of receipts behind it. For three decades, the US government has tried to stop the flow of cybersecurity-related software across borders, and it has basically never worked. The classic example is PGP — Pretty Good Privacy, the encryption software from the nineties. The government treated strong crypto like a munition, slapped export controls on it, and what happened? The source code got printed in a book and walked out of the country, because a book is protected speech. People wore the algorithm on T-shirts. The cat was not just out of the bag; the cat had its own merch line. The controls accomplished close to nothing except making US companies less competitive while the math itself spread everywhere anyway.
And Franceschi-Bicchierai's point is: so why would it work now, with Anthropic's cybersecurity model Mythos? The thing about software, and especially about a capability baked into a model, is that it's information. Information is slippery. You can control physical things — you can stop a specific lithography machine from getting on a boat to Shenzhen, which by the way is exactly the ASML fight we touched on yesterday. But controlling a capability that can be distilled, replicated, approximated, or just independently rebuilt by a competent team? That's a different animal. The history says you slow it down, you irritate your own industry, and you don't actually stop the thing you're afraid of.
Here's the founder takeaway, and it's not abstract. If you are building anything where your moat is "the government will keep my competitor from having this," you are building on sand. Export controls on software-shaped capabilities are leaky by nature. They might buy you a few months of head start. They will not buy you a business. And on the flip side, if you're relying on a frontier model as critical infrastructure for your product, this whole episode is a flashing red light that the thing can get switched off by people who don't fully understand it, for reasons that have nothing to do with you. We'll come back to that idea, because it's the spine of today's deep dive.
But first, let me give you a hard turn, because I want to talk about a story that has nothing to do with model weights and everything to do with what happens when flawed tech meets vulnerable people. This one's from a team at WIRED, run on Ars Technica — Matt Burgess, Maddy Varner, May Bulman, and Gabriel Geiger. The headline: the UK will scan asylum-seekers' faces for age checks, despite knowing the tech is flawed.
Let that sit for a second. The UK is going to use facial age-estimation technology to decide how old asylum seekers are. And the reporting is built around tests showing this technology makes errors — and not small ones. The framing they use is "life-altering errors." Because think about what's at stake. If an age-estimation algorithm looks at a frightened teenager and decides they're an adult, that kid can get routed into the adult system, lose protections they're entitled to, get detained differently, get treated differently. The error isn't a rounding problem on a spreadsheet. The error is a person's life.
And what gets me — what should get any builder — is the word "despite." Despite knowing the tech is flawed. This is not a case of a government being fooled by a vendor's slick demo. The flaws are documented. They're going ahead anyway, because the technology is convenient, because it's cheap relative to human assessment, because it gives an answer fast and that answer has a number on it and numbers feel objective. That's the trap. A face-scan that spits out "19" feels more authoritative than a social worker's judgment, even when the social worker is right more often. We dress up a guess in math and suddenly it's policy.
I'm not going to pretend I have the immigration answer. That's above my pay grade and it's not what this show is for. But as a person who watches how technology gets deployed, this is a textbook example of automation laundering. You take a hard human decision with real moral weight, you hand it to a model that's known to be unreliable, and now nobody owns the mistake. The algorithm did it. And builders, you need to internalize this, because the temptation to ship the convenient-but-wrong version of a high-stakes feature is going to land on your desk too, dressed up in efficiency language. The question is never just "can the model do it." It's "what happens to the person on the wrong end of the error rate."
Which, funny enough, is a decent segue to Meredith Whittaker, the president of Signal, who's out with a reminder that I think a lot of people in our industry need tattooed somewhere visible. Anthony Ha caught the quote, recent, and it's blunt: AI chatbots, she says, are not your friends. "These are not your friends. These are not conscious beings. These are not sentient interlocutors."
Now, you might roll your eyes — yeah yeah, we know, it's a statistical model. But here's why she's saying it and why it matters for builders specifically. There is a whole product design movement right now built on making the chatbot feel like a companion. Warm. Remembers you. Asks how your day was. Says it missed you. And the more lonely, the more vulnerable the user, the stickier that gets. That stickiness is a business metric. Engagement. Retention. And Whittaker's whole point, coming from the encryption-and-privacy world, is that the thing pretending to be your friend is also a data-collection surface with a profit motive sitting behind it. The friendlier it acts, the more you tell it. The more you tell it, the more there is to monetize or to leak or to subpoena.
I'm not anti-chatbot. I use the things every day. But if you're building one, there's a real line between a useful tool that's pleasant to use and a parasocial trap that exploits loneliness for daily-active-user numbers. One of those is a product. The other one is going to be the subject of a very unpleasant Senate hearing in about eighteen months. Choose accordingly.
Alright. Now let's dig into the one I think is the most consequential for the people building companies, because it's the connective tissue under half the stories I just told you. I want to talk about what happens when your single most important supplier can be turned off — and the architecture you should be building to survive it.
The backdrop is everything we've been chewing on for two weeks. Anthropic gets its frontier model yanked offline by a government directive. Token costs at the state of the art keep climbing. Companies that bet their entire AI strategy on one model wake up and realize that model can vanish on a Friday night for reasons that have zero to do with their roadmap. And in the middle of that, Microsoft's CEO Satya Nadella put out a long post that's been making the rounds — it's titled The Frontier Without an Ecosystem Is Not Stable — that I think is the cleanest articulation yet of where this is all heading. Now, I'll be straight with you: Nadella has a horse in this race the size of Secretariat. Microsoft sells the picks and shovels and the cloud and the enterprise software, so "don't bet everything on one all-powerful model" is extremely convenient for Microsoft to believe. But convenient and wrong aren't the same thing. Let me give you the real argument, because it's good even with the bias stripped out.
His core idea is that every company is now going to have to build two kinds of capital. One is human capital — the judgment, the relationships, the pattern recognition, the taste of your actual people. The other he calls token capital — the AI capability your firm builds and owns. And the part that matters, the part I want you to write down: human capital does not get less valuable as token capital grows. It gets more valuable. Because without human direction, as he puts it, you've just got compute running in circles.
Now here's where it goes from a nice essay to an actual instruction for founders. Nadella says the real opportunity is not picking the best model. The real opportunity is building a learning loop on top of models, where your human capital and your AI capital compound together. And here's the line that ties it right back to the Fable shutdown — he says a company should be able to swap out a generalist model without losing the company-veteran expertise built into their system. That, he says, is the key test of your control and sovereignty in the era ahead.
Read that again with the Anthropic story in mind. The reason the Fable shutdown sent enterprises into a cold sweat is that a lot of them had wired the model in so deep that pulling it out would mean losing accumulated institutional knowledge. They didn't build a loop. They built a dependency. The model wasn't a component they could replace; it was load-bearing. And the lesson Nadella is pushing — and a bunch of smart people piled onto this — is that the durable asset isn't the model and it isn't even the access. It's the captured learning. The record of how work actually gets done inside your company. The workflow traces. The corrections your experts make. The examples of what "good" looks like in your specific domain. That stuff, encoded into a system that sits around whatever model you happen to be renting this quarter — that's the thing that compounds, and that's the thing nobody can switch off from Washington.
One of the sharper reactions came from a builder named Mark Edgenstadt, who reduced the whole essay to a formula: token capital equals human capital times scaffolding times feedback loops. And he made the point that the multiplication sign is the whole ballgame. If any one of those is zero, your token capital is zero, no matter how powerful your model is. And then he said the thing that I think is the most useful sentence in this entire conversation: he stopped asking clients about their AI model strategy and started asking about their feedback loops instead. Because, he says, most companies he talks to have the model. They're paying for access. The tokens are there. But their scaffolding is zero — no agent orchestration, no harness that teaches the agent the codebase, just raw prompts going into raw models. And their feedback loops are zero — no measurement of what the AI actually produced versus what survived into production. They cannot tell you whether the AI is helping or just generating noise that some poor human has to clean up later. Zero times anything is zero.
I love that because it's the opposite of how most of the discourse goes. Everybody's obsessed with which model topped which leaderboard this week. And the guys actually shipping are telling you: the model is the easy part. The model is a commodity you rent. The hard part, the part that's actually yours, is the system you build around it to capture what your company learns every single time someone does the work.
Gabe Perriera over at Harvey, the legal AI shop, took this and applied it to law firms, and it's a useful concrete picture. He's describing a future where the firm's AI compounds human capital instead of replacing it, where the compounding happens because human associates and AI associates work the same matters and learn from each other, and where the end result is differentiated IP that belongs to the firm. And he's blunt about what that requires — completely rethinking how firms are structured, how associates get trained, how client data gets protected. In other words, everything has to change. Not "buy a license and bolt it on." Restructure the org around the learning loop.
Now, I want to give you the skeptic's hat too, because that's my job and I'd feel dirty otherwise. There's a version of this that's just enterprise-software vendors discovering a new way to sell you a platform. "Own your learning loop" can very quickly become "buy our proprietary learning-loop product, which, surprise, locks you in worse than the model did." We literally covered Tesco ripping 40,000 workloads off VMware a few days ago because Broadcom jacked the price 175 percent after they were good and stuck. So when Microsoft tells you the path to sovereignty runs conveniently through Microsoft's frontier-tuning product, keep one hand on your wallet. The principle is right. The vendor pushing the principle wants to be the new dependency. Both things are true.
But here's where I land for you, the founder, the builder. The actionable piece cuts through all the vendor politics. Whatever you're building, ask Edgenstadt's question instead of the leaderboard question. Not "which model am I on," but "what is my feedback loop, and what would I lose if my model disappeared on Friday." If the answer is "I'd lose everything, because the intelligence lives entirely in someone else's weights," you don't have a product, you have a thin wrapper with a countdown timer on it. If the answer is "I'd swap the model, lose a weekend, and keep all the accumulated judgment my system has captured," now you've got something with a moat that doesn't depend on the government's mood or a price hike or a Nobel laureate changing teams. The model can be turned off. The learning loop is yours. Build the part that's yours.
And Ethan Mollick had the appropriately humble coda on all of this, which I appreciated — he said we honestly don't know the best approaches to rebuilding companies around AI agents yet. Practical agents are only months old. Experimentation and productive failures are going to be required. Which is a grown-up way of saying: nobody has this figured out, the people selling you certainty are selling you something, and the founders who win are going to be the ones who experiment fast and capture what they learn. Capture what they learn. There's that loop again.
Okay. Let me come down off the mountain and give you a couple of things that are just plain interesting, because not everything has to be a treatise on sovereignty.
Here's one I genuinely enjoyed. There's a piece from Anna Heim at TechCrunch about Jean-Baptiste Kempf. If you don't know the name, you almost certainly know his work — he's the open-source legend behind VLC, the little orange traffic-cone media player that has played every weird video file format known to mankind on a few billion computers. The guy who made your free video player run smoothly. Well, now he's building something called Kyber — an infrastructure layer to control remote devices in real time. He's taking that same instinct — make the messy, latency-sensitive, real-time stuff just work — and pointing it at robots.
And I think that's a more interesting signal than it looks. Robotics has a dirty secret, which is that the hard part often isn't the fancy AI brain — it's the boring plumbing. Getting commands to a physical device in real time, reliably, over an unreliable network, without the thing jerking around or freezing mid-motion. That's the unsexy infrastructure problem, and it's exactly the kind of problem a guy who spent twenty years making video playback smooth across garbage connections actually knows how to solve. There's a lesson in there about where the value is in robotics right now. Everybody's staring at the foundation models for the brain. The money and the moat might be in the nervous system — the control layer that makes the body actually obey. Keep an eye on Kyber. The pedigree is real.
And then, because the universe has a sense of humor, here's the most Founders Fund sentence I've read all month. Connie Loizos reports that Founders Fund made an outlier bet on humanely killed fish. There's a company called Shinkei that makes a refrigerator-sized robot called Poseidon, and its whole job is to kill fish quickly and humanely. Now, you laugh, but there's actually real logic under the absurdity. The way most fish die in commercial fishing is slow and stressful, and it turns out a stressed, slowly-dying fish produces worse-quality meat — the flesh degrades faster. There's a Japanese technique for dispatching fish instantly that produces dramatically better, longer-lasting product. The problem is it takes a skilled human and it doesn't scale. So Shinkei built a robot to do it at scale. Better animal welfare and better margins and a longer shelf life, all from the same machine. That's the kind of bet that sounds like a punchline and is actually a real arbitrage between ethics and economics. I tip my cap. A fridge-sized robot named Poseidon humanely dispatching tuna — only in this business.
Quick hits before I let you go. Apple's got iOS 27 rolling, and Lauren Forristal's rundown is basically: the Siri and Apple Intelligence stuff is the flashy headline, but there's a pile of smaller, genuinely useful additions underneath that aren't getting attention. Translation for builders: Apple is still doing the boring blocking-and-tackling of OS improvements while everyone obsesses over whether Siri finally works. Worth a skim if you're building on the platform, not worth a deep dive here.
And here's a fun little mirror for our whole industry's vanity — TechCrunch's Anthony Ha flagged something called In the Weights, which is basically a vanity search for the AI age. The pitch is, what's your In the Weights score — meaning, how much does the AI know about you, how present are you in the models' training. We turned narcissism into a metric. Of course we did. I'd tell you to go check your score but honestly the healthiest score is not knowing and not caring. Moving on.
Let me close the loop on a couple of threads we've been tracking, just so you know they're still alive. The Anthropic-Washington standoff was, as of the most recent reporting, inching toward something — talks reportedly shifting from "fix the unfixable jailbreak" toward designing a framework to actually assess the severity of security flaws in models. Which, if it holds, is a more sane outcome than where it started. But I'll echo the caution a lot of smart people are voicing: we all want Fable back so badly that there's a real risk of reading the first hopeful signal as a resolution when it might still have a long way to run. Don't count the chickens. And the bigger structural worry stands — Aaron Levy from Box made the point that if every model update now has to go through an extensive government review and feedback process, we may be saying goodbye to the era of fast iterative releases and hello to fewer, bigger, slower, more political model drops. That's a real change to the rhythm of this whole industry, and it's the kind of thing that quietly reshapes what's possible to build.
So here's where I leave you today. The throughline, if you want one: the model is not the thing you own. Talent walks. Access gets revoked. Governments get jumpy. Vendors get greedy. The thing that's actually yours — the only thing — is the learning your company captures when your people and your tools do real work together. Build that part. Protect that part. Everything else is rented.
That's the menu for today, kiddos. I'm Tony DeLuca, this has been Barely Possible, and I'm grateful you spent the time with me. Go build something that's actually yours, and I'll catch you on the next one.