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, pour yourself something with a little heat in it, because your boy Tony DeLuca has got a fresh menu of AI morsels laid out on the counter and I want to talk about something that doesn't usually make the front page. Welcome to Barely Possible.
Here's where I want to start today. Not with a model release, not with a chip war, not with somebody's trillion-dollar round. I want to start with a leatherworking school in San Francisco and a 24-inch touchscreen table that knows when you put a toy spaceship on it.
Because there's a report out this week — a piece from Connie Loizos — about a small but growing group of founders who are placing a very specific, very stubborn bet. And the bet is this: that after twenty-some years of technology doing everything it can to keep you staring at a screen, alone, there is now real money in doing the opposite. In getting people back into the same room. Not a group chat. A room. With chairs. And other humans in them.
Now I've been doing this long enough to be skeptical of anything that smells like a trend piece. "Founders discover human connection" — okay, sure, and I discovered the diner has good coffee. But when I actually sat with this one, there's something underneath it that matters for you, the person building a company. So let's dig in.
The two people at the center of this piece have both already done the thing every founder dreams about. Brynn Putnam built the connected-fitness company Mirror — the mirror you work out in front of — and sold it to Lululemon for five hundred million dollars in cash in under three years. Tristan Walker built the men's grooming brand Walker & Company and sold it to Procter & Gamble. These are people who guessed right once. They read where the puck was going and they got there first.
And here's the part that stopped me. They're both back at it, and their new companies could not look more different on the surface — and yet they're chasing the exact same thing.
Putnam's new company is called Board. It's a big touchscreen table. But the whole design philosophy is the reverse of Mirror. Mirror was one person, alone, working on — and I'm quoting her here — "myself, my body, my fitness, my reflection." She built that when she was pregnant with her first kid, trying to bring the studio into her living room through a screen. Now she's a mother of five in a blended family — household went from three people to seven — and she says she kept hitting the same wall: there was no product that let people of wildly different ages and abilities play together. A six-year-old and a seventy-year-old at the same table, same game, no shared learning curve. That's why the table recognizes physical objects instead of using a controller. You put a spaceship down, something happens. You put a knife down, it cuts. A kid and a grandparent don't have to learn the same buttons.
And the way she frames the whole pivot is the line I keep coming back to. She said it wasn't a "movement from digital to analog," so much as — quote — "how can we use tech to bring people together rather than take them apart." That's not anti-technology. That's a different design objective for the same technology.
Now flip to Tristan Walker. His new thing, Heirloom Craft, is buying up and scaling fine-craft trade schools. First one is a leatherworking school in San Francisco, founded by the first American artisan-ambassador Hermès ever had. And Walker's motivation is where the AI angle sneaks in the back door. He didn't start this because he loves leather. He started it, he says, because he was worried about AI — quote — "stealing knowledge work" and pulling people away from each other. His argument is that America has spent fifty years breaking the chain from master to apprentice — offshoring the production, stripping the prestige out of working with your hands, until an apprenticeship became the thing you did when you couldn't get into a four-year college.
And he's got a number he threw at a room full of investors: roughly half of Gen Z and Gen Alpha say they don't want a four-year degree at all. Here's the kicker that a builder should really chew on. At his first San Francisco location, more than half the seats are filled — not by career-changers, not by retirees — by young tech workers. Software people. Looking for what he calls a "phones-down creative outlet."
Think about that for a second. The people building the machine that's automating knowledge work are the ones lining up to learn how to hand-stitch a wallet. That's not nothing. That's a signal.
Now — and this is where the OG in me kicks in — I want to put the brakes on the romance and look at the money, because that's the part that separates a movement from a magazine cover.
The dollars here are small. Board has raised thirty-five million. Pie — that's Bonobos founder Andy Dunn's "social life operating system," his words — has raised twenty-four. Six Bells, the countryside inn Audrey Gelman runs where they do a monthly murder-mystery dinner and somebody "turns up dead" by dessert — that one's raised about three-point-eight million. Heirloom hasn't disclosed. So we're talking about a category, if it even is a category, that in total is a rounding error next to a single AI infrastructure round.
And there's the exception that proves the rule. Adam Neumann — yeah, the WeWork guy — his new residential real estate play, Flow, marketed around the same idea of solving loneliness through physical space, has raised more than four hundred and fifty million dollars, mostly from Andreessen Horowitz. That one number dwarfs everyone else in the story combined. Which tells you something honest: when "togetherness" gets attached to real estate, capital shows up, because real estate is a thing investors understand how to model. When it's attached to getting strangers to actually show up for each other — as good as that is for society — the money is cautious. Because, as the piece puts it plainly: getting strangers to show up for each other isn't nearly as scalable as shipping software.
That's the whole tension for a builder. And I want to be straight with you about it. This is not me telling you to go quit your SaaS company and open a whittling school. The economics are genuinely harder. You can't copy-paste a room full of people the way you copy-paste a server.
But here's the useful part, and it's grounded, not sentimental. There's a real demand signal underneath all this. Last year the World Health Organization's Commission on Social Connection formally classified loneliness as a defining global health challenge — one in six people worldwide affected, linked to roughly eight hundred and seventy thousand deaths a year. And in a Harris Poll this year, two-thirds of Americans said they're prioritizing experiences over stuff, and the same two-thirds said retail shopping feels too generic — everything looks and feels the same. That's a market telling you something about what it's tired of.
So if you're building, the question this story actually poses isn't "should I go analog." It's "what am I optimizing my product to do to the human using it." Because Putnam made the exact same category of product twice — a screen with software behind it — and pointed it in two opposite directions. One isolated people by design. One tries to gather them by design. Same tech. Different objective function. And she's betting the second objective is where the next decade of demand lives.
And you'll notice these are all people who've built a real brand once before. That's not a coincidence either. That's what makes an investor willing to write a check on something this hard to scale — a founder who's already proven they can build something people actually buy. We'll know in a year or two whether the customers follow. But the bet itself is worth watching, because it's a bet about what people are hungry for that a spreadsheet doesn't fully capture.
Alright. Let me pull the thread from that story into the next one, because there's a real connection here and I don't want to fake it. Walker's whole anxiety was about AI "stealing knowledge work" and moving too fast. Well, funny thing — the people who make the AI are having that exact argument out loud right now, and it got heated this week.
There's a piece off the latest TechCrunch Equity podcast — Anthony Ha writing it up, with Kirsten Korosec and Sean O'Kane — asking a very simple question: is the AI industry actually ready to slow down? And I want to be careful here, because we've circled the "pace the frontier" theme a couple times recently on this show. Amodei's plan, Huang's speakerphone call with Trump — we've been here. So I'm not going to re-run that tape. What's genuinely new and useful in this conversation is one specific argument about why a slowdown probably can't happen even if everybody says they want it, and it's an argument that matters for anybody selling into enterprise.
The setup, quickly: Anthropic's Dario Amodei put out a plan to "pace the frontier." Sam Altman's using similar language. Even Elon, to a degree. And a surprising number of industry people coalesced around it fast — which Korosec makes a sharp point about. She says consensus can be a good thing, but — quote — "it can also be the beginnings of forming a cartel." Hold that thought, because it's exactly right and we'll come back to it.
But the meat is Sean O'Kane's argument, and it's the one worth your time. Korosec asks him the obvious free-market question: look, we've got existing regulations, we've got a competitive market — if some AI company ships something genuinely unsafe, won't the market just punish them? Won't customers leave, won't regulators come down, won't they stop being a viable company?
And O'Kane basically says: in a clean market, sure. But we're not in a clean market. Two reasons. One — and I'll quote him — "we have a federal government at the moment that certainly does not seem eager to enforce regulations broadly speaking, let alone specifically speaking here." Okay, that's the political layer. But the second reason is the one that should make a founder sit up. He says there isn't much consumer choice driving this market anymore, because these companies have moved into enterprise. His words: "Corporations are not going to pack up and move from Codex to Claude Code just because they disagree on principle with something that OpenAI did."
Sit with that, because that is the whole game. The old theory of tech accountability was you vote with your dollars — company does something ugly, users churn, stock drops, behavior changes. But when your revenue is enterprise contracts, not fickle consumers, that feedback loop basically goes dead. A Fortune 500 doesn't rip out its entire coding infrastructure over a moral disagreement. The switching cost is enormous, the integration is deep, and the procurement department does not care about your Twitter feelings. So the market's self-correction mechanism — the thing we all assume is quietly keeping these companies honest — O'Kane's argument is that it's mostly not operating. And on top of it, he points out these companies are floated by so much investment money that they can absorb whatever losses a scandal might cause anyway.
Now put that next to Korosec's cartel line, and you get the real picture for a builder. If the frontier labs coordinate on "safety" — which sounds noble — and there's no meaningful consumer pressure and no aggressive regulator, then what you might actually be watching is a small number of very well-capitalized incumbents agreeing on the rules of a market where they already hold the switching costs. That's not a knock on safety work — real safety work matters, and we've covered it. It's a caution about who benefits from the framing.
And here's why this lands on your desk specifically. If you're building on top of these labs, the lesson isn't "pick the moral one." The lesson is that you are the fickle consumer they no longer have to worry about. Your leverage is at the moment you architect your dependencies — not after. Which, as it happens, is exactly why we keep hearing enterprises talk about keeping optionality, running multiple models, not marrying a single API. That's not paranoia. That's the rational response to a market where the exit door is expensive.
O'Kane's own summary of the whole slowdown idea: "It's a nice idea in theory, but I feel like in practice, there's none of that really happening." That's the sober read, and I think it's the right one.
Alright, let me shift gears entirely, because not everything today is a philosophy seminar. Let's talk about cars and robots that drive them, and one very specific date on the calendar you should know about.
Kirsten Korosec's mobility rundown had a question in the title that I think is more important than it sounds: how do we know when a self-driving car is actually safe enough? And the honest answer right now is — we don't, not in any standardized way. Here's the thing that got me. We make a sixteen-year-old take a driver's test before we hand them a license. There is no equivalent test for a robot driver. None. Whether an autonomous system is a safe driver has, up to now, largely been a question the companies get to answer about themselves. The homework grades itself.
The federal safety agency — NHTSA — says it's working on that. The administrator, Jonathan Morrison, said this week they don't want to be stuck playing "whack-a-mole" with policy, and they've struck an agreement with the engineering standards body SAE to launch something called ASCEND — a government-industry effort to build data-driven performance standards for autonomous driving. The stated goal is an actual competency standard — a driver's test for the machine — that could replace the current patchwork of inconsistent state and local rules.
Now, no timeline yet on when this produces anything real. And Korosec makes the dry observation I'd make too: the AV industry is absolutely already in there weighing in on what "safe enough" means. When the referee and the players design the rulebook together, you watch that rulebook closely.
But for a builder, this is the shape of a whole opportunity forming. Because the current setup is a mess of federal rules that literally weren't written for these vehicles — the Tesla Cybercab and the Zoox robotaxi don't have pedals, don't have steering wheels, don't have side mirrors, which are all things federal rules currently require. When the framework doesn't fit the product, somebody has to build the connective tissue — the testing, the validation, the monitoring, the compliance tooling. That's a business.
And the deal that caught my eye in the same rundown: an AV company called May Mobility is planning to go public through a SPAC merger, a deal that could raise more than three hundred million at a one-point-four billion dollar valuation. May Mobility runs autonomous Toyota Siennas in three locations, and the claim is they'd be the first U.S. public company focused purely on autonomous ride-hailing. Sean O'Kane's note on it is the interesting part — it's a test of their "asset-light, partnership-first" approach. Meaning: can you win in robotaxis without owning the whole stack, the whole fleet, the whole capital-heavy nightmare? Wall Street's about to tell us whether it'll fund a pure-play robotaxi bet. That's a real signal for anyone thinking about capital-light versus capital-heavy in a hardware-adjacent space.
And the one date I promised you: Zoox — that's the Amazon-owned one — is currently capped at a hundred vehicles in Las Vegas, its only commercial market. That cap expires September 25th. After that, they can expand the fleet. So if you're watching how fast robotaxis actually scale in a real city, mark it.
Also in there, quickly, because it's the kind of thing that's easy to miss and matters: a Joby Aviation aircraft flew more than 3,100 miles across the country with nobody touching the controls. Korosec's framing — this is more than a splashy demo. Autonomous flight over that distance is a different order of problem than a robotaxi doing loops downtown. Keep an eye on it.
Now, from the road to the classroom, or — let's be honest — to the algorithm that ate the classroom. There's a startup called ScrollEd, out of Palo Alto, pitching at the Disrupt Startup Battlefield, and their whole idea is one of those things that's either genius or the end of civilization depending on what mood you're in.
The pitch: kids won't open a textbook. So instead of fighting the scroll, make the textbook scroll. You feed it any text file — a dry PDF, a whole textbook — and it turns into a feed that looks and feels like TikTok or Reels. AI-generated video, audio, quizzes. Swipe up, new topic. Swipe sideways, go deeper on the one you're on, quiz at the end.
Now my first instinct — and probably yours — is: great, we've invented a way to doomscroll your chemistry final. But the founder, Utsav Gupta, said something that I actually respect. He said they're being — quote — "deliberate about how we use it and choose not to maximize engagement on the platform." And this line: "A short video can make you curious; following that curiosity can mean searching for another explanation or leaving the app. We wanted going deeper to be as easy as moving to the next video."
That's the interesting inversion. Every social platform is engineered so the easiest possible action keeps you swiping sideways through junk. ScrollEd's claim is they're engineering it so the easiest action is going deeper into one subject. Same mechanic, opposite objective. Which — hey — is the exact same thing Brynn Putnam said about her table. Same tool, pointed at a different goal for the human holding it. I didn't plan for these two to rhyme, but they do.
Do I believe an 18-year-old's brain will reliably choose "go deeper" over "next"? I have my doubts, and I've got a nephew who'd prove my doubts right in about four seconds. But it's a real product thesis, and the business model underneath is the grown-up part: freemium for consumers, but the money's in institutional licenses — schools and corporate training pay to deliver material through it and get back reports on engagement and progress. That's a defensible enterprise wedge dressed up in a consumer costume. Watch whether the deeper-not-wider promise survives contact with actual teenagers. That's the whole ballgame for them.
Let me stay on the hardware-meets-privacy beat for one more, because it's short and it's the kind of thing a founder should think hard about before shipping. Ivan Mehta reviewed a new gadget from a company called Vocci — it's a ring, a piece of jewelry, that records your meetings. Double-tap to start, double-tap to stop. Under six grams, titanium, records about eight hours on a charge. And by the transcription account, it works well — pulled clean transcripts out of hour-long conversations in loud cafés.
Here's the wrinkle, and it's a real one. The software's half-baked — the app's confusing, the AI-generated "insights" were sometimes longer than the note you took, and it took weeks before you could even connect it to other assistants. Fine, early product. But the thing that should stop you is the form factor itself. It looks exactly like a normal ring until somebody notices the button. There's a recording indicator light — but it faces the wearer. Not the person being recorded. And Mehta's blunt about it: it's best practice, and in some places outright the law, to tell someone you're recording them. A device that hides the fact that it's recording is a device built, whether the company means it or not, to make non-disclosure easy.
Now the builder lesson here isn't "don't make cool hardware." It's that when you design a capture device, the defaults you choose are an ethical statement whether you want them to be or not. Where does the indicator light point? Who does the product make it easy for — the wearer, or everyone else in the room? A pendant, a pin, a card — those at least look like devices. A ring that passes for jewelry crosses a line the company can't fully hand-wave with "we encourage people to disclose." At two hundred and forty-nine bucks, this is a niche product. But the category — always-on personal capture — is coming fast, and the companies that think about disclosure by design are the ones that won't get eaten alive when the regulators show up. And on the recording-disclosure beat, they are showing up — California just this week put teeth into a law fining influencers up to five thousand bucks a pop for not disclosing paid political posts. Disclosure is having a moment. Build accordingly.
Alright, let me do a couple quick ones and then get out of your hair.
Yann LeCun had a jab out this week that made me chuckle, and I'll frame it right because it references older launches. He was comparing two secretive AI labs. His own outfit, AMI Labs, launched February 2026, raised about a billion. Safe Superintelligence — Ilya Sutskever's SSI — launched back in June 2024 and has raised eight billion. And LeCun's dig: AMI Labs has "released exactly as much as SSI" — which is to say, both have shipped basically nothing public — "except that we've actually spoken about what we're doing and have published a few papers." It's a shot, sure. But there's a real point buried in the trash talk about a market where you can raise billions on pure reputation and radio silence. Eight billion dollars for a company whose entire public output is the promise that it's working on something big. That's the era we're in. File it under: capital is not the same thing as evidence.
And one more, because I said I'd tell you what to watch. There's an older report resurfacing about the open-weights safety problem — Base Labs, from the inference company Baseten, partnering with Hugging Face and Goodfire on safety infrastructure for open models. The number in there is the thing to hang onto: Hugging Face currently lists over six thousand "abliterated" models — that's open models where somebody's stripped the safeguards out. Six thousand. That's the actual live tension under all the polite "pace the frontier" talk we walked through earlier. The frontier labs can coordinate all they want in their conference rooms — but the open-weight world doesn't sit at that table, and it's not waiting for permission. Whatever safety cartel forms up top, six thousand un-safetied models are already in the wild. Keep that in the back of your mind whenever somebody tells you the labs are going to handle this.
So let me tie the bow on it. If there's a thread running through today, it's not "AI good" or "AI bad." It's a question about intent. Putnam building the same screen twice and pointing it in opposite directions. ScrollEd taking the doomscroll mechanic and trying to aim it at depth. The Vocci ring making a design choice about who its little light faces. And the whole slowdown debate coming down to whether the people building this stuff actually face any pressure to point it anywhere good — with O'Kane's hard truth that when you sell to the enterprise, the crowd that used to keep you honest isn't in the room anymore.
Same tools. The whole thing is where you point them. That's the founder's actual job, and no model release is going to do it for you.
That's the menu for today. Your boy Tony DeLuca — go build something you'd be willing to put your name on. Talk soon.