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, welcome back. I'm your boy Tony DeLuca, and this is Barely Possible, where we sort through the AI news pile so you don't have to eat the whole thing yourself. We've got a leaner menu today than some of the frenzied model-drop weeks we've been chewing through, and honestly, I'm glad for it. Because when the headlines slow down a notch, you can actually see the machinery underneath. And the machinery today is telling me something interesting about where the money and the muscle in this business are actually pointing. Buckle up, let's have at it.
Let me start with the one that jumped out at me, because it's not a chatbot story and it's not a lawsuit story, and Lord knows we've had a run of lawsuit stories lately. Reed Jobs, son of you-know-who, sat down with TechCrunch's Connie Loizos in a piece that ran on the eleventh. And the headline gets the framing right: Reed Jobs would rather talk about curing cancer than his last name. Fine. Let's talk about the cancer, and let's talk about the business.
Here's what's actually in the piece. When TechCrunch last sat down with Jobs, at TechCrunch Disrupt, that was nearly three years ago. His venture firm Yosemite was brand new back then, and biotech, as they put it, was still reeling from its post-pandemic crash. Now? Yosemite has a team of seventeen. And two things have changed in the water around them that matter for anybody building in this space.
First thing: a cluster of blockbuster drugs are all losing patent protection in roughly the same window. That's the so-called patent cliff, and it's a real thing on the pharma calendar — big money-making drugs going off-patent around the same stretch, which cracks open all kinds of new opportunities for anybody who wants to move into that space. Second thing, and this is the one for us: AI has gone, in Jobs's own words, from a curiosity to a huge part of what Yosemite does. And the line that stuck with me, direct from Jobs: "I didn't expect Yosemite to be moving this fast."
Now I want to be careful here, because the piece is an interview and I'm not going to invent a bunch of drug-pipeline specifics that aren't in front of me. But I'll tell you why this is worth your two minutes even if you never touch biotech. Because what you're hearing is a venture investor describing the exact shift that everybody keeps yapping about in the abstract, but happening in a domain where the stakes are life and death and the regulation is real. Three years ago, AI was a curiosity at a cancer-focused venture firm. Now it's a huge part of the operation, and the firm is moving faster than its own founder expected.
That's the tell. When the technology goes from "interesting demo we should keep an eye on" to "central to how we allocate capital and pick shots," in a field as conservative and evidence-bound as drug development — that's not hype talking. Pharma does not get excited easily. These are people who spend a decade and a couple billion dollars to bring one molecule to market and half the time it fails in phase three. If those people are telling you AI has become core to how they operate, that carries a different weight than some consumer app founder telling you the same thing.
And here's the founder lesson buried in it. The opportunity Jobs is circling isn't "AI." It's the patent cliff plus AI. It's a real business dislocation — a wave of drugs coming off exclusivity — showing up at the same moment a new tool arrives that changes your cost structure and your speed. That intersection is where the money is. Not the tool alone. The tool landing on top of a market that was already about to move. If you're a builder, that's the pattern to hunt for. Don't go looking for where AI is cool. Go looking for where some big boring structural change is already forcing an industry to rethink itself, and then bring the tool to that fire. That's what a smart venture shop does, and that's why Yosemite's moving fast.
Now let me connect that to something we hear constantly and rarely with numbers attached: the gap between companies saying they're all-in on AI and companies that actually have people using it well. That belief that AI is suddenly core — Reed Jobs feeling it at a seventeen-person biotech fund — has a much messier cousin inside big enterprises, and there's a recent piece of commentary that put some hard figures on that mess.
This came through a recent episode of the AI Daily Brief, Nathaniel Whittemore's daily show, and I'm going to treat it the way I always treat another podcast — pull the one useful argument, add my two cents, and get out. The number that stopped me: a proficiency report they cited found that while sixty-nine percent of workers surveyed said their organization had taken some action on AI agents, only sixteen percent actually use an agentic tool at work. And fewer than ten percent — one in ten — could define what an AI agent even is in their own words. And here's the kicker that explains the whole thing: only thirty percent of employees at organizations that have AI agents had actually received any training on them.
Chew on that for a second. Sixty-nine percent of companies "taking action." Ten percent of employees who can tell you what the thing is. That is the entire enterprise AI story in one gap. The press release says transformation. The floor says, what agent? So when you hear a CEO on an earnings call brag about their AI initiative, remember that "we took action" and "our people can use it" are two completely different sentences, and the distance between them is measured in training that mostly didn't happen.
But the part of that discussion I actually want to hand you is the counter-example, because it's a playbook you can steal. The show quoted Uber's CTO, Praveen Napali, describing what they call "agentic pods." And the setup is simple enough that you could run a version of it next month. You take your most AI-proficient engineers — people who deeply know your systems — and you pair each one with a domain expert from an actual business function. Finance, legal, ops, customer support. Then you give the pod exactly ten days. Days one and two, the engineer just shadows the expert. Watches every step, documents the workflow, asks dumb questions, builds intuition. Day three, prioritize what's worth attacking. Days four and five, build a working agent sitting right next to the person who does the job. Days six through nine, validate it against other people doing that same work — does it generalize, does it actually make the job better. Day ten, ship it.
And the results Napali laid out were not subtle. Capital allocation across a hundred and fifty cities went from fifteen hours to thirty minutes. Financial pacing reports from two days to ten minutes. One QA workflow from two weeks to fifty minutes. They ran sixteen of these pods across sixteen business functions in two months.
But here's the line from Napali I want you to write on a sticky note: "The best AI opportunities are rarely visible from the outside. You discover them by sitting next to the people doing the work." That is the whole enchilada right there. The reason ninety percent of enterprise AI is stuck is that somebody in a conference room looked at a process diagram and tried to automate a task off the chart. And Napali's point is that the diagram lies. The real friction, the handoffs, the pointless approvals, the legacy tool everybody hates — you only see that stuff by sitting in the chair next to the person doing it. And once you see it, you're not automating a task anymore. You're redesigning the whole workflow. That's the difference between fifteen hours to thirty minutes and some chatbot that summarizes meeting notes nobody reads.
So my takeaway, and I'll bring it back to Jobs before we move on: the winners in this cycle, whether it's a cancer fund or a ride-hailing giant, are the ones putting the technology directly next to the actual work. Yosemite putting AI next to the science. Uber putting engineers next to the finance team. The losers are the ones putting a press release next to a procurement contract. Same technology. Wildly different outcome. And the variable is proximity to the real work.
Alright, let me shift from the enterprise floor to the open road, because there's a mobility story that ties right into who's got the muscle in this business.
TechCrunch's Kirsten Korosec ran her Mobility newsletter on the twelfth under the headline "A robotaxi ultimatum." Now I'll be straight with you — the source I've got is the top of the newsletter, the welcome-back framing, not the full guts of whatever the ultimatum is. So I'm not going to pretend I can read you the specifics of some deal I can't see. What I'll do is tell you why the word "ultimatum" in a robotaxi context is worth pausing on.
The robotaxi business has quietly become one of the few places in AI where the rubber literally meets the road — where the model has to work in the physical world, in real traffic, with real liability, every single trip. There's no "it hallucinated, oops" when the thing is two tons moving through an intersection. And the framing Korosec leads with — that AI is now, more than ever, part of the future-of-transportation story — that's the thing to sit with. We spent a couple years treating self-driving as its own weird island, separate from the chatbot boom. That wall is coming down. The same companies, the same compute, the same frontier-model muscle, is increasingly the thing steering the cars. When a mobility writer says the AI piece matters more than ever, that's the merge happening in real time. I'll flag the full breakdown when there's more meat on it, but keep robotaxi on your radar as an AI story now, not a separate car story.
Now let me pull us over to the physical world in a different way, because I want to talk about a music tool that I think is a genuinely nice little example of a narrow model doing real work.
There's a recent post — this went around on the tenth, with Yann LeCun amplifying it — from a shop called MireloAI, working together with Kyutai Labs, introducing a new Audio-to-MIDI model. And the pitch, from what I can see of it, is exactly what it sounds like: you feed it a finished recording, and it identifies and pulls out the musical information — turning a finished audio track back into MIDI, which for the non-musicians out there is the editable, note-by-note representation of what's being played. Think of it like taking a baked cake and getting the recipe back out.
Now that is hard. Real audio is a soup — multiple instruments, overtones, reverb, all mashed together into one waveform. Separating that back into "here's the actual notes each instrument played" has been a stubborn problem for a long time. And I bring it up not because it's going to move a trillion dollars, but because it's the kind of thing I keep telling you to watch for: a specialized model that does one clearly-defined transformation extremely well, in a domain where the output is immediately useful to a working professional. A producer, a composer, somebody scoring to picture. That's not AGI. That's a tool. And tools that a professional will actually pay for and use every day are, frankly, a better business than half the AGI fan-fiction floating around. I flagged the demo's the interesting part — link'll be in the show notes. Just a clean little reminder that "narrow and excellent" beats "broad and mediocre" almost every time you're building something people actually depend on.
Speaking of the open web and who controls it, let me get into a fight I've got real feelings about: the crawler wars.
There's a recent report from The Verge that Patreon is working with Cloudflare to squash AI crawlers. Now I've only got the headline framing on this one, so I'm not going to invent the technical particulars of the deal. But the shape of it is a story we've been circling for a while, and it's worth naming plainly. Patreon is a platform built on creators putting work behind a paywall so their audience pays them directly. And the AI crawlers — the bots that scrape the web to feed training data — are, from a creator's point of view, showing up to eat that work for free and give nothing back. So Patreon teaming with Cloudflare, which sits in front of a huge chunk of the internet's traffic and can identify and block those bots, is the platform effectively drawing a fence around its creators' stuff and saying: you want this, you negotiate, you don't just take it.
And here's why this matters to you as a builder even if you never touch Patreon. The free-scraping era of the open web is closing. Slowly, unevenly, but closing. Cloudflare has been building tooling to let sites detect and block or charge AI crawlers, and platforms are lining up to use it. If your product's business model quietly assumes you'll be able to hoover up open web content for free forever, that assumption is decaying under you right now. The web is being partitioned into stuff you can take and stuff you have to pay for, and the paywall is getting smarter about telling the difference. Plan for a world where data has a toll booth. Because the toll booths are going up.
Now, that theme — who gets to take what, and who gets squeezed — actually runs straight through a couple of the harder stories in today's pile. Let me get into one that has nothing to do with AI on its surface and everything to do with the pattern.
Ars Technica ran a piece, reported by Mark Olalde of ProPublica and Jimmy Tobias for High Country News, on the Bureau of Land Management rewriting its grazing regulations. And the headline is the story: the overhaul seeks to cut public involvement. This is the first time since 1995 — three decades — that the BLM is rewriting the rules for how livestock grazing works on public lands. That's federal land, land that belongs to all of us in the technical and legal sense, being used by ranchers to graze cattle under a permit system.
And the through-line the reporting flags is that the rewrite is aimed at reducing the public's ability to weigh in on how that land gets used. Now why am I putting a cattle-grazing story in an AI show? Because it's the same move, in a completely different arena, that I keep pointing at. When you're rewriting the rules that govern a shared resource — whether it's public grassland or the open web — one of the quietest, most consequential levers is who gets a seat at the table when the decision gets made. Cut public involvement, and you've changed the outcome before anybody votes on it. It's a process story dressed up as a paperwork story, and the process is where the power lives. I'd keep an eye on that one, because "we streamlined the public comment period" is one of those phrases that sounds like efficiency and functions like a locked door.
Let me stay in the public-health-and-policy lane for one more, because it's a recent report worth flagging.
Ars Technica's Beth Mole, writing on the tenth, covered a study on the changes RFK Jr. has made to the measles vaccine schedule. The framing from the piece: the study confirms the change will hurt vulnerable toddlers. The specific mechanism, as Ars lays it out, is about the combination shot — the kids who get a combination shot are, in the words of the piece, some of the most vulnerable. Now I'm going to stay in my lane here. I'm a radio guy, not a pediatrician, and I'm not going to freelance on the medical particulars beyond what the reporting says. But the reason it belongs in a briefing for builders and decision-makers is this: it's another instance of a policy change at the top rippling down to the people with the least ability to absorb it. That's a theme with legs whether you're talking vaccine schedules, land access, or web data. The decision gets made way upstream, and the cost lands on whoever's standing at the bottom of the hill. If you're building products that touch health, families, or anything regulated, the regulatory ground is moving, and it's not moving in a straight line. Watch it.
Alright, let me lighten the load, because it's summer and not everything's a policy fight.
TechCrunch's Lauren Forristal wrote up a review — during last weekend's brutal heat wave, she put the new Ninja Slushi Twist to the test instead of trudging out for a Slurpee. Ninja's latest slushie machine, building on the popularity of the original. Now look, this is a countertop appliance review, and I'm not going to pretend it's a signal about the future of intelligence. But I'll give you the honest builder-brain read on why a machine like this sells: it takes a small, real, recurring annoyance — it's a hundred degrees out and I want a frozen drink and I don't want to leave the house — and it kills it, reliably, for a price people will pay. That's it. That's the whole business. No AI, no agents, no moat made of magic. Just a clear job that a customer feels in their bones every hot afternoon. And I mention it right after all this frontier-model talk on purpose, because sometimes the most durable product in the room is the one that solves the dumbest, most human problem completely. Keep that humility handy when you're tempted to bolt a large language model onto something that just needed to make a good slushie.
Now here's a product story that's more in our wheelhouse. TechCrunch's Sarah Perez, in a recent report, wrote up a new app called HyperTexting, which aims to turn the open web into a scrollable, social-media-like feed. The idea is that it takes websites, blogs, newsletters, and podcasts and stacks them into a feed you scroll like Instagram, while also making it easier to post to your own website.
Now, my first reaction was, didn't we already invent this? It's called RSS, and it's older than some of you listening. But I want to be fair to it, because there's a real thesis underneath. The open web — the blogs, the independent sites, the newsletters — lost the attention war to the algorithmic feeds. Not because the open web had worse content, but because it had worse ergonomics. Scrolling a feed is frictionless. Checking twelve bookmarked blogs is a chore. So HyperTexting's bet is that if you give open-web content the same slick, scrollable wrapper that the social platforms use, people will actually consume it, and creators will get an audience without renting it from a platform that can change the rules on them Tuesday.
And that connects straight back to the Patreon and Cloudflare story, right? Both are, at bottom, about the open web trying to reclaim ground from the walled gardens. One's about keeping the scrapers out. One's about making independent publishing feel as good to consume as the feeds do. Whether HyperTexting works, who knows — the graveyard of "RSS but nicer" apps is deep and full. But the direction of travel is real: a bunch of smart people are trying to rebuild a version of the web where the creator, not the platform, holds the relationship with the reader. As a builder, if you've been assuming the platforms own distribution forever, notice that a lot of money and effort is now betting against that. Might be worth thinking about where you'd stand if the feed weren't the only game in town.
Now let me get to a cluster of OpenAI-flavored items, and I'm going to be careful here because a couple of these are thin on detail and I'm not going to build castles on sand.
There's a Verge item pointing at GPT-5.6, Codex, and ChatGPT Work. Now, we already covered the GPT-5.6 and ChatGPT Work launch back on the tenth — Altman did the livestream, the desktop app, the whole enterprise push against the incumbents. So per my own rule, I'm not going to re-run that story or re-quote anybody. If there's a genuinely new wrinkle in this particular piece about Codex and how it plugs into the Work product, I don't have enough in front of me to responsibly tell you what it is, so I'll just note it's a continuation of that same enterprise-knowledge-layer play we talked about a few days back and leave it there. And a quick callback while we're here — remember Altman's line we flagged, that the benchmarks suggest 5.6 is the best model in the world "right now"? The operative words there being "right now," which in this business means "until roughly Thursday." Nothing new to add on that today.
There's a second Verge item in the pile with a headline suggesting OpenAI's Fidji Simo is stepping down from a CEO role into an advisor role. And here's where I'm going to be a hard-nose, because I've got standards on this show and one of them is: I do not report executive departures off a headline I can't verify with a body. I've got a URL slug and nothing underneath it — no article text, no date I trust, no confirmation of what actually happened or when. Personnel moves at a company like OpenAI are exactly the kind of thing that gets garbled in the retelling, and "steps down as CEO to advisor" could mean a dozen different things depending on which CEO role and what the actual transition is. So I'm going to do the responsible thing and not narrate it as fact. If it firms up with real reporting, we'll cover it properly with the details. Until then, file it under "heard something, waiting on the real story." That's the honest move, and you deserve the honest move.
Same discipline on a GitHub item floating in the pile referencing a Claude Code issue involving Anthropic and Microsoft. It's an issue tracker link with no substance I can responsibly read to you — could be a bug report, could be a feature request, could be nothing. I'm not going to spin a corporate-intrigue narrative out of an issue number. If the developer-tooling relationship between those two shifts in a way that matters, it'll show up somewhere I can actually stand on. Not going to guess.
I want to flag one more that I'm treating with real caution, and then I'll tell you the deeper point. There's a New York Times piece, from the tenth, on AI-enabled terrorism — the reporting looks at Boko Haram in Nigeria and how AI tools factor in. And there's a related report from an outfit called CASP on AI-enabled terrorism as well. Now, I only have the pointers to these, not the full bodies, and the CASP report in particular I can't reliably date, so I'm not going to imply any of this happened today or read you specifics I can't back up. But I'll name the category, because it's important and it's coming: as these tools get cheaper and more capable, the same capabilities that let a small startup punch above its weight also let bad actors punch above theirs. That's not a reason to panic, and it's definitely not a reason to trust a scary headline I can't read the guts of. It's a reason to understand that the misuse conversation is going to get louder and more concrete, and it's going to shape regulation that lands on everybody building in this space — including the people building nothing but boring, useful, legitimate tools. When the misuse stories get specific, the rulemaking gets specific right behind them. Watch that space, and demand the actual reporting before you believe the framing.
Now, since I brought up regulation, let me note one more marker on the map. There's an item pointing to the AI Now Institute — that's a research and policy shop focused on the social implications of AI. I've just got their homepage in the pile, no specific new report I can read to you, so I'm not going to manufacture a finding. But I'll say this: the fact that a policy-research institute keeps surfacing alongside the terrorism reporting and the crawler wars and the land-use rules tells you the same thing every strand of today's episode is telling you. The governance layer is filling in around this technology. Not in one dramatic law, but in a hundred small fights over who gets to take what, who gets a seat when the rules get written, and who eats the cost when something goes wrong. If you're building, that layer is now part of your terrain. You don't get to pretend it's someone else's problem anymore.
Let me tie a bow on the tech pile and then give you the two odd-and-ends I couldn't resist.
Ars Technica's Kyle Orland, in a recent report, wrote about Valve's new Steam Machine verification system, and the gist is that it raises more questions than it answers — dozens of titles that are too taxing for the Steam Deck are still sitting there unrated for the new hardware. If you're not a PC gaming person, the short version: Valve slaps a "verified" label on games to tell you they'll run well on their hardware, and Orland's point is that the labeling for the new Steam Machine is conspicuously silent on exactly the demanding games where people most want the answer. And I mention it in a builder context because it's a clean little lesson in trust and labeling. A "verified" badge is only worth anything if it covers the hard cases. The moment your quality signal goes quiet precisely where the stakes are highest, people stop trusting the badge entirely. If you ship any kind of certification, rating, or "works with" label in your product — and a lot of you do — the Valve situation is your cautionary tale. Cover the hard cases or don't bother with the badge.
And now the two that have nothing to do with any of this, which I keep in the show on purpose because a steady diet of corporate maneuvering will rot your brain.
Ars Technica ran a piece — this was authored by David Sear, Manoj Joshi, and Mark Peaple, writing for The Conversation — on what they call the real mystery behind Moana. The question is a genuinely great one: after roughly seventeen hundred years, why did Polynesians suddenly start sailing east? These were some of the greatest open-ocean navigators in human history, and there's this long puzzle about the timing — they settled a big chunk of the Pacific and then, after a very long pause, made this dramatic push eastward across enormous distances. And the new angle the piece brings is climate evidence — the idea that shifting climate conditions may have opened windows that made those long eastward voyages possible in a way they weren't before.
Now, I'm not going to overclaim the science, because what I've got is the framing, not the full data set. But I love this story for the same reason I love the good AI stories: it's about timing. Why then? Why not two hundred years earlier or two hundred years later? The answer these researchers are reaching for is that the capability — the boats, the navigation skill — may have been there for a while, but the conditions had to align to make the leap worth taking. And that, my friends, is the same lesson Reed Jobs was giving us at the top of the show, dressed up in a very different outfit. The tool alone doesn't make the move happen. The tool plus the moment does. The Polynesians had the canoes. They needed the wind to change. Yosemite had the AI curiosity. It needed the patent cliff. Timing is the whole game, whether you're crossing the Pacific or building a company.
And the last one, because I'm a sucker for a shipwreck. Ars Technica's Jennifer Ouellette, in a recent report, wrote up the first images of the Quest — that's Shackleton's last ship. The expedition captured the first images of the wreck, and the finding is bittersweet: the wreck is in worse shape than expected, but it's turned into a thriving marine ecosystem. And there's something almost poetic in that, isn't there? The thing falls apart, and life moves into the ruins and makes something new. I'll let you draw your own metaphor for the tech industry. I've got a few, but I'll keep them to myself.
Let me pull it all together before I let you go, because I think there's a real thread here and it's not the one you'd expect from an AI show.
The loud stuff today — the model launches, the executive shuffles — was mostly thin or already covered, and I'm not going to pad it out to sound busy. The stuff with actual meat on it was quieter and it all pointed the same direction. Reed Jobs telling us AI became core to a cancer fund because it landed at the exact moment the industry was already about to move. Uber showing that the enterprise AI win comes from sitting in the chair next to the person doing the work, not from a diagram. Patreon and Cloudflare fencing off the open web. The BLM quietly narrowing who gets to weigh in on shared land. The whole day was about proximity and timing and who gets a seat at the table when the resource gets divided up.
If you're building, the takeaway isn't "AI is powerful," you already know that. The takeaway is that the winners are the ones close to the real work and close to the real moment — and the fights that'll shape your next three years are increasingly about access, permission, and who eats the cost. Less about the model. More about the ground it's standing on. Watch the ground.
That's the menu for today, kiddos. Go make something that solves a real problem for a real person, and if you can't do that, at least go make yourself a decent slushie — it's hot out there. I'm Tony DeLuca, this has been Barely Possible, and I'll see you next time.