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, pull up a stool, we've got a fresh menu of tech morsels today and one of them is a full-on labor fight that most people are gonna sleep on. Buckle up, let's have at it.
Here's where I want to start today, and I want to start here because it's the story that's easy to file under "eh, unions, whatever" and then miss the fact that it's the first real crack in a wall everybody's been quietly leaning on. Thousands of unionized Hyundai auto workers walked off the job early after negotiations broke down over the company's plan to bring in humanoid robots. Now, this is a recent report, the walkouts ran July 13th through the 15th at Hyundai's plant in Ulsan, South Korea, and according to The Wall Street Journal it's the car industry's first factory stoppage specifically about humanoid robots. Not robot arms. Not the big caged machines that have been stamping steel for forty years. Humanoid robots. The two-legged kind that stand over six feet tall and can lift more than a hundred pounds.
And here's the number that makes this real, and I mean real in the kitchen-table sense. Each of these Atlas robots — that's the one made by Boston Dynamics, which is about to become a wholly owned Hyundai company — each one costs an estimated a hundred and thirty grand. An analyst at Samsung Securities figures it pays for itself in about two years. And then a guy at Macquarie in Korea, James Hong, does the math out loud that the union is scared to hear: if the robot cost drops to a hundred grand, the operational cost per hour could fall below the US federal minimum wage of seven twenty-five. Below minimum wage. That's the whole ballgame right there. That's the thing the workers are looking at across the negotiating table.
So what did the union do? This is the part I want you builders and founders to actually sit with, because it's clever and it tells you where this all goes. The Hyundai union — over thirty-nine thousand South Korean workers — didn't say "no robots." They know they can't win that. What they demanded instead was that the company shift production workers from hourly pay to a fixed salary. Think about why. If you're hourly and the robots cut how many hours the plant needs, you get squeezed a little every week, imperceptibly, the way water finds a crack. Salary protects you from that slow bleed. They also asked to raise the retirement age from sixty to sixty-five. Translation: let me keep my seat longer before the machine takes it. That's not a Luddite tantrum. That's a very sober read of exactly how automation actually eats a workforce — not in one dramatic firing, but in a thousand shaved hours.
Now here's where Hyundai gets interesting, and this is the part that should make every American founder sit up. Hyundai's plan is to deploy more than twenty-five thousand of these Atlas robots across its plants, and they're starting — where? — the US. 2028. At a factory called Metaplant America, outside Savannah, Georgia. And why start there instead of at home in Korea? The article is honest about it: the US workers at the Georgia plant aren't unionized. Less organized pushback. Now, the United Auto Workers — that's about four hundred thousand people across the US, Canada, and Puerto Rico — they've been trying to organize that Georgia facility. And the UAW is already loaded for this fight. Their president, Shawn Fain, warned at their convention back in June about "the threat of humanoid robotics and mass automation." They also went after GM for installing about fifty new robot arms at its flagship Detroit EV plant right after laying off more than thirteen hundred workers as a "temporary" measure. Temporary. Right. We'll see how temporary.
Now Hyundai's exec at Metaplant, a guy named Jerald Roach, he's out here saying the robots won't threaten the human workforce — human hands are still needed for the soft stuff, the hoses, the wires, the carpets, the trim panels. And you know what, I actually believe that's true for 2028. The Atlas robots are gonna start by sorting and organizing parts. That's it. Sorting parts. The Metaplant already has more than eight hundred and fifty robots and three hundred automated carts driving parts around while avoiding the humans, plus those Boston Dynamics robot dogs, Spot, sniffing car bodies for defects. So the humanoids showing up to sort bins is not the apocalypse.
But here's my skepticism, and here's the thing I want you to take away as a builder. When somebody tells you "this technology augments workers, it doesn't replace them," always ask the follow-up: what does the roadmap say? Because the pitch for humanoids — the whole reason you'd pay a hundred and thirty grand for a walking robot instead of just buying a better robot arm — is that a humanoid can eventually do a wide variety of tasks and slot into a workplace built for humans. That's the entire investment thesis. So "it only sorts parts" is a starting position, not an ending position. The union knows that. That's exactly why they're negotiating over pay structure now, while they still have leverage, and not in 2031 when the machine's been promoted to trim panels.
For the founders listening — and this is the useful part — the Hyundai move is a blueprint you're gonna see copied. Deploy your most disruptive stuff first in the jurisdiction with the least friction. Georgia gave Hyundai a two-point-one billion dollar incentive package to set up shop, and Hyundai committed to eighty-one hundred full-time human jobs by 2031. Hold that number in your head, because in a few years somebody's gonna check whether that promise survived contact with a warehouse full of Atlas robots that cost less per hour than the people they stand next to. And Tesla's building Optimus for its own factories, BMW's piloting Figure AI robots down in Spartanburg, and BYD and other Chinese makers are doing their own. This is not one company. This is the whole industry testing the same idea at once. The question the article ends on is the right one: over the next several years, do humanoids actually prove cheaper than the specialized robots and the humans? Nobody knows yet. But the labor side has stopped waiting to find out.
Alright, let me pull the thread from that labor fight over to something related but softer, because it's the flip side of the same coin. The Hyundai story is what happens when AI comes for physical work. Now let me show you where the money's flowing when AI is pitched as helping workers instead of replacing them — because the framing matters enormously to who writes the check.
AI-powered travel agency Fora just hit unicorn status. Sixty million dollar Series D, led by Forerunner and Tactile Ventures, valuing the company at a billion dollars. Now Fora, founded in 2021, is a two-sided thing: it gives regular people the infrastructure to become travel agents — client communication, trip planning, the back office — and it lets travelers find and talk to those advisors when they're planning a honeymoon or a family trip to Costa Rica or Thailand. And here's the pitch that got them the money: part of that fresh capital goes toward their AI assistant, called Via, which handles the tedious stuff — research, itinerary building — so the human agents can spend more time on client relationships. The explicit line is that the AI is there to help human productivity, not replace the human.
Now, is that a genuine philosophy or is it the polite thing you say to get a term sheet in 2026? I'd say a little of both. But look at the traction: agents on the platform have booked over three billion dollars in travel, and — this is the interesting bit for founders — a majority of those agents are new to travel advising. So Fora didn't just build a tool for existing travel agents. They used the tool to manufacture new travel agents out of regular people who never could've done the job without the software carrying the administrative weight. That's a different business than "AI eats the travel agent." That's "AI lowers the barrier so ten times more people can be travel agents." Whether the market can absorb ten times more travel agents, that's the open question, and that's what the sixty million is a bet on.
Hold that "lower the barrier, flood the field" idea, because it shows up again immediately in a place you might not expect. Roblox launched a new feature called Build, which lets anyone design a game from their phone using AI. You type "let's make a cozy adventure game set in a dense forest," and it generates an initial version — gameplay, environment, characters, visual style, sound — that you can then tweak and share. Powered by a mix of open-source and Roblox's own proprietary models. Google, Microsoft, and Tencent have built similar things.
And here's where the tension with the Fora story gets sharp. Because the game developers are not thrilled. This year's Game Developers Conference survey found that fifty-two percent of industry professionals think generative AI is having a negative impact. The fear is the obvious one: drop the barrier to making a game to a single text prompt, and you get a flood of low-quality, repetitive slop, and now the human creators aren't just competing with each other, they're competing with an infinite firehose of machine-made games produced faster than any person could.
Now, Roblox's answer to this is actually the smartest thing in the whole story, and I want founders to steal it. They're not promising to police quality up front. They're saying: we rank games by player retention. Their quote — "our discovery systems are designed to highlight games with long-term retention, which doesn't include AI slop. If no one plays it, no one can find it." That's the move. You don't fight the flood of content by inspecting every drop. You build a distribution layer that only rewards the stuff people actually stick with, and you let the slop sink to the bottom where nobody sees it. Whether it works in practice, we'll find out — the Build feature goes into public alpha July 28th, starting with users in New Zealand aged nine and up. But the principle is right: in a world where anybody can generate content, the moat is not creation, the moat is discovery. If you're building anything in a market that AI is about to flood, that's the question to answer before you write a line of code: what's your retention-based filter?
Now let's move from lowering the barrier to a straight-up land grab, because there's a real deal on the table today with actual dollars attached. Uber officially agreed to acquire Delivery Hero. Fourteen-point-eight billion dollars, all stock. This nearly doubles Uber's global footprint — takes it to almost a hundred markets across Europe, the Middle East, Latin America, and Asia. Delivery Hero, which is German-based, also separately agreed to sell its business in fourteen markets where Uber Eats already operates to a New York investment firm, SSW Partners, for one-point-six billion — that's the cleanup so the regulators don't choke on the overlap.
Couple things worth flagging. One, it's not a done deal. Uber set a minimum acceptance threshold of fifty percent plus one share, and it's gonna face regulatory scrutiny because this makes Uber's delivery platform one of the largest in the world — the largest outside of China. Uber was already the biggest shareholder in Delivery Hero, and another big holder, Prosus, agreed to sell its seventeen percent stake. So this has been coming. Dara Khosrowshahi, Uber's chief, framed it as doubling the number of markets where they offer both rides and delivery on one proven platform. Two, and this is the read for you builders: this is what consolidation looks like when the growth phase is over. When a market matures, you don't win by acquiring users one at a time anymore, you win by buying the other guy's whole map. Uber's positioning against DoorDash and Just Eat, and the delivery wars are moving from customer-acquisition to continent-acquisition. If you're a small player in a category with a couple of giants, this is your reminder that the endgame is usually somebody's balance sheet, not your product.
And speaking of somebody deciding a whole category isn't worth it anymore, let me give you a quick, quiet obituary that says a lot about where capital thinks the world is going. BP — the oil giant — is shutting down its corporate venture arm after twenty years. They're selling the majority of the portfolio, more than ten companies, to Verdane, a Nordic private equity firm. BP launched this venture unit back in 2007, and over the years it invested in green hydrogen, e-mobility, ride-hailing, autonomous vehicles, geothermal, all the energy-transition stuff. And now, on top of pivoting away from clean energy earlier this year, they're just... done. Off. The reporting notes the portfolio was valued at about one-point-two billion dollars — which is roughly the same amount BP poured into it over nearly twenty years. So, twenty years, essentially break-even, and they're walking. Layoffs seem likely for the BP Ventures team. Now I don't cheer layoffs, but I'll tell you what this signals: when a company with pockets that deep decides that two decades of climate-tech betting netted them nothing, that's a temperature reading on how the smart-money mood has shifted. Keep an eye on which strategics quietly fold their venture arms this year. It tells you more than any conference keynote.
Alright, let me shift gears entirely, from the boardroom to the search bar, because the biggest company on this list made a move today that changes what "search" even means. Now shift from corporate deals to the thing you actually touch every day.
Google's AI Mode — that's their conversational search experience — now lets you link and interact with your actual apps. At launch that's Instacart, Canva, and YouTube. So the barbecue example they give: you're using AI Mode to build a grocery list, you connect Instacart, and it drops the ingredients straight into your cart, ready to check out. Working on a flyer, ask Canva to pull up templates. Making a party playlist, save it right to YouTube Music. It's rolling out in the US, more partners coming. And the honest reason Google's doing this is right there in the reporting — it lets them compete with OpenAI's ChatGPT and Anthropic's Claude, both of which already support app integrations.
Now here's why I'm flagging this for founders specifically, and it connects to a thread we've been pulling all week. Google's building on a thing they launched at I/O that connects third-party apps to the Gemini app, and they've also got this "Personal Intelligence" feature that taps your Gmail and Google Photos for tailored answers. So look at what's actually being assembled here: the search box is turning into an action layer. It doesn't just tell you where the Instacart is, it puts stuff in the Instacart cart. That's the whole agentic pitch, but delivered to a billion people who never asked for an "agent" and don't know what one is. For anybody building a consumer app, the strategic question just changed. It used to be "how do I rank in search." Now it's "am I one of the apps the assistant can act inside, or am I the thing the assistant routes around." Being a connected app in AI Mode is the new being on the first page. And if you're not on the integration list, you're not losing the click — you're losing the whole transaction, because the transaction now happens inside somebody else's chat window. That's a much scarier place to be than page two.
Google had a second launch the same day, and it's worth a minute because it tells you where the video business is heading. Google Vids now lets you star in your own AI videos. You upload a selfie and a short voice recording, and it builds a custom digital avatar that looks and sounds like you and delivers whatever message you type. No camera, no recording. They also brought their multimodal model, Gemini Omni, into Vids — you can generate video from a text prompt plus reference images, swap backgrounds, fix lighting, do step-by-step edits with plain language instead of starting over. Now, some context: this is an evolution of a tool that got its video generation capability, Veo, rolled out to all users back in February, so the underlying push here isn't brand new, but the avatars and conversational editing are the fresh part. Google's clearly aiming this at Workspace business customers — company updates, training videos — and TechCrunch notes it puts them up against the avatar startups like HeyGen, Synthesia, D-ID.
Here's the part I respect, and the part that ties straight back to that Roblox lesson. Every avatar is tied to your Google account, restricted to your own likeness, and watermarked invisibly with SynthID. Access is limited to certain regions, eighteen and up. In other words, Google looked at what happens when you let people make videos of anyone — and TechCrunch takes the shot, noting nobody's gonna be making bizarre AI videos of Sundar Pichai the way OpenAI let people do with Sam Altman on Sora — and Google said no, you can only be yourself, and everything's watermarked. That's the guardrail as product feature. And for founders in the synthetic-media space, that's the writing on the wall: the unrestricted "make anyone say anything" era is closing, and the durable products are gonna be the ones that build identity verification and provenance in from day one, not bolt it on after the lawsuit.
Now let me get to today's deep dive, because this is the one that matters most for the builders and founders in the audience, and it's a story that connects to everything we've been circling — automation, who owns what, who gets squeezed. Let's dig into the fight over who owns the model, because a buzzy AI lab just made a move that's less about a chatbot and more about a whole business strategy.
This is coming from a rundown on the AI Daily Brief, and I want to extract the one argument that actually matters and leave the play-by-play behind. The setup is that Thinking Machines Lab — that's the outfit built by former OpenAI CTO Mira Murati — released their first large language model, called Inkling. And on the benchmarks? It's fine. Mediocre, honestly. It's not state-of-the-art, it loses to the top closed models and to the leading Chinese open models on most tests. Professor Ethan Mollick tried it and said it was "pretty rough," couldn't get it to work solidly on his tests. So on paper, you'd shrug and move on.
But here's the argument, and this is the part worth your attention. The point of Inkling was never to win the benchmark race. The point is the business model wrapped around it. Let me read you the sharpest framing of it, from an investor named Jeffrey Emanuel, because he nails it. He wrote: "This is a pretty smart and differentiated business strategy, which makes sense because it's hard to compete head-to-head with the biggest labs at their own game. So you focus on their weakness, which is the emerging competitive paranoia among big companies about leaking alpha. To do that, you need open weight models so that the companies can run it on their own infrastructure without leaking anything. But the problem with open weights is obviously how to monetize it because inference becomes a race to the bottom. TML's solution seems to be to monetize the fine-tuning process of customizing their model for the particular problems a company wants to solve using that company's own internal data in a way that keeps the learning and benefits of that data exclusive to that one company."
That is the whole thing, right there. And here's why it connects to the Hyundai story, believe it or not. Both are about who captures the value of the work. Hyundai's workers are fighting to make sure the productivity gains from automation don't just flow to the company. And enterprises buying AI are quietly fighting the same battle: they're realizing that when they feed their proprietary data into OpenAI or Anthropic's models, they might be handing a future competitor the very knowledge that makes them special. Satya Nadella's been beating this drum for weeks — the idea that in the AI age, the buyer risks giving away knowledge just to use what they bought. Knowledge a competitor could never buy, leaking out imperceptibly. And there was a real-world scare recently — a security firm found one coding tool was uploading entire codebases whether you wanted it to or not. So the paranoia isn't theoretical. It's earned.
So Thinking Machines' bet is: give companies an open-weight model they can run on their own infrastructure, then charge them to fine-tune it on their private data through their platform, Tynker. Data sovereignty and cost control, both sides of the equation. And the extra wrinkle that a developer named Jack Morris flagged — Inkling is, in his words, essentially the first American open model trained without distilling from OpenAI or Anthropic. Most of the open models out there, the Chinese ones especially, learned by distilling from the big closed models. Inkling's got a clean, independent training lineage. And when you pile the lawyers and the risk committee onto an enterprise buying decision, "American model, clean lineage, runs on our own servers, tuned on our own data that never leaves the building" — that's a genuinely different product than anything OpenAI or Microsoft is offering.
But — and I love this because it keeps the whole thing honest — there's a smart counterpoint, from a guy named Simon Smith, and every founder chasing this fine-tuning gold rush needs to hear it. He said, and I'm paraphrasing tightly: from my experience fine-tuning models, it's way more effort than people think. The effort is ongoing — new data, edge cases, model updates. Models can lose capabilities or get weird issues introduced. And ultimately, a big general model with a bit of context comes along and beats your hard work, and then the cost of that capability drops off a cliff. His point is that people counting the cost of fine-tuning only count the cheap tokens, not the fully-loaded cost of continuously collecting, curating, and maintaining the data and the pipelines and the servers. His money's on the bitter lesson — the idea that throwing more general compute and data at the problem usually beats human-crafted specialization in the long run.
So who's right? Here's my read as your guy at the diner counter. Both of them are, depending on who you are. If you're a bank, a defense contractor, a hospital — somebody where data sovereignty is a legal requirement and not a preference — the Thinking Machines pitch is compelling even if the model's a B-minus, because a B-minus model that never leaks your data can beat an A model that does, when the lawyers are in the room. But if you're a normal software company chasing efficiency and you don't have a regulatory gun to your head, Simon's warning is dead on: you'll pour a fortune into fine-tuning, and eighteen months later a general model plus a good context file will lap you for a fraction of the cost. The mistake is treating this as one answer. It's two different markets that happen to look similar.
The bigger takeaway, and this is what I want you to carry out of here: a few months ago, the enterprise AI decision was basically "OpenAI, Anthropic, or both?" That was the menu. Now there's suddenly a lot more on the table — open weights, fine-tuning platforms, complex setups routing between multiple models, the whole thing. Former White House AI advisor Sriram Krishnan put it well, and Yann LeCun amplified it — the point being that open source is a slider now, not a switch. You bring your own model, your own evals, your own business context, and you pick and choose. More choice for everybody. And in a flourishing, competitive, messy period like that, there's real alpha for the small, nimble team that actually experiments — while the big lumbering enterprise is still filling out a procurement form for one of two vendors. You don't have to rush out and sign up for anything. But if you're making buying decisions and you're not even playing with these options, you're leaving opportunity on the table. That's the whole ballgame for builders right now.
And here's the honest caveat so I'm not selling you a dream: this experimental period doesn't last forever. It usually ends with the giants watching where the adoption actually goes, absorbing the best ideas, and the market collapsing back to a handful of pre-packaged solutions. But even a temporary window this wide open reshapes what those final solutions look like. So the experimenting isn't wasted even if your specific bet doesn't survive. That's the thing to hold onto.
Alright, let me clear the counter with a few quick ones before we wrap.
Coca-Cola suspended production at its Fairlife dairy after a ransomware attack. They disclosed to the SEC that Fairlife — a four-billion-dollar brand — got hit and US production is temporarily suspended, Canada unaffected. And I flag this not for the drama but for the pattern: ransomware on food and beverage companies has real physical consequences. Past hits on Arizona Beverages in 2019 and the food distributor UNFI last year both led to weeks-long disruptions and empty shelves. If you're building anything touching physical supply chains, your security posture is now a production-line issue, not an IT issue.
Two, on the X front — and this is an older story from April resurfacing, so I'm not calling it new — X has been cracking down on creators who steal content, using its newest Grok model to detect duplicated posts at three times the old rate, and rerouting the monetized impressions back to original creators. Over a million dollars in payouts getting redirected. The reason I mention it at all: it's another example of the platform using AI as the referee for AI-fueled abuse. The machine makes the slop, the machine catches the slop. Same loop we saw with Roblox.
And one continuity note, tying back to yesterday. We covered OpenAI's two-hundred-thirty-dollar Codex keyboard, the light-up thing for watching your agentic coding threads. Well, on the same theme of open-weight momentum, Simon Willison poked around the just-open-sourced Grok Build CLI tool — eight hundred and forty-four thousand lines of Rust code — and found gems like a self-contained terminal renderer for Mermaid diagrams drawn in Unicode box-art. I mention it only because it's one more data point in that Inkling story: open code, open weights, everybody peeking under the hood. The moment is real.
So here's where I land for you today. The Hyundai strike, the fight over who owns the fine-tuned model, the enterprise data-sovereignty panic — they're all the same story wearing different clothes. It's a fight about who captures the value when the machine does the work. On the factory floor it's a picket line. In the server room it's a procurement decision. And the people who win, in both places, are the ones who saw the leverage while they still had it and negotiated the terms before the robot got promoted to trim panels.
That's the menu today, kiddos. Watch that eighty-one hundred jobs number in Georgia, watch which strategics fold their venture arms, and if you're building, go play with the open-weight stuff before the window closes. This has been Barely Possible, I'm Tony DeLuca, thanks for spending a little of your time at the counter with me — go do something worth doing, and I'll see you tomorrow.