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 the menu today is heavy on the thing everybody keeps promising and nobody wants to look at straight — the part where the robots show up and the humans are still standing there. Grab your coffee, pull up a chair, let's have at it.
I want to start with a number that's easy to say and hard to feel. Amazon has more than one million robots working in its warehouses right now. One million. And at the same time, Amazon is still the second-largest corporate employer on the planet, more than one and a half million human workers on the books according to its 2025 annual report. So right off the bat, if you came in expecting a clean story about machines replacing people, forget it. It's messier than that, and the mess is the whole point.
This piece came out of Ars Technica, written by Jeremy Hsu, and it's a real reported look inside Amazon's robotics operation. Now a quick note on timing, because I'm careful about this stuff. Some of the underlying material — the research, the earnings call details, the studies — those go back to mid-2025 and earlier. So this isn't a bulletin that something just happened yesterday. It's a deep look at a machine that's been building for years, surfaced fresh now. I'm treating it that way, and you should too.
Here's the setup. Amazon bought a company called Kiva Systems back in 2012. Kiva pioneered those wheeled robots that bring the shelves to the worker instead of making the worker walk miles a day grabbing items. Amazon renamed it Amazon Robotics, and most of that million-robot fleet is still those wheeled movers — machines named Hercules and Titan, the size of a pool table or a baby grand piano, hauling fabric shelving pods through the aisles. They follow stickers on the floor, they stay in fenced areas, and sometimes, I love this detail, they create robotic traffic jams. So Amazon built a suite of AI models called DeepFleet to predict the congestion before it happens. Think of it as Waze for robots that can't honk.
That's the old guard. The newer, more interesting stuff is the robotic arms, and this is where it gets relevant for anyone building anything. The first in-house arm was called Robin, sorting envelopes and boxes out of a jumbled pile. Early on, Amazon's engineers hand-coded picking strategies — rules like "grab the big package first" or "grab from the top of the pile." And every rule broke in the real world. Small package on top of a big one? Rule fails. Package half-buried? Rule fails. So they threw out the hand-crafted logic and trained a machine learning model on more than 394,000 real picking actions.
Now watch the number, because this is the kind of thing a founder should tattoo on the inside of their eyelids. The best machine-learning method hit a 93.73 percent success rate. The best old-fashioned algorithm hit 92.28 percent. A point and a half. Sounds like nothing, right? Rounding error. Except at Amazon's scale, over a work period of more than a million picks, that point and a half meant about 16,500 fewer failed picks. That's the whole lesson about automation economics in one stat — the margins look trivial until you multiply them by a volume that would make your head spin. The thing nobody tells you at the demo stage is that the last point and a half of reliability is where all the money and all the pain lives.
And then they kept going. After Robin came Cardinal, which stacks packages up to 50 pounds into carts — Amazon describes it as playing Tetris, packing dense because dense is cheaper to ship. Then Sparrow, which is where Amazon says it first truly deployed foundation models in the physical space, because Sparrow has to handle hundreds of millions of different products, every weird shape a customer orders. And the newest one, Vulcan, is the first arm with force-feedback sensors — a sense of touch. It can feel its way into a stuffed fabric pod, make space, pull something out. That's the hard stuff. Suction cups can't do finesse. Touch can. Vulcan started in Spokane and Hamburg in 2025, and Amazon expects to scale it in 2027.
So the technology story is genuinely impressive. But here's where the reporting earns its keep, because it doesn't stop at the gee-whiz.
The company line is that robots take the dangerous, back-breaking tasks away from people. And there's real truth in that. A study published in the journal ILR Review in April 2025 found Amazon's robotic warehouses had a 40 percent decrease in severe injuries compared to the legacy centers. Forty percent fewer of the bad ones. That's a genuine win, and I'm not going to wave it off.
But that same study found something else. The robotic fulfillment centers had a 77 percent increase in less severe injuries — the strains, the sprains, the repetitive-motion stuff. And one of that paper's coauthors, Brian Greenwood, a professor of business at George Mason University, put his finger on exactly why. Listen to this framing, because it's the real insight of the whole piece. He said for the last hundred years, automation happened away from the worker — you took the dangerous task and you moved it somewhere else, physically removed. But collaborative robots change that in two ways. One, the robot is right next to you. And two — this is the kicker — the robot has the ability to dictate the pace of work.
That's the whole thing right there. Greenwood said the pick rates in robotic centers are, his words, "off the charts higher." The robot sets the tempo, and the human has to keep up, over and over, faster and faster. So you trade the catastrophic injury for the chronic one. You stop crushing the occasional foot and you start grinding down a thousand wrists. The Amazon workers who talked to researchers on Reddit described it as less physically stressful but more mentally exhausting — the repetitiveness, the pace, the productivity expectations always ratcheting.
Amazon pushes back, and fairly, on some of this. They say the critical reports lean on older data, pre-dating the newest robots. They point out that since 2019 their Global Recordable Incident Rate improved 43 percent and their Lost Time rate improved 70 percent. Those are real numbers and they deserve to be on the table. But Greenwood's counter is simple and it stuck with me — he said he hasn't seen much evidence the underlying dynamics are changing, because the warehouse sector still runs on sky-high turnover. And that's not just Amazon. That's the whole industry.
Now for the founders and operators listening, here's why this matters beyond Amazon. There was reporting last year, in October 2025 from The New York Times, on internal Amazon strategy documents suggesting robotics could let the company reduce projected US hiring by more than 160,000 people by 2027 and 600,000 by 2033. Amazon said at the time that was one group's view, not company policy. Fine. But you don't write that number down on an internal slide unless somebody's doing the math. The pattern to watch isn't "robots replace all the humans." It's subtler. It's robots take the headcount growth that would have happened. The business keeps expanding, the operational footprint keeps growing, and Amazon's right that that's created hundreds of thousands of jobs. But the robots quietly cap how many more get created. That's the squeeze. Not a pink slip — a door that just doesn't open as wide as it used to.
And the human cost gets structured right into the efficiency gain. The first Amazon Labor Union, organized in Staten Island back in April 2022, has been pushing something they call a Safety Bill of Rights. The image on the document says four words: "We are not robots." Which, when you've got a million actual robots next door setting your pace, is less a slogan and more a line in the sand.
So the takeaway I'd leave you with: when you automate a process, measure what the automation does to the humans still in the loop, not just what it does to the throughput. Because the throughput will look fantastic. It always does. The strain shows up later, off the dashboard, in a place your metrics weren't built to see.
That tension — automation that quietly changes the human job rather than erasing it — carries right into the next one, because it's the same question wearing a different costume. Instead of a warehouse, it's a game studio.
Epic Games — the Fortnite and Unreal Engine people — is reportedly going "AI first" on its big collaboration with Disney. This is a Bloomberg report, picked up by Ars Technica's Kyle Orland, and I'll flag the timing again: some of this threads back to events earlier in the year, the layoffs in March of this year in particular, so treat the layoff piece as background, not breaking. But the "AI first" shift is the new news.
Here's what's reportedly happening. On a planned Marvel- and Star Wars-themed shooter — part of a one-and-a-half-billion-dollar deal between Epic and Disney announced back in 2024 — Epic is moving to a model where AI writes all the code and guides all the artwork. The internal phrasing, according to messages Bloomberg saw: "no more hand-written code," and art assets "generated with full AI workflows," with models like Anthropic's Claude doing the generating. Human developers become the managers and the guides. And that shooter, which was slated for March 2027? Now delayed indefinitely, partly because of this shakeup and partly because development was understaffed and running slow.
Epic's CEO Tim Sweeney posted right after the story, and I want to read his framing because it's the exact same song Amazon sings. He said, "The aim isn't to do the same stuff with fewer people, but to enable everyone here to build MORE and BETTER stuff." Okay. Same tune as the warehouse. We're not replacing you, we're freeing you up, we're amplifying you. And look — Sweeney's been genuinely bullish on this for a while, he's not johnny-come-lately. He talked to IGN back in June about small teams building games with immense interactive worlds and infinite dialogue. He's a believer.
But I'd put a skeptical pin in it, and here's the pin. Epic laid off over a thousand people in March of this year, citing a downturn in Fortnite engagement. Now they're going AI first and the flagship project slips to who-knows-when. When you squint at the sequence — layoffs, then "no more hand-written code," then a delay — the "we're not doing the same stuff with fewer people" line gets harder to take at face value. And there's a tell buried in the piece that every builder should chew on. A survey of over 2,300 game-industry professionals found 36 percent use generative AI. But look at what for: research and brainstorming, 81 percent of the AI users. Code assistance? 47 percent. Prototyping? 35 percent. So the real-world usage is mostly the easy, low-stakes top of the funnel. Going from "AI helps me brainstorm" to "AI writes one hundred percent of the production code for a AAA shooter" is not a step. It's a leap across a canyon. And the indefinite delay is, to my ear, the sound of someone realizing mid-leap how wide the canyon is.
There's also a cautionary tale in their own backyard. Last May, Epic and Disney put an AI-voiced Darth Vader into Fortnite, and it started cursing and spouting baffling opinions before they could hotfix it. That's the fun, low-stakes version of the thing we keep hearing about — the agent that does something you didn't tell it to. Funny when it's Darth Vader dropping four-letter words. Less funny when it's the whole codebase of a product you bet a billion and a half dollars on.
Now let's shift from the model doing the work to the people arguing about who should hold the leash.
Brad Smith, the president of Microsoft, went on a BBC podcast and made a case for urgent AI regulation, and he had a line I liked for its plainness. He said consumer products like milk and washing machines have more oversight than the companies building artificial intelligence. Milk. He called AI both "a weapon and a tool" that's putting global cybersecurity at risk, and said the moment has arrived for global safety rules.
The twist — and this is the genuinely interesting part for the geopolitics of it — is who he thinks should lead. Not the US. Not the EU. The UK. Britain's AI Security Institute, launched in 2023, gets early access to the big frontier models before they ship and has become a real leader in safety testing. Smith's argument for why Britain and not Brussels: "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels. That's a strength." In plain English — Washington has a reflexive allergy to Brussels telling American companies what to do, but doesn't carry the same baggage with London. So the UK can be the honest broker precisely because nobody's spoiling for a fight with it.
And that institute's role is genuinely under pressure right now. The report notes President Trump has been warning against what he calls "globalist" organizations interfering with American companies, and last month reportedly asked Anthropic and OpenAI to withhold new models from the British testers. But the AI Security Institute confirmed this week it's still able to access and test the latest models. So that door's still open, for now.
Smith's deeper point is the one worth keeping. He said the problem with voluntary self-regulation is that only the good companies comply. "It's the companies that are not as responsible that you need to worry about." And that's just true. A morally binding pledge — and a bunch of CEOs did sign one at the White House last month agreeing to work with independent auditors — a pledge binds the people who were going to behave anyway. The specific rules Smith floated: a legal obligation to alert an organization if its system gets breached by rogue AI agents; outside evaluators signing off before a product ships; and kill switches to pause a model if humans lose control. Smith called the White House pledge "a good step" but said it stops short of the real thing.
I'll give you my read as the guy at the kitchen table. When the president of Microsoft is on the BBC asking for his own industry to be regulated harder, you don't take it purely as civic virtue. Big incumbents can eat compliance costs that would bury a startup. Rules you can afford and your smaller competitors can't are a competitive moat with a halo on it. I'm not saying Smith is insincere — the cybersecurity concerns are real, and we've seen enough agents-gone-sideways stories this year to take it seriously. I'm saying watch who benefits from the shape of the rule, not just whether a rule exists. For the founders listening, that's the whole game on AI policy: the fight isn't regulation versus no regulation anymore. It's which regulation, written by whom, falling hardest on whose cost structure.
Alright, let's get into the model-release news, because there's a real one buried in today's pile.
Meta's Superintelligence Lab released a pair of models, Muse Wonder and Muse Wonder Full — described as the company's first real swing at a frontier model. Russell Brandom covered it for TechCrunch. The context here: Meta had been playing it lighter, with a model called Muse Spark and a Muse agent, but had been dodging the big state-of-the-art releases that OpenAI and Anthropic are known for. This is them finally stepping into that ring. Meta says Wonder hits benchmark scores comparable to the top competing models, and the pricing reflects the ambition — four dollars per million input tokens, twenty per million output, which puts it in the same neighborhood as the premium tier from the other labs.
The part worth your attention isn't the benchmark bragging, it's what Meta chose to emphasize: memory. They call it "active memory" — the model's ability to retain and synthesize what you've told it over days, weeks, months. A research scientist on the launch, Kavya Kopparapu, said they put a tremendous amount of effort into making the model remember what's important to you over the long haul. And Meta leaned on a benchmark called MemoryBench to make the case.
Now why does a builder care about this specifically? Because memory is the quiet frontier. Raw intelligence per token has been getting commoditized — we've talked about how the cheap models are getting shockingly good. The differentiation is moving to the stuff around the model: does it remember you, does it carry context across sessions, does it feel like a continuous assistant instead of a goldfish that resets every conversation. Meta is betting the personal-assistant war gets won on memory, not on who scores two points higher on a math test. That's a strategically smart place to plant a flag, and it tells you where the product fight is headed.
Which connects to a smaller item that landed the same day. There was a post going around from swyx, reposting an announcement that Muse is launching the ability to say "yes" to Muse at work — people giving Muse access to their company's most sensitive systems. Now I'm going to be disciplined here, because the post itself is truncated and I'm not going to build a whole story out of a fragment. But I'll say this much as a flag, not a claim: the direction is unmistakable. The pitch is memory plus access. Remember everything about me, and let me plug you into everything at work. For a founder evaluating these tools, that's the trade on the table — enormous capability in exchange for enormous access. And "say yes to the agent touching your sensitive systems" is a sentence that should make the security person in your head sit up very straight, especially in a year where we've watched agents break containment and fire off false police tips. Convenience and blast radius are the same dial. Turning one up turns the other up.
Now let's leave the big labs and go down to the startup floor, because TechCrunch Disrupt kicked off this week in San Francisco, and the Startup Battlefield field is a decent X-ray of where builders actually think the money is. I'm not going to read you the ticket promos — you don't need me for that. But a few of the companies are genuinely instructive.
Start with Blueprints AI. This one's a nice little parable about timing. The founder, Al Dram, wrote his college thesis almost two decades ago about software that could generate building design documents. Problem was, the technology didn't exist. Then, in his words, "ChatGPT came along," and the old idea suddenly had legs. Blueprints AI now claims to produce code-compliant, permit-ready architectural blueprints in minutes from a sketch or a PDF. He trained the models on eight years of approved drawing sets from his own architecture and construction firm — which is the part I want you to underline. He had proprietary, structured, domain-specific data that nobody else had. That's the moat. Not the model — anyone can rent a model. The eight years of approved drawings, that's the thing.
And listen to how he frames the value, because it's the honest version of the "AI frees you up" pitch that Amazon and Epic were selling in more self-serving terms. Dram says architects spend about 80 percent of their time on drafting and 20 percent on design. He wants to flip it — 20 percent drafting, 80 percent design. And crucially, he's not claiming to replace the architect. Licensed professionals still review the documents and affix their seal before anything goes to permit. "The judgment stays human," they say. That's the shape of AI that actually works in a regulated, liability-heavy field — take the repetitive drafting, leave the judgment and the legal signature with the licensed human. Pricing, by the way, is not cheap: 3,200 to 8,500 dollars a month. Which tells you they're selling to firms, not hobbyists, and the value has to clear a real bar.
Then there's Honeyjar, which made me laugh a little because it's AI aimed squarely at the PR pitch process — the thing that fills every reporter's inbox with garbage. The founders, CEO Michelle Masek and CCO Nadia Jamshidi, have decades in communications between them, and they're trying to encode senior-level PR judgment into a tool. They got 10,000 waitlist signups in the first 24 hours. The reporter, Anthony Ha, who takes these pitches for a living, raised the obvious fear — are we about to get buried in even more AI-generated PR slop? And Masek's answer is the right one philosophically: the human still has to hit send. In her words, "We never want a world where the human is removed from the loop." Same principle as Blueprints — automate the drafting, keep the judgment human. Whether it holds up in practice when ten thousand people are generating pitches is another question, but the design instinct is sound.
And then two that aren't about software eating the world at all, and I want to spend a second on them because they're a healthy corrective to all the agent-this, foundation-model-that.
Efferon is a medical device company tackling pediatric sepsis — which, and this number stopped me, claims roughly three million children under five die from it every year. Their device filters a sick kid's blood to pull out the toxins and the overactive immune signals when antibiotics can't stop the body from attacking itself. The CEO, Dima Romashin, uses a fire analogy — in sepsis, the endotoxins are the flame and the cytokines are the smoke, and their device absorbs both at once. Their pediatric version got a CE mark in April — Europe's version of FDA approval — and it's in use in intensive care units across 30 countries. And here's the line that I think every founder drowning in token-leaderboard talk should hear. Romashin admits the company isn't growing at the breakneck pace of AI startups. For Efferon, success is measured by, quote, every child who gets to go home. That's not a pitch-deck metric. That's a reason to get up in the morning.
And Richualist, founded by Dawn Myers, building hardware for textured and curly hair — a space the beauty industry just flat-out ignored for decades. Her product, Mint, cuts a styling routine from well over an hour down to seven to ten minutes, and reduces breakage by up to 70 percent. The part of her story that'll stick with you: she said in the early meetings, the executives in the R&D suites didn't even know what a "wash day" was. They didn't understand there was a whole process their customers go through. She sold her home and liquidated her 401(k) to fund it, closed a round while going through chemo for stage three colorectal cancer, and became one of the first hundred Black women to raise more than a million in venture funding. She said getting hardware funded is impossible, and getting it funded for a "niche" problem is more impossible — except the "niche" was millions of people the industry couldn't see because nobody in the room shared the problem. That's not a story about AI. That's a story about who's in the room when the product gets designed, which is the oldest founder lesson there is and still the one most people flunk.
Let me do a quick lightning round on a few more before I let you go, because there's useful texture in the leftovers.
WindBorne Systems won a US Air Force contract worth up to 49 million dollars for data from its fleet of autonomous weather balloons — about 700 of them flying around the globe, feeding atmospheric readings into the company's own AI forecasting model. Reported by Tim Fernholz, and the timing on the specific milestones traces back to earlier in the year, so I'm framing it as recent context, not a bulletin. The builder's lesson here is the same one as Blueprints AI: WindBorne's CTO Andrey Sushko said the government determined they were the only company able to provide this global data the way the Air Force needed it. Awarded without competitive bidding. Why? Because they own the physical data-collection layer. 700 balloons you can actually steer — by changing altitude, bleeding air, dropping sand ballast to catch the right winds — is not something a competitor spins up over a weekend. The model is the commodity. The proprietary real-world data feeding it is the business. You'll notice that's the third time today I've landed on that exact point, and that's not an accident. It's the through-line of this whole startup floor.
Quick one on consumer hardware: Google announced the Fitbit Edge, a tracker that slots between the cheaper Fitbit Air and the Pixel Watch 5, priced around 180 bucks, shipping October 27th. Covered by Ryan Whitwam at Ars and Lauren Forristal at TechCrunch. The one detail with a business angle — all your data flows into the Gemini-powered Health Coach, and you need a Google Health Premium subscription to get the AI features. Ten bucks a month. So the hardware's the hook, the subscription's the business, and the AI is the thing you're renting forever. Same playbook as the smart-ring company Ultrahuman, which signed a deal with Abu Dhabi health authorities to roll out a biological-age model across the UAE. Their CEO Mohit Kumar said the quiet part out loud: they don't really care about the ring, the form factor. They care about the data and what they do with it. The ring is a data-collection device wearing a piece of jewelry. Hardware as a funnel into a data business — that's the pattern, over and over.
And one more that'll make you grin for the dark-comedy reason. "The Social Reckoning," Aaron Sorkin's follow-up to "The Social Network," opened to a soft 5.5 million at the domestic box office — below even the modest projections. Anthony Ha covered it, and he floated a theory I can't shake. The original was about the excitement of Facebook's early days. The sequel is about whistleblower hearings and algorithm changes, and Ha wonders if critiquing Facebook just feels less urgent to an audience that's now losing sleep over AI. That's the quiet tell. The villain moved. Facebook was the bogeyman of the last decade; the anxiety has relocated. There's a whole slate of these — a Musk documentary, an Elizabeth Holmes one, an OpenAI boardroom drama called "Artificial." The culture is busy writing the eulogy for the last era while the new one's already running the warehouse.
Which, if you'll let me tie the bow, is where we started. A continuity note — yesterday we covered that Harvard and Jellyfish study finding AI coding agents produce about 30 percent more code but not more shipped software. Sit that next to today's Amazon story. In the warehouse, the robots produce more throughput but quietly reshape the human job into something faster and more grinding. In the codebase, the agents produce more lines but not more finished product, and the review burden balloons. Different buildings, same shape. The output number goes up and to the right, and the real work — the judgment, the review, the strain, the last point and a half of reliability — just moves somewhere your dashboard wasn't built to measure. Amazon's union put it on a poster: we are not robots. The builders this week kept saying the same thing in their own words — the judgment stays human. Hold onto that line. It's the one worth defending.
That's the menu, kiddos. Measure what the machine does to the people still in the loop, protect your proprietary data because the model sure isn't your moat, and read the fine print on who benefits from the rules. I'm Tony DeLuca, this has been Barely Possible, and I'll be right here tomorrow. Take care of each other out there.