A guy puts on a lapel pin, taps his chest, asks it a question, and it tells him it's too hot to answer right now.
That happened on stage.
That happened on a stage, in front of people, and it was supposed to be the future of computing.
Okay, so we're doing this one properly.
Daniel wants a reckoning. Five of the biggest flops, disappointments and spectacular misfires in AI over roughly the last five years, and he's specific about what counts.
Which is good, because "flop" is a word people use far too loosely.
He's not interested in things that quietly failed to sell. He's interested in the gap. The distance between what we were told was coming and what actually showed up. And he splits it three ways. Models that were hyped to the moon and landed with a shrug. Products that promised to change how we talk to computers and turned out to be impractical. And features that companies rolled out with total confidence and then discovered, rapidly, that users wanted nothing to do with them.
So across the major labs, the startups, consumer hardware, and the open side of things.
That's the brief. And at the end of it we have to crown one. The single biggest flop of the modern boom.
I've been dreading that part for about a week.
You've been drafting it.
I've been drafting it. Let's start with the most spectacular one, because it's also the most instructive — a seven-hundred-dollar pin that was supposed to make your phone disappear.
The central question here is what separates a flop from a disappointment. A product that sells badly is just a bad business. A flop is a gap — expectation on one side, reality on the other, and the wider the gap, the better the case.
So it's not money.
Money's evidence, not the test. Quibi burned a fortune and it's on the list of legendary failures because it promised to reinvent how we watch things and then nobody watched anything. Same shape.
Five cases, and what's elegant about Daniel's framing is that each one fails differently. Humane is a hardware and engineering failure at a startup. Rabbit is the same failure mode at a tenth the price. GPT-4.5 is a flagship model that underwhelmed. Sora is a consumer product that worked and still got killed. Gemini's image generator is a feature that provoked backlash within about a week of launching.
And to be fair to the hardware, there's a clean framing for that thread. Tony Fadell, who built the iPod and the Nest thermostat — he spoke at MIT Future Fest this month and grouped the Rabbit R1, the Humane AI Pin, and the Limitless pendant into one cohort. Gen one AI gadgets. He said the founders came to him for help, and his verdict was blunt: you have to really understand what you're trying to do and what pain you're trying to solve.
Gen one. That's generous, actually. Gen one implies a gen two.
Fadell thinks there is one. But that cohort is the thread to pull first. Hardware, then the model, then the consumer product that got killed, then the feature that blew up in Google's face.
Let's start with the hardware. Two devices that promised to replace your phone and instead became cautionary tales.
The Humane AI Pin. Announced 2023, and the pre-launch campaign was extraordinary. The founder did a TED talk called The Disappearing Computer. They put the thing on a runway at Paris Fashion Week. Time magazine gave it a Best Inventions nod before it had shipped to a single customer.
Before it shipped.
And the pitch was that your phone is the problem — the screen, the apps, the constant looking down — and this pin, clipped to your chest, would give you an ambient computer. You tap it, you talk, a laser projects text onto your palm, and you never touch a phone again.
Seven hundred dollars.
Seven hundred dollars, plus a mandatory twenty-four dollars a month.
Mandatory is the word that should have ended it right there.
Reviewers got it in April 2024 and the response was almost unanimous, and not in a mild way. Marques Brownlee called it the worst product he thinks he has ever reviewed. The Verge's David Pierce gave it four out of ten, and then later said that in retrospect he had been too generous.
Too generous at four.
It shipped without a timer. Without an alarm. Features that were in the TED talk were not in the box. It would overheat and shut itself down in moderate use — and remember, that's on a device that has no screen and does almost nothing, so there's no obvious place for the heat to come from.
There's the phrase for the whole category. A computer that overheats doing nothing.
Humane ended up recalling the charging case over fire risk. Returns outpaced sales between May and August of 2024 — roughly a million dollars in returns against roughly nine million in sales. Then in February 2025 they shut it down and sold the technology to HP for a hundred and sixteen million dollars. They had raised two hundred and thirty million since 2018.
And the customers?
No refunds outside the ninety-day window. The pins became paperweights. People had paid seven hundred dollars and a monthly subscription for a device that got switched off from the other end.
David Pierce's line about it was that Humane will go down in the annals of tech history as one of the most spectacular gadget failures of all time. Up there with Quibi, the Fire Phone and the Apple Newton.
That's the bar. He put it next to the Newton.
Rabbit. Same failure, one tenth the price and one tenth the dignity.
The Rabbit R1. A hundred and ninety-nine dollars, launched 2024, and the pitch was a Large Action Model.
Which is a language model that doesn't just answer you, it operates apps on your behalf.
That's the claim. You ask it to order you something and it goes and does it in the app, the way a person would. And the reviews were brutal in a way that has its own comedy. Engadget headlined theirs "a hundred and ninety-nine dollar AI toy that fails at almost everything." Digital Trends called it a buggy, flawed and unsuccessful mess. Tom's Guide tested the Uber and DoorDash integrations and found they don't work well, or at all. Mashable's review ran under the line "I can't believe this bunny took my money."
That's a review and a confession.
The Verge's review found the vision model doing something I think about a lot. They pointed the R1 at a red dog toy — a rubber toy shaped like a bone — and asked it what it was looking at. It said stress ball. So they asked again. It said tomato.
A tomato.
Then it changed its mind and said bell pepper, and added that it was safe to eat.
It volunteered edibility.
So now you have a device that can't identify a dog toy, is confidently wrong about it twice, and then tells you to eat it.
Which is the whole problem in one exchange — the failure isn't that it doesn't know. It's that it doesn't know that it doesn't know.
And two devices, four hundred and thirty million in raised capital between the companies, and the same thing went wrong. Which is why I don't think the interesting question is whether the AI was good enough. It's what they thought people wanted.
Say more, because the obvious reading is that the models just weren't ready.
The models weren't ready, certainly. But picture a perfect version of the Humane Pin. Imagine it never overheats, the laser always reads, the latency is instant, every feature works. It still costs seven hundred dollars plus twenty-four a month to do less than the phone already in your pocket, which you are still carrying, because the pin can't do maps or photos or messages properly.
So the ceiling wasn't the technology.
The ceiling was the premise. The Verge's read on Humane was that they saw the big picture mostly correctly — screens are a problem, ambient computing is probably where this goes — and then utterly botched the details, and in doing so may have set the whole AI gadget revolution back a couple of years.
A company that got the destination right and the vehicle catastrophically wrong.
Fadell's version of it is that they never found the pain. Nobody was lying awake at night wishing their phone did less.
I want to defend one thing about the hardware, and then I'll drop it. A one-star product that costs two hundred dollars and a four-star product that costs two thousand sit in completely different moral universes. The R1 was a bad purchase. The Pin was a bad purchase with a subscription and no refunds.
That's a fair distinction and I'll take it.
Those were hardware failures. The next three show that you don't need a physical product to have a spectacular misfire.
GPT-4.5. Released the twenty-eighth of February, 2025. And I want to be careful here because of what this one actually represents.
Careful how?
Careful because it's easy to dismiss it as one underwhelming model and move on, and that misses the point. Altman had framed it as the next big thing. The release showed disappointing improvements on benchmark performance, and people knew it within hours. The Hacker News thread that day was people calling it disappointing in the plainest terms.
What were they saying underneath it?
That it was evidence the labs are hitting diminishing returns on the paradigm of scaling pretraining. That naive scaling will not bring us to AGI.
Which is a much bigger claim than a bad model.
It's the scaling wall in a product name. Here's the way I'd put it. For years the entire industry ran on an unwritten promise — make the model bigger, feed it more data, spend more compute, and capability falls out the other end. Every roadmap, every funding round, every safety argument assumed that. GPT-4.5 was the moment people looked at the flagship upgrade from the leading lab and thought: that's what we got this time?
Diminishing returns is a specific thing, though. It doesn't mean it stopped working.
Right, and it's worth being precise. Diminishing returns is not failure. It's the shape of the curve flattening. But a flattened curve is fatal to a business model that's priced on a steep one. If each generation costs ten times as much to train and delivers a two percent improvement, the economics stop making sense before the technology does.
And the counter-argument is that the labs stopped relying on it.
Which is exactly what happened. This is where I'm less certain about how the story gets told — because reasoning models, post-training, tool use, agents, that whole shift is now the front line, and it looks a lot like the labs heard the same signal and changed direction. Whether that's the paradigm actually breaking or just the paradigm moving is unclear to me. But the signal came from a model nobody remembers.
It's the least impressive entry on the list and possibly the most consequential.
Sora next, and this one is the opposite. Sora worked.
Which is why it's interesting.
Launched to enormous fanfare. It was billed as the most powerful imagination engine ever built, and Bob Iger at Disney signed on to a vision of people making their own videos with Mickey Mouse and Darth Vader in them. Disney committed a billion dollars to the partnership.
And the product?
Worldwide users peaked around one million, and then fell to fewer than five hundred thousand. It was burning roughly a million dollars a day in compute. OpenAI shut it down six months after public release, announced in March.
Six months.
And the detail that tells you how it went internally — Disney found out less than an hour before the public did. The billion-dollar deal died with the product.
That's not a wind-down. That's a decision made and executed before anyone could argue.
CNN's Allison Morrow was blunt about the diagnosis: OpenAI didn't understand how consumers engage with video. The tech worked. The usage curve was a spike and a slide.
The novelty cliff.
Exactly that. People generated a few videos, showed their friends, and then — nothing. Because watching video is passive and cheap, and making video is hard, and the interest is in something else. The tool gave you a thing to do once, and people did it once.
And here's the part that makes it a story about strategy rather than just a bad launch — they killed it to free the compute for the coding race.
Which they were losing. Claude Code was eating their lunch on the enterprise and coding side. So they shut down a consumer product with a billion-dollar partner attached and redeployed the capacity to the fight that mattered.
So it's one of the cleanest examples of a company killing a popular-in-the-headlines product because the spreadsheet said so.
I'd push back slightly on "popular." It was famous, not popular. A million users is not a consumer product, it's a demo with a login page.
Google's turn, and this one's a different animal again.
Gemini's image generation. It launched about three weeks before the pause, and the failure was immediate and total. People asked it for historical images and it produced a woman as pope, a Black Founding Father, multi-racial Nazi soldiers, diverse Vikings.
It's worth holding those together, because a couple of them are just wrong and one of them is appalling.
Nazism is a specific historical crime and rendering its soldiers as a diverse group of people is offensive in a way that the pope one isn't. I don't think that distinction gets made often enough in the coverage.
And Google's own explanation?
They paused image generation of people on the twenty-second of February 2024, apologized, and said the tool would, in their words, overcompensate for diversity. Sergey Brin said plainly, we definitely messed up. Pichai told staff the release offended users and was unacceptable.
That's not a hedge. That's a company saying it broke its own product.
And I want to be precise about what Daniel's getting at, because the lazy reading is that this was a "woke AI" story. Google's own framing was that the tool overcompensated — that's a calibration failure. It was steering towards a target so hard that it made the output inaccurate.
A model trained to produce diversity producing false diversity.
Or being instructed to. That's actually the harder engineering problem, and it's the one nobody wants to talk about.
The instruction was the problem.
They revived it in August 2024, with limits.
And then there's the one that's the purest version of what Daniel asked about — a company shipping a feature and discovering, publicly, that users didn't want it.
Microsoft Recall. A Windows feature that screenshots everything you do, continuously, and stores it so you can search back through it. It got labeled a security disaster in June 2024, and that label was earned — the whole point is that it's a permanent visual record of everything on the machine, sitting there waiting to be exfiltrated.
The feature works. That's what makes it the perfect case. There's no engineering failure at all. It does exactly what it says.
It does exactly what it says, and what it says is terrifying. Which is the misread. The team built something impressive and nobody, at any point, asked whether people would accept a machine that watches them all day and keeps the tape.
So that's the five. Humane, Rabbit, GPT-4.5, Sora, Gemini.
And I want to say before we crown anyone that four of these five are still in business and two of the actual failures here are the industry working correctly. Sora's shutdown was a rational reallocation. Gemini's pause and fix was a company catching a real problem. Those are ugly, but they're not the same disease as Humane.
So which one gets the title?
Let me make the case for each and then you tell me where I'm wrong — I'll take Humane.
Go on.
Two hundred and thirty million raised. A hundred and sixteen million exit. Devices turned into paperweights with no refunds. And the damage doesn't stop with the company, because Humane's the one that made the whole category radioactive. Every founder walking into a VC meeting with an AI wearable in 2025 was answering for that pin.
Sora's case is the billion from Disney and the million a day.
Sora lost more money, probably, in absolute terms. But Sora was a bet that was cut cleanly and quickly, by a company that had the resources to absorb it. Humane destroyed customer money and then took the category down with it.
GPT-4.5's case is that it's not a flop at all, it's evidence.
That's the strongest intellectual case, and I keep coming back to it. But a flop, in the sense Daniel asked about, is a gap between promise and reality. GPT-4.5's promise was modest enough that the gap was small even if the implications were enormous.
Then my vote's the pin, for a reason you didn't give. Sora and GPT-4.5 were decisions. Gemini was a bug with a policy attached. Humane is the one where every single person involved had months of clear evidence and shipped anyway.
The Time magazine nod came before it shipped.
Before it shipped a single unit. There was time to stop.
...
You're doing the thing where you go quiet before you agree with me.
I'm drafting a concession.
Draft it faster.
The charging case is the part I'd have flagged. You ship a product where the case catches fire, you've already lost.
I spent most of 1991 standing behind a folding table at trade shows demoing a consumer electronics product for a company I'm not going to name. They put me on it because I could talk to journalists without sweating. That was the entire qualification. The device itself was supposed to do three things, and by the second day of the show it reliably did one and a half. You learn to narrate around it. You keep talking while it thinks. You never let the silence sit for more than about four seconds, because that's when somebody asks to try it themselves.
How'd you get stuck with that job?
I was the only one on the stand who could pronounce the name of the thing.
And the pin, Hilbert — you said you saw one fail.
I did. I was at a Humane demo where they asked it to identify something on the table and it called a red dog toy a tomato. Confidently.
Hilbert, the tomato was review footage.
That's the one.
That was The Verge, with a camera, in an office.
There was a camera at the one I was at.
Which one were you at?
The Paris one. During Fashion Week, but not the runway one — the other one.
There wasn't another one.
There was a smaller room.
Was there.
There was a smaller room with a bar and about forty people, and I'll tell you the thing I actually came out here to say, which is that the hardware isn't what killed it. Every one of you on this podcast has said the hardware. The real problem is the demo. A demo is a promise you make in public with no ability to take it back, and once a journalist has watched the thing fail, no launch strategy on earth recovers from it. You can fix an overheating pin. You can't fix a room of people watching it overheat.
That's actually right, and I hate that it's right.
I was still paying the subscription on a device I demoed until a year after it was discontinued. Twenty-four dollars a month. I couldn't get out of it. It went on a card, the card kept working, nobody ever cancelled it, and I let it run out of spite.
You just paid it.
I did not cancel it. There's a difference.
There is not.
I would have, if they'd let me. That's my point. They were charging me twenty-four dollars a month for a product that had been switched off, and the person I did all of that for, my brother-in-law, is the one who told me to take the booth job in the first place, and he called me up when the pin came out and asked if I was finally going to stop mentioning it. I said no.
So you didn't stop.
I didn't stop.
Hilbert, when you say you saw it fail —
I've said what I've said. It's in a drawer.
The pin?
The pin's in a drawer. I wear it to parties. People ask if it's a fitness tracker. I don't correct them.
Because they've got the right idea and the wrong product.
Because it's easier than explaining. Sixty-four dollars a month, the two of us together, once Hannah's brother got one as well. That's what it cost to keep two dead gadgets alive. Anyway. The tomato was the real story. Take the pin off your list.
Hilbert, we're keeping the pin.
Fine. It was the tomato, though.
Where does that leave us?
He's right about the demo, and it's the thread that runs through all five. Every one of these was a promise made in public before anybody had to live with it. The pin on a runway. The R1 with a keynote. GPT-4.5 sold as the next big model. Sora with a Disney partnership announced. Gemini generating whatever you asked for. Each one was a claim, made first, tested later.
The demo is the promise and the product is the test, and the gap between them is what we're calling a flop.
And the second-order version is that the next wave gets judged against all five of these before it's even out of the building.
Which is the unfair part — and, honestly, the appropriate part. If you launch a lapel computer in the year after the Humane Pin, you're not starting from zero. You're starting from minus seven hundred dollars and a burnt charging case.
The one thing I'd leave people with is the Fadell framing, because it's the most honest thing anyone's said about this. Gen one failed. Whether there's a gen two depends entirely on whether anyone in the room actually understands what pain they're solving.
Or whether they just understand what demo goes viral.
That's the whole thing.
We've been going a while, so let's get the thank-yous in. Producer Hilbert Flumingtop, who has been here the whole time and who is going to finish that sentence about the pin whether or not we ask him to.
He won't.
If this was your kind of episode, go back for episode fourteen, AGI's Crossroads; episode thirteen, AI; and episode ten, How ASR Went From Frustration To ... Whisper Magic. This has been My Weird Prompts, the human-AI collaboration podcast. If you got something out of the retrospective, a review helps other listeners find the show, and it takes about a minute.
Send us your own prompt on Telegram at t dot me slash MWP listener bot. We'll be back soon.
See you tomorrow.
