For a couple of years now, the tech industry has poured hundreds of millions into a wave of ambitious AI gadgets — screenless, standalone devices you'd talk to instead of your phone. Where did all those gadgets end up? Probably a landfill. But a cheap-ish device — $159, comparatively — that was built to do just one thing, single-serve, quietly became the world's number-one AI note-taking device. It records your meetings and summarizes them. That's the whole job. So why did the boring one succeed where the flashy ones failed? That question is the story of where physical AI is actually going.

Hashi's take: The future of AI in the physical world isn't a gadget that tries to do everything. It's a hundred small ones that each do a single thing well, tucked into objects you already own. Businesses fixated on building the impressive do-it-all device are chasing the graveyard. The winners are picking one job.

If your company is being pitched an "AI-powered everything device," forward this first.

Rewind to 2023. ChatGPT had just exploded, and suddenly we were flooded with announcements for AI-powered devices — gadgets you'd carry around and talk to, so you'd never have to look at a screen. Two of them led the pack, both from big-name teams with serious money behind them.

The Humane AI Pin was the flagship. Founded by former Apple executives, it raised roughly $230 million and promised a post-smartphone future, projecting information onto your palm with a tiny laser. It launched at $699 plus a monthly fee. Reviewers savaged it — one prominent critic called it the worst product he'd ever reviewed. By early 2025, Humane sold its remains to HP for $116 million and bricked every Pin it had sold. The servers went dark. The devices became paperweights.

The Rabbit R1 was the other. A charming orange square that sold 100,000 units on launch hype — I was one of the early adopters, and mine's been sitting in a drawer gathering dust ever since. Within five months, 95% of buyers had stopped using it. I had company. Then came the Friend pendant — a $129 device that listens all day and sends you "friendly" texts. Its million-dollar subway ad campaign got defaced by New Yorkers scrawling "AI is not your friend" across it. The company reportedly sold 3,000 units against 1,000 shipped.

These weren't just individual flops. Industry counts put the failure rate of AI hardware startups in 2025 around 85%. The pattern is consistent enough to be a lesson: the gadget that tries to do everything keeps dying because it solves a problem almost nobody has, while competing with the excellent AI device already in everyone's pocket — the phone.

While those devices were dying, a far less exciting product was doing fine. Plaud makes an AI voice recorder about the size of a credit card. It does one thing: records a conversation and turns it into a clean transcript and summary. That's it. No laser, no pendant, no promise to replace your phone. And it worked — it's sold over two million units while the flashier gadgets went to the landfill, precisely because it does one boring, genuinely useful job instead of trying to do everything.

That's the model I'd call single-serving AI — a device built to do one thing well, powered by AI under the hood, that slots into a workflow you already have instead of asking you to change your life around it. The intelligence isn't the product you interact with. It's an ingredient baked into an object with a clear purpose.

That inversion is the contrarian bet, and I think it's exactly why Plaud is winning. Everyone else led with the AI — "look what our assistant can do." Plaud led with the job — "you need meetings summarized; here's a thing that does that" — and hid the AI inside. Nobody buys it because they want to talk to an AI. They buy it because they're tired of taking notes. The AI is invisible, and that's the whole trick.

Under the hood, this shift has a name the industry is converging on: small language models, or SLMs — sometimes called micro LLMs. These are compact, task-specific models with a few billion parameters or fewer, small enough to run on modest hardware instead of a data center. Gartner projects that by 2027, organizations will use small task-specific models three times more than general-purpose large ones.

The logic is straightforward. A giant model that knows everything is overkill for a device that only needs to do one thing. A pill dispenser doesn't need to discuss philosophy — it needs to recognize a pill, track a schedule, and notice when a dose was missed. A fridge doesn't need to write poetry — it needs to know what's inside it and when it's going off. For narrow jobs like these, a small, focused model outperforms a giant general one, at a fraction of the cost, power, and size.

This is what makes the coming wave of objects possible. The medicine box that reminds an elderly parent — and calls someone if two doses are skipped. The fridge that plans meals around what's actually inside and expiring. The door lock that recognizes a family member's gait. None of these needs a supercomputer. Each needs a small brain that does one job, and small brains are now cheap enough to put almost anywhere.

The one device that tries to do everything ends up master of none. The smarter path is the opposite: single-serving devices that each nail one job, and can later be centralized — fed into a single orchestration layer that ties them together. Which raises the real question for any business: what's the one job you'd hand to a small, embedded intelligence, done well enough that nobody notices the AI at all?

A hundred small devices, each running its own AI and quietly learning a piece of your life, raises an obvious privacy question — one we can't ignore. But it's not one to be paralyzed by either. Some of these objects will keep your data in a corporate cloud; others will run the model locally and never send anything anywhere. The trend that makes single-serving AI cheap also makes on-device, private versions possible — Apple's models already work this way. The job isn't to fear every smart object. It's to understand what each one actually does with what it learns, weigh that against what it gives you, and decide for yourself whether the trade is worth it.

It's easy to look at a smart recorder, a smart pill box, and a smart fridge and see three unrelated gadgets. My opinion? They're not going to stay unrelated for long.

As we start embedding small, single-serve AI — micro LLMs — into more and more everyday objects, the next purpose becomes obvious: feed what each one learns into something more centralized, something that can put it all together and actually benefit us. Each device learns a slice of you. The recorder learns how you think and talk. The pill box learns your body. The fridge learns your habits. Some of these already exist, each locked in its own silo, owned by a different company. Some haven't been invented yet. But none of that lasts — it's only a matter of time before they connect.

The single-serving devices arriving now are the sensory organs of an intelligence that doesn't have a body yet. The real question isn't what each object does. It's what happens when they all report to the same place.

I get excited thinking about the future. A central intelligence layer. Not another gadget — a private intelligence sitting on top of all the small ones, taking every fragment they collect and assembling it into one continuous picture of you. It grows with you over time. Every dumb-looking object you add makes that picture sharper. A hundred tiny intelligences, all pointed at one life, stop looking like appliances and start looking like the raw material for a model of you.

This is a big idea and a big vision — and I don't think it's as far off as it sounds. But it raises a stack of questions we need to sit with before we sleepwalk into it.

Start with readiness. Are we, as a public, actually prepared for this? Sentiment toward AI is fragile, and privacy concerns are real and rising. Then there's the question of what that central layer would even orchestrate. Picture the good version: it coordinates with your healthcare providers — maybe even their AI agents — to proactively schedule appointments, chase follow-ups, manage medication, and shift the whole system toward preventative care instead of reactive treatment. That's a genuine improvement in quality of life, longevity, and how productive we get to be.

But the same picture has a shadow. Who actually owns that data — you, or the company that built the layer? Are we comfortable with the answer? And what's the worst case if someone got hold of a complete, centralized model of a person — every habit, every prescription, every conversation, in one place? That's not a reason to stop. It's a reason to build carefully.

Because that's the part that matters most: this is coming, and the upside is real, but the groundwork has to be laid thoughtfully — while it's still early enough to shape.

That's what Beyond the Chatbox is really about. And we're just getting started.

I keep a Rabbit R1 in a drawer — a monument to the idea that AI had to arrive as one dazzling device you'd hold up and marvel at. It didn't. It's arriving as a recorder, a pill box, a lock, a lamp: things so unremarkable you'll forget they're intelligent at all.

- Hashi