Last November, Humane unveiled its first product - the Ai Pin. For those of you who aren't familiar, Humane is a hardware company started in 2018 by former Apple directors Imran Chaudhri and Bethany Bongiorno. The company remained in stealth mode until 2021 - yet it raised nearly $150 million in VC funding without so much as a product announcement.
To be fair, they had a decent track record: the two founders were directors of design and software engineering at Apple, and both have their names on dozens of patents from their time there. But after years of cloak-and-dagger, the company finally showcased the Ai Pin, a wearable device with a camera, a speaker, and, interestingly, a projector that uses your hand as a display and an input device.
The tech seemed pretty futuristic and next-gen, but so did the price tag. When the product was unveiled, Humane announced that it would sell for $699 plus a monthly data subscription. Now, five months later, the first reviews are coming in for the device - and they're brutal. The Verge titled its review "not even close," and in its opening paragraphs, called out some pretty big issues:
Should you buy this thing? That one’s easy. Nope. Nuh-uh. No way. The AI Pin is an interesting idea that is so thoroughly unfinished and so totally broken in so many unacceptable ways that I can’t think of anyone to whom I’d recommend spending the $699 for the device and the $24 monthly subscription.
...
As the overall state of AI improves, the AI Pin will probably get better, and I’m bullish on AI’s long-term ability to do a lot of fiddly things on our behalf. But there are too many basic things it can’t do, too many things it doesn’t do well enough, and too many things it does well but only sometimes that I’m hard-pressed to name a single thing it’s genuinely good at. None of this — not the hardware, not the software, not even GPT-4 — is ready yet.
Over on YouTube, Marques Brownlee, a tech reviewer with 18.6 million subscribers, has called the pin "the worst product I've ever reviewed... for now."
To do a "normal" hardware startup, you need a hefty dose of grit and/or naivete - probably a bit of both. To do an AI hardware startup, though, it seems like you need a hefty dose of ambition and/or insanity.
Hardware is hard.
An old startup cliche is "hardware is hard," but it's a cliche for a reason. When it comes to building hardware products versus building software ones, there are many, many more challenges on the hardware side:
Higher development costs and capital requirements: Hardware startups typically require significantly more upfront investment for things like prototyping, tooling, manufacturing setup, inventory, etc. compared to software startups.
Longer development timelines: Designing, prototyping, testing and manufacturing physical products takes months or quarters compared to days or weeks for software. There are also the challenges of mass manufacturing, quality control, inventory management, and complex supply chains.
Less flexibility to iterate: Making changes to hardware is much slower, costlier and more difficult than pushing software updates. The entire process is less forgiving of mistakes, and lower product margins can also constrain product flexibility.
Physics: Tangible products, unlike all but the lowest-level software, have to deal with the laws of physics. There are hard limits to how much stress plastic and glass can withstand, how hot transistors can get, and how much wear and tear components can endure.
Even in the best of conditions, hardware startups are incredibly difficult. And working with generative AI should in no way be considered "the best of conditions."
LLMs are weird. And kind of bad.
What makes things messy is the fact that generative AI is still a big unknown. We're still in the early stages of understanding what LLM-enabled software is uniquely good at, let alone hardware.
Prompting best practices are so, so weird. UX patterns, like Copilots and streaming responses, are only starting to emerge. And if you’re a regular reader of this publication, you’ll know that generative AI is still an unpredictable, flawed technology.
Wharton professor Ethan Mollick describes some of this weirdness in more detail:
LLMs are made of software, but they do not work like most software. They are probabilistic and mostly unpredictable, producing different outcomes given the same input. Though they don’t think, they produce simulations of human language and thought that are, as far as we can tell, original enough that they out-invent most humans. The are perceived as more empathetic, and more accurate, than human doctors in controlled trials. And yet they are also limited in odd ways that are surprising, like an inability to do backward reasoning.
LLMs are essentially just a really fancy autocomplete. So how can a fancy autocomplete do these things? The answer so far, as described in an excellent overview in the MIT Technology Review is “nobody knows exactly how—or why—it works.”
The result is that working with these systems is just plain weird.
What makes things even messier is how fast the industry is advancing. The original industry-leading GPT-4 model, now a whopping 13 months old, is considered old news. Although we’re far from mastering the current state of LLMs, we keep seeing more features and better capabilities. We have no idea what the “best model” will be a month from now, let alone how best to work with it.
This creates a technology trap, which I’ve dubbed “AI quicksand.” One of the ways of avoiding the trap is to be wary of integrating too deeply with any given AI platform:
While tight integrations can lead to better UX, know that you're making a technology bet. And until the dust settles, it's better to stay agile with lightweight integrations with the various large language models.
With that in mind, building a hardware product whose value is tied to the latest LLMs seems a little bonkers. LLM software may look radically different a year from now, but hardware startups are also experimenting with form factors, UX, input methods, and more.
Given these challenges, a more agile, experimental, and humble approach - insofar as you can do that with hardware - seems like it might be a better strategy.
Taking the road less hyped
Perhaps the biggest factor in the AI Pin's lackluster reception was Humane's own hype machine. The startup was intentionally mysterious and opaque. It showcased its first product for the first time at a TED talk. And it explicitly positioned the product as trying to (eventually) replace the smartphone, even going so far as to film an iPhone-esque launch video.
If you're Apple or Google, you might be able to get away with that level of ambition and secrecy. But hardware upstarts will have a much more difficult time. After spending tons of marketing dollars on photoshoots and custom vending machines, even Snapchat still had to throw away $40 million of unsold smart glasses. Hardware is hard.
But there may be a different approach. The AI startup Rewind (now rebranded as Limitless) just pivoted from software into hardware by introducing a $99 pendant to record and remember your conversations. They started with an AI-powered recording app focused on a well-defined use case and expanded into hardware to further that solution. They’re not promising a crazy vision, nor are they aiming to disrupt the best-selling product of all time1.
I genuinely wish Limitless and Humane the best of luck, because I'm honestly torn on how to feel about the coming wave of AI-enabled devices. While it's easy to dunk on a bad product launch, I'd much rather cheer for bold experimentation in hardware. We often complain that modern smartphones all look the same, but we're also quick to ridicule any unconventional design that doesn't immediately catch on. Though, to be clear, we're definitely going to see some wild AI hardware in the near future.
The Ai Pin likely won't be the last hardware flop in the coming months2. As The Verge put it: "AI gadgets might one day be great. But this isn't that day, and the AI Pin isn't that product." Yet I’m excited about the prospect of new, fresh ideas in hardware when it comes to working with LLMs. Humane (or any other startup) shouldn’t get a pass on a bad product for being an AI hardware pioneer, but we also should expect (and understand) a lot of flaws in the short term with new devices.
We’re undoubtedly going to deal with a lot of messiness in the short term, but that’s okay - messiness has been the nature of startups for a long time, and AI isn’t going to change that.
That said, there are some significant privacy questions to consider with always-on recording wearables.
The Rabbit R1 is another high-profile device I've got my eyes on. The first units will be delivered later this month, but it also has a lot of hype to live up to. We'll soon see how it compares to the company's well-produced demo videos.
Even though I was skeptical (to put it mildly) about the Ai Pin from the start, I still binge-watched the reviews from Marques Brownlee, The Verge, and Mrwhosetheboss out of curiosity.
It's almost painful, considering how much care seems to have gone into the hardware and the overall build. They all left me feeling like the Ai Pin could've been a pretty successful product if it was a much cheaper, less overengineered alternative to the Apple Watch / smarwatch that lived on your shirt and still offered instant access to the phone's AI features, apps, etc, while sporting its own camera for scanning the environment and taking photos/videos. That would also not require any monthly subscription, etc.
Let's see where Rabbit R1 lands!