MKBHD on AI Hardware: The Humane AI Pin, Rabbit R1, and Why AI Gadgets Keep Failing
The internet's most trusted tech reviewer delivers the most devastating consumer verdict on AI hardware — the Humane AI Pin is the 'worst product I've ever reviewed,' and it exposes why standalone AI gadgets keep failing.
Top Claims — Verdict Check
The Humane AI Pin is the worst product MKBHD has ever reviewed — slow, inaccurate, and solves no problem that a phone doesn't already solve better
🟢 Real“This is the worst product I've ever reviewed. It's slow, it gets answers wrong, it has terrible battery life, and it does nothing that your phone can't do better. I wanted to love this. I don't. [representative paraphrase]”
Standalone AI hardware devices face a fundamental problem: they're competing with a phone that already does everything they do, plus more
🟢 Real“Every AI gadget faces the same question: why wouldn't I just use my phone? If your answer requires a 10-minute explanation, your product has failed the simplest test. [representative paraphrase]”
AI product demos consistently overperform relative to the actual shipped product — the gap between demo and reality is the central problem in AI hardware
🟢 Real“The demo was incredible. They showed this seamless AI assistant that translates in real time, answers questions instantly, and replaces your phone. The reality is a device that takes 10 seconds to respond, gets half the answers wrong, and overheats in your pocket. [representative paraphrase]”
The Rabbit R1 has the same fundamental problem — a $200 device that does less than a free app on your phone
🟢 Real“The Rabbit R1 is a solution looking for a problem. Everything it does, your phone already does. And your phone does it better, faster, and without carrying an extra device. [representative paraphrase]”
The best AI products will be software features in existing devices, not new hardware categories
🟡 Partially True“The future of AI isn't a new gadget. It's better software on the phone you already have. Apple Intelligence, Google Gemini in Android — that's where AI actually works for consumers. [representative paraphrase]”
What's Real
MKBHD's review was confirmed by every subsequent assessment. The Humane AI Pin launched at $699 plus a $24/month subscription, and within three months, Humane was reportedly exploring a sale after returns exceeded sales. The device's response time averaged 6-10 seconds in real-world testing (versus the sub-second demos), its projector was unreadable in daylight, and it hallucinated answers confidently — telling a reviewer that a tennis match was at a different score than displayed on the TV in front of them. The Rabbit R1 fared similarly: sold 100,000 units on launch hype, then received a 2/10 from The Verge. Its core 'Large Action Model' technology was shown by teardown analysts to be essentially an Android app wrapper. The broader pattern — AI hardware demos that wildly outperform shipping products — matches Gebru's and Narayanan's critiques of AI overclaiming. MKBHD's 18.6 million YouTube subscribers meant his review reached more consumers than any academic paper ever could, making it arguably the single most impactful piece of AI criticism published in 2024.
What's Hype
MKBHD's conclusion that 'the best AI will be software on existing devices' is probably right for 2024-2026 but may age poorly. The smartphone was also a 'solution looking for a problem' before the App Store ecosystem made it indispensable. AR glasses (Meta Orion, Apple Vision Pro successors), AI earbuds with real-time translation (Google Pixel Buds, AirPods with conversation awareness), and wearable health AI (continuous glucose monitors with AI interpretation) represent hardware categories where AI might eventually justify a new device form factor — not because the phone can't do it, but because the form factor enables use cases the phone's form factor doesn't serve well. MKBHD evaluates hardware through a consumer electronics lens (does it work today, right now, out of the box?) which is the right frame for a review but potentially the wrong frame for assessing a category's future. The original iPhone was worse than a BlackBerry at email and worse than an iPod at music. The App Store changed the equation — an AI hardware device that finds its equivalent unlock could prove the category viable.
What They Missed
The venture capital dynamics that produce AI hardware failures. Humane raised $230 million before shipping a single unit — that funding came from investors (including OpenAI's Sam Altman and Salesforce's Marc Benioff) betting on the category, not the product. The VC funding model incentivises grand visions and demo-ready prototypes over working products. The Rabbit R1's business model was even more questionable: $200 hardware at low margins with no subscription revenue to fund ongoing development. These aren't product design failures — they're business model failures that product reviewers don't typically analyse. The accessibility angle is also absent: for users with visual impairments, motor disabilities, or conditions that make smartphone use difficult, a voice-first AI device could be genuinely valuable — but none of the reviewed products were designed with accessibility as a primary use case. The emerging market perspective matters too: in markets where smartphone penetration is lower or data costs are prohibitive (parts of rural ASEAN, Sub-Saharan Africa), a purpose-built AI device with offline capability could serve a real need that MKBHD's San Francisco perspective doesn't capture.
The One Thing
If your AI product can't answer 'why wouldn't I just use my phone?' in one sentence, it will fail — and that test applies to AI software features, not just hardware.
So What?
- Apply the 'phone test' to every AI feature you're building: if a user can get the same result by opening ChatGPT or Claude on their phone, your feature needs a stronger reason to exist — speed, integration with their data, or a UX that the phone app can't match
- The demo-to-reality gap is the biggest risk in AI product development. Before shipping any AI feature, test it on the 50 worst-case inputs, not the 5 best-case demos. If it fails more than 20% of the time, it's not ready.
- For Malaysian tech companies: don't build AI hardware. Build AI software that runs on the phones your market already has. Malaysia's smartphone penetration is 97% — the device problem is solved. The software opportunity is wide open.
Action Items
- 1Watch MKBHD's Humane AI Pin review (25 minutes) and his Rabbit R1 review as a masterclass in honest AI product evaluation. Then apply the same evaluation framework to your own AI features: does it actually work in real conditions, or does it only work in the demo?
- 2Run a 'phone test' on every AI feature in your product roadmap. For each feature, ask a team member to accomplish the same task using ChatGPT or Claude on their phone. Time both approaches. If the phone approach is faster or equivalent, your feature needs rethinking.
- 3Build a 'worst-case demo reel' for your AI product: record the 10 worst outputs from real user sessions and show them at your next product review meeting. This is the antidote to the demo-culture that killed Humane and Rabbit.
Tools Mentioned
Humane AI Pin
AI wearable ($699 + $24/mo) — cancelled by most reviewers as a failed product. Response times of 6-10 seconds, frequent hallucinations, unreadable projector.
Rabbit R1
$200 AI hardware device — sold 100K units on hype, rated 2/10 by The Verge. Core technology was essentially an Android app wrapper.
Apple Intelligence
Apple's on-device AI features for iPhone — cited by MKBHD as the more sensible approach to consumer AI integration
Workflow Idea
Build a 'real conditions' testing protocol for your AI features. Every week, have one team member use the AI feature for their actual work — not scripted test cases. They log every failure, every slow response, every wrong answer. After four weeks, you have a real-world reliability score that's far more useful than any benchmark. The Humane AI Pin scored well on internal tests and failed catastrophically in real use — this protocol prevents that pattern.
Context & Connections
Agrees With
- gary-marcus
- arvind-narayanan
Contradicts
- sam-altman
Further Reading
- MKBHD's Humane AI Pin review — youtube.com/@mkbhd (April 2024)
- MKBHD's Rabbit R1 review — youtube.com/@mkbhd (April 2024)
- The Verge's coverage of Humane exploring a sale — reporting on the business model failure behind the product failure