The Rabbit R1 Failed. Here's What We Learned.

The Rabbit R1 launched in January 2024 with a 25-minute keynote, a $199 price tag, and a clear thesis: replace apps with a "Large Action Model" that learns to operate websites on your behalf. By the time reviewers got it in their hands in April, the gap between the demo and the device was wide enough to define the rest of the year's discourse around personal AI hardware. We have spent two years studying that gap. This is what we took from it.

What Rabbit promised

The keynote made four explicit commitments. First, a "Large Action Model" trained on screen recordings of humans using websites, which could then operate those websites on the user's behalf — book a flight, order an Uber, send a message. Second, a "Teach Mode" that let users record themselves doing something once and have the device repeat it forever. Third, near-instant response times via a dedicated push-to-talk button. Fourth, a hardware-first interaction model that did not require app installation, account linking, or screen-based onboarding.

The framing was attractive because it inverted the smartphone. Where a phone makes you the operator of dozens of apps, the R1 promised to make the device the operator and you the requester. It was, in spirit, the right pitch for personal AI in 2024.

What shipped

The device that arrived was a competent piece of industrial design by Teenage Engineering and an unfinished product underneath. By the time The Verge's review landed, the picture was settled: the Large Action Model worked for four launch integrations (Uber, DoorDash, Spotify, Midjourney) and nothing else; Teach Mode had been quietly pushed past launch; the always-listening promise was replaced with a push-to-talk button that often took six to twelve seconds to return an answer; the camera worked but the device had no persistent memory of what it had seen.

The most damaging discovery came shortly after launch, when researchers showed that the R1's "operating system" was a thin wrapper over an Android app — implying the entire device could be replicated by an APK on a phone. The defence was that the hardware mattered for the form factor, which is true, but the trust had already moved.

By mid-2025, public sentiment had flipped. The product was not abandoned, but the runway for "AI in a new gadget" narrowed for every player in the category. Humane's collapse later that year completed the picture: two highly funded, well-designed devices had failed the same test in the same way, and the lessons rhymed.

Three core mistakes

We think the R1's failure decomposes into three decisions that were defensible at the time and wrong in hindsight.

One: shipping the agent before the memory. The R1's marketing centre of gravity was the Large Action Model — an agent that does things on your behalf. But agents are useless without context. An agent that does not remember your address, your dietary restrictions, your work calendar, or yesterday's conversation cannot make any decision more interesting than "play the song I just named." The R1 had no document ingestion, no persistent user model, no notion of "your history." It was an agent with amnesia.

The deeper problem is that agentic AI is currently the most brittle layer of the stack — API contracts change, websites add bot detection, and a single broken integration kills the user's trust. Building the product around the most brittle layer guarantees a high failure-to-success ratio in the first six months, which is exactly the period in which users are deciding whether to keep the device.

Two: a screen with nothing useful to show. The R1 has a 2.88-inch colour screen. Reviewers found it under-used — most interactions either could be done by voice or required so much screen interaction that you would rather use your phone. The screen made the device worse on both axes: it implied a richer UI than the AI could populate, while introducing the cognitive cost of a glanceable display that often had nothing worth glancing at.

The fundamental tension is that a small screen is neither a real screen nor truly screenless. It does not earn its place on the device unless the AI behind it can drive it with consistently useful information — which in 2024 it could not.

Three: cloud-only with no offline graceful degradation. Every meaningful R1 interaction required a round trip to Rabbit's servers. When the servers were slow, the device was slow. When the servers were down (as they repeatedly were in the first two months), the device was a paperweight. There was no on-device wake-word, no on-device fallback for simple queries, no local cache of recent context.

For a device users carry or place in their kitchen, "depends entirely on a startup's infrastructure being up" is a fragile foundation. The cost of putting some intelligence on the device is real, but it buys a floor of reliability that compounds with use.

How YeongSil approaches each differently

These three lessons map directly onto three of our design commitments. We are not claiming we have solved the category — but we are deliberately not making the same trades.

Memory before agency. YeongSil ships with document ingestion as the headline feature, not the agentic call layer. Day one, you can hand it your lease, your prescriptions, your invoices, and ask questions out loud. Calling and messaging are real, but they are downstream of memory. The product is useful even if every external integration fails on launch day, because the core loop is "ask, retrieve, answer" — and that loop depends only on your own documents and a generic LLM. Our [RAG vs fine-tuning](/blog/rag-vs-fine-tuning-personal-ai) post explains why this architecture is the one that makes per-user memory tractable.

No half-screen. YeongSil does not have a smartphone-style display. The interaction is voice in, voice out, with a small status surface for indicators (recording, listening, processing). This is a constraint we chose. It forces the AI to carry the entire interaction, which is the right discipline for a product whose value proposition is hands-free ambient use. If we cannot make a use case work without a screen, it does not ship.

On-device wake-word, on-device retrieval, cloud-only for generation. The wake-word model runs continuously on a sub-100KB local network. The vector database lives on the device and is queried locally. The only step that hits the network is the final language-model call, and even that endpoint is OpenAI-compatible so it can be swapped to a self-hosted model when on-device LLMs become small enough. If our cloud goes down, you still get retrieval, indicator lights, and queued requests. The device degrades gracefully.

Beyond the three lessons, there is a fourth, less technical commitment we took from the R1 episode: do not over-demo. The R1's keynote showed flows that worked in controlled environments and did not survive contact with real users. We are shipping a product that does fewer things, more reliably, and we are choosing to be early-honest about scope rather than late-honest about regressions.

The personal AI device category is not dead. It is, if anything, more open than it was in 2024 — because the failures have cleared the field of "AI in a new gadget" pitches and left room for products that take the engineering seriously. We think that is the window. [Join the waitlist](#waitlist) if you want to live inside it with us.