The AI landscape is vast, and everyone, from Google to Microsoft and OpenAI to Anthropic, is trying to stake their claim. Today, much of AI is computed in the cloud; while these workloads are expected to shift towards a more hybrid model that balances computing between local and cloud, there is also a push for more agents to help manage the many different LLMs and SLMs that exist today.

Some companies like OpenAI and Google already have their own agents, while others like Cohere help you build your own AI agents with RAG (retrieval augmented generation). Stanford’s HAI (Human-Centered Artificial Intelligence) has a 2024 report detailing how AI agents are likely to move the industry forward and improve human-to-AI interactions. Advanced AI assistants have also become important enough for Google to publish a 273-page paper on The Ethics of AI Assistants. Much research is already being done in the space, including by companies like NVIDIA, which has developed its Voyager agent in partnership with universities. Many of these companies are building complex AI systems to run generically through a smartphone or browser app. We’ve even seen investments in custom hardware from AI players such as Midjourney or OpenAI, which is investing in many companies building hardware for AI.

None of these companies, however, is like Rabbit, which launched its handheld device (co-designed by Teenage Engineering) as a conduit to help scale its own cross-platform, cloud-based generic agent system called LAM. LAM aims to enable AI agents to perform tasks quickly and inexpensively.

Introduction to Rabbit, Rabbit OS, and Rabbit R1

The Rabbit r1 is a handheld device about the size of a folded Samsung Galaxy Z Flip. It’s meant to be carried as a personal AI-powered assistant that can help answer complex questions, solve math problems, and translate speech. The r1 came to market with much hype, and since its release earlier this year, it has wildly surpassed Rabbit’s expectations of selling a few thousand units; by now, there are over 100,000 units sold. This has been both a blessing and a curse, because the level of demand forced Rabbit to separate pre-orders into different waves of fulfillment. (I experienced this as someone who pre-ordered a Rabbit r1; I got lucky enough to be part of the earlier waves.)

The user experience early on was buggy and inaccurate, but it quickly improved to a respectable state, with new features and fixes added weekly. Rabbit has deployed significant quality updates since April’s launch to ensure that the product meets people’s expectations. Since launch, Rabbit has already pushed out 16 OTA updates to the r1, fixing bugs and adding features, and I don’t expect that pace to slow down much.

One of the things that sets the Rabbit r1 apart from the rest of the industry is that it is an entirely custom piece of hardware; this allows Rabbit, even as a startup, to independently run its custom Rabbit OS, which is built based on Android. Both now and in the future, this approach minimizes the potential for confinement by the existing app-based ecosystems of smartphones and tablets. This means that Rabbit has complete control of all the sensors and information coming in from the wireless subsystem, which includes a 4G modem.

Rabbit’s LAM Approach

Rabbit OS is also different from other Android-based operating systems because it is heavily tied to the cloud to run LAM. LAM creates infrastructure that helps solve the problems that users might encounter when trying to use AI models on asynchronous problems. These problems manifest themselves in the form of understanding human intentions that might be communicated imperfectly, in pieces, conversationally, or over many steps. LAM can translate those intentions into actions on a webpage or webapp that require a sequence of steps that unfold over time rather than in an instant. LAM was originally introduced as a “large action model,” but the company’s definition has expanded beyond being just a model and into more of a system of agents.

Rabbit’s approach to Rabbit OS also allows it to be more flexible in addressing different problems using different AI tools. For example, Rabbit OS uses a combination of GPT 4o, Perplexity Pro, Claude, and Wolfram|Alpha to tackle different user queries. Rabbit OS chooses the best LLM to use based on the capabilities of each model. Once the LAM understands the user’s needs, it can address the user’s request by tapping into the appropriate LLM and performing the necessary action. This approach has evolved over time as LAM has evolved.

With companies discontinuing APIs for their apps and making access more complex and more expensive, LAM can simplify a user’s life by performing a task with a single voice command. Through Rabbit’s research and users’ feedback, the company has continued to evolve how LAM works, with new capabilities and modalities created through new features like teach mode. Teach mode will enable Rabbit to offer new capabilities as modular features that its creators could potentially sell in a marketplace.

Rabbit’s AI go-to-market is unique mainly because it has taken the opposite path of many of its competitors by building both its hardware and software together, then evolving its software rapidly. Using LAM, the company hopes to become more future-proof to changes to web application user interfaces or LLMs. The company’s use of diverse models and its multiple potential revenue streams allow it to continue to grow alongside the AI industry’s boom. Rabbit’s approach also allows it to be hardware-agnostic and potentially launch new devices beyond the Rabbit r1 with different form factors.

What’s Next for Rabbit and LAM?

Recently, I got a chance to see the future of Rabbit’s LAM-powered generic agent, which is powered entirely independently of any kind of pre-training. This agent uses a combination of a planner and a controller, allowing the agent to think, look, and act in real-time, which makes it more resilient to UI changes. It could easily demonstrate booking flights and making restaurant reservations on almost any site. While at this earlier stage, LAM might be slower, the system has already demonstrated the ability to perform a full series of tasks on behalf of a human using only voice as an interface. Eventually, I could see a future where LAM could perform these tasks faster than a human.

Rabbit’s unique approach to owning the entire vertical stack, from the cloud down to the device, enables it to innovate rapidly in ways that its competitors might find difficult. The LAM at the center of Rabbit’s innovations is constantly evolving, and the company is continually updating the whole vertical stack to increase capabilities and drive down latency.

The entire industry is moving in the direction of agents, with all the big players investing heavily in advanced agents and how they might better use LLMs to address users’ needs; in some ways Rabbit is already ahead of them. With the rapid pace of innovation in AI, Rabbit is one of the companies I’ll continue to keep a close eye on for its latest developments.