The Rabbit r1 device, which was highly anticipated at the start of 2024, fell short of delivering its initially advertised features, leading to a rapid decline in market enthusiasm. Jesse Lyu, CEO of Rabbit, admitted, "We set our expectations too high at the product launch." However, he also announced that an upcoming update next week will introduce the much-awaited "Large Action Model" (LAM), enabling universal functionality on the web.

Although skeptics view this as a potential pivot or a belated remedial measure, Rabbit's vision to develop cross-platform agents compatible with both web and mobile applications remains theoretically significant.

According to reports, Lyu stated that over the past six months, the Rabbit team has been focused on fixing bugs, enhancing response times, and adding minor features. However, the r1 device is still primarily restricted to interacting with large language models (LLMs) or accessing seven specific services, such as Uber and Spotify.

"This is the first version of LAM, trained using data collected by data laborers, but it is not universal and can only connect to those specific services," Lyu explained. Whether or not it is called LAM, the model has not met the functionality levels that Rabbit promoted at launch.

However, Rabbit is set to release a universal version of LAM, which is not tailored to any specific application or interface. This new iteration will be a web-based agent capable of performing everyday tasks such as purchasing concert tickets, registering on websites, or playing online games. Lyu mentioned that the agent will be able to analyze buttons, fields, and images on webpages and interact based on the learned functionalities of the web.

During demonstrations, the LAM agent successfully registered a domain name for a film festival website and located the r1 device on eBay, although it was redirected to purchase directly from the official website after explicit instructions. Additionally, it completed Dictionary.com's daily word game, although some instruction optimization was required during the process.

The LAM agent currently operates using a new cloud-based browser, but Rabbit is developing a local version, such as a Chrome extension, allowing users to leverage existing sessions without needing to log in again.

However, the LAM agent is still in the "beta" phase. While it possesses complete functionality, improvements are still needed. For instance, it can outline task steps but lacks the intelligence to bypass unnecessary ones. Additionally, Rabbit has stated that it will not be enhancing the model by collecting user data but will introduce a "teaching mode" that allows users to demonstrate how to perform specific tasks.

Rabbit is also developing a desktop agent to interact with applications such as word processors, music players, and browsers. Although this feature is still in its early stages, preliminary results have been achieved.

While the LAM agent shows potential, its real-world applications remain unclear, lacking a "killer app" comparable to robotic vacuum cleaners.

Addressing questions about why applications are not being developed, Lyu explained that creating apps would be constrained by the limitations imposed by platforms like Apple and Google, making it difficult to surpass their built-in voice assistants. Moreover, these platforms take up to 30% of the revenue. Therefore, Rabbit has opted for a cross-platform web agent approach to seize a larger market opportunity.