← back

An engineering lesson from Rabbit

Created on: 2026-10-27

I was really excited about the idea of a device like Rabbit. The ability to talk to an assistant and avoid sitting in front of a screen sounds great. I really wanted to believe.

To make this work, they needed a way to map voice intent into actions to control different UIs, hence their LAM model (Large Action Model).

If you haven’t been following the news, it turns out that there is no such LAM, and they are simply using a wrapper around ChatGPT and a few python scripts with pupetter(similar).

This isn’t great: say that you are going to do one thing and do another to get sales for your product is a shady business practice, and they are getting their good share of backlash because of this.

Yet, there is something to learn about this. It is possible that they never intended to build LAM, it is possible that they didn’t have the time to develop it and had to go to market with a workaround, but it is also possible that building a LAM that is reliable is hard!

De-tour

Everyone is talking about AI these days. We have all used ChatGPT and is quite useful for drafting that dreadful email, or summarizing the report that you didn’t want to read.

If you are in the space, you might have seen some very fantastic demos in which you drop a couple of files into a chat and you get an Excel spreadsheet with formulas and all.

So, when we come to one of this large conferences, when everyone is buzzing AI, what do we talk about?

Sure, the talks are about demos, are all shiny and cool. But that’s not what the engineers are saying.

Engineers are worried.

Large Language Models (LLMs), powering chatbots like ChatGPT, and Gemini are unruly beasts: they hallucinate, they are inconsistent, and they are expensive.

If you are building with LLMs and you can’t seem to get that LLM to behave, you are not alone.

You might be there together with ML Engineers at Rabbit. Realizing that extracting a few keywords from a voice message and executing a script that performs search using perplexity, is easier than teaching an LLM to search in google.

We will probably get better with time, but until then, hang-tight, and don’t trust demos until you try them!