Sometimes I like to make a post (like this one) fleshing out something I wrote elsewhere with a little extra detail that didn’t fit into the original, for whatever reason.
I recently wrote a post for Hackaday.com featuring a voice-controlled, locally-installed AI agent on a Raspberry Pi 5.
It’s a great showcase for how much better these tools (both hardware and software) are getting. What’s available to hobbyists is so far beyond what it was ten (or even five) years ago.
Only a couple of years ago roughly a 3 billion parameter LLM model would be considered the minimum for coherent talk, this model uses a 1.3 Billion Qwen model. That’s not even getting into the support for other functions. The “AI in a Box” project I covered is a good example of what this scene was like two years ago.
I’d like to briefly mention that the term “agent” (when it comes to AI systems) is currently one of those terms whose meaning depends on who is speaking.
The definition of agent seems to settling on “calling tools in a loop [to accomplish a particular task]”.
The other meaning of agent is roughly “an AI system that can use a browser (for example) and click on stuff as if it were a person”.
I cannot resist saying that I am personally more a fan of the latter definition (when an entity is takes action based on your instructions – for example using a browser as if it were you – it is literally acting as your agent) but the definition seems to be settling on the former.
So just be mindful that you can run into both definitions at this time, depending on who you are speaking to.