Hacking on Llama3

The Llama 3 model was released by Meta into the world just over a month ago. The latest model to be offered by Meta, it is free and has a relatively open license that lets developers deploy it into most commercial applications and services. It's also amongst the most capable model available that you can run locally. So at least at the moment Llama 3 is at the bleeding edge of Small AI, and just few days ago five hundred developers gathered and spent twenty four hours hacking on it to find out what it could do.

Amongst the projects that ranged from security research on how to jailbreak the Llama model, to getting the models to debate each other, the overall winner of the hackathon was Open Glass AI.

OpenGlass is a twenty dollar pair of smart glasses that can answer questions about not just what you're looking at right now, but anything you've seen since you put them on.

I put on the OpenGlass, and walk around Shack15 for a few minutes. However, I cannot find where I left my phone, so I say, 'Hey OpenGlass, where is my phone?'. To which, it would respond with a description of where it last saw my phone.

If you're interested in the project the developers are going to ship a limited number of pre-built kits, and you can fill out the interest form to get notified when they're available. Although if you miss out, or don't manage to snag one when they're released, the design and source code has also been released on Github.

But I actually thought one of the more fascinating projects at the hackathon was LlamaFS. Files systems are something of a controversial topic, and how you view your file system is probably going to depend — at least somewhat — on how old you are. Because over the last couple of decades we've seen a move from structure-first filesystems to search-first filesystems. If you're old enough that you grew up with UNIX or MS DOS, then your files are probably sorted hundreds of folders. Where a file is, and how it's named, tells you what is in the file. But, if you're young enough that you grew up in a post-Google, and post-iPhone era, your files are probably going to be in a big pile — cloud services like Google Drive almost demand that approach — and you rely on search to find things rather than structure.

LlamaFS moves beyond existing search-based file systems into the realms of self-organising. It's a file system for us older folks that need structure in our lives. It uses the model to look inside the files and figure out the contents, and then derives and name and folder structure in a semantically meaningful way. It automatically renames and organizes your files based on their contents. It supports many file types, including images and audio, and like all the projects from the hackathon, the source code is available on Github.

If you want to learn more about the projects that happened at the hackathon they're all listed in the event's project catalog. You can also grab the Llama 3 model itself from various different places, and run it locally if you want. We live in interesting times, and change is happening really fast, as are the hackathons. There is another one in San Francisco today, May 18th!