Pixie: Entry to Mistral x a16z Hackathon

Oct 21, 2024 • Yousef Amar • 4 min read

This post documents my participation in a hackathon 2 weeks ago. I started writing this a week ago and finished it now, so it's a little mixed. Scroll to the bottom for the final result!

Rowan reached out asking if I'm going to the hackathon, and I wasn't sure which one he meant. Although I submitted a project to the recent Gemini hackathon, it was more of an excuse to migrate an expensive GPT-based project ($100s per month) to Gemini (worse, but I had free credits). I never really join these things to actually compete, and there was no chance a project like that could win anyway. What's the opposite of a shoo-in? A shoo-out?

So it turns out Rowan meant the Mistral x a16z hackathon. This was more of a traditional "weekend crunch" type of hackathon. I felt like my days of pizza and redbull all-nighters are long in the past by at least a decade, but I thought it might be fun to check out that scene and work on something with Rowan and Nour. It also looked like a lot of people I know are going. So we applied and got accepted. It seems like this is another one where they only let seasoned developers in, as some of my less technical friends told me they did not get in.

Anyway, we rock up there with zero clue on what to build. Rowan wants to win, so is strategising on how we can optimise by the judging criteria, researching the background of the judges (that we know about), and using the sponsors' tech. I nabbed us a really nice corner in the "Monitor Room" and we spent a bit of time snacking and chatting to people. The monitors weren't that great for working (TV monitors, awful latency) but the area was nice.

Since my backup was a pair of XReal glasses, a lot of people approached me out of curiosity, and I spent a lot of time chatting about the ergonomics of it, instead of hacking. I also ran into friends I didn't know would be there: Martin (Krew CTO and winner of the Anthropic hackathon) was there, but not working on anything, just chilling. He intro'd me to some other people. Rod (AI professor / influencer, Cura advisor) was also there to document and filmed us with a small, really intriguing looking camera.

We eventually landed on gamifying drug discovery. I should caveat that what we planned to build (and eventually built) has very little scientific merit, but that's the goal of a hackathon; you're building for the sake of building and learning new tools etc. Roughly-speaking, we split up the work as follows: I built the model and some endpoints (and integrated a lib for visualising molecules), Nour did the heavy-lifting on frontend and deploying, and Rowan did product design, demo video, pitch, anything involving comms (including getting us SSH access to a server with a nice and roomy H100), and also integrated the Brave API to get info about a molecule (Brave was one of the sponsors).

You're probably wondering what we actually built though. Well, after some false starts around looking at protein folding, I did some quick research on drug discovery (aided by some coincidental prior knowledge I had about this space). There's a string format for representing molecules called SMILES which is pretty human-readable and pervasive. I downloaded a dataset of 16k drug molecules, and bodged together a model that will generate new ones. It's just multi-shot (no fine-tuning, even though the judges might have been looking for that, as I was pretty certain that would not improve the model at all) and some checking afterwards that the molecule is physically possible (I try 10 times to generate one per call, which is usually overkill). I also get some metadata from a government API about molecules.

On the H100, I ran one of those stable diffusion LORAs. It takes a diagram of a molecule (easy to generate from a SMILES string) and then does img2img on it. No prompt, so it usually dreams up a picture of glasses or jewellery. We thought this was kind of interesting, so left it in. We could sort of justify it as a mnemonic device for education.

Finally, I added another endpoint that takes two molecules and creates a new one out of the combination of the two. This was for the "synthesise" journey, and was inspired by those games where you combine elements to form new things.

Throughout the hackathon, we communicated over Discord, even though on the Saturday, Rowan and I sat next to each other. Nour was in Coventry throughout. It was actually easier to talk over Discord, even with Rowan, as it was a bit noisy there. Towards the end of Saturday, my age started to show and I decided to sleep in my own bed at home. Rowan stayed at the hackathon place, but we did do some work late into the night together remotely, while Nour was offline. The next day, I was quite busy, so Rowan and Nour tied up the rest of the work, while I only did some minor support with deployment (literally via SSH on my phone as I was out).

Finally, Rowan submitted our entry before getting some well-deserved rest. He put up some links to everything here, including the demo video.

Not long after the contest was over, they killed the H100 machines and seem to have invalidated some of the API keys also, so it looks like the app is also not quite working anymore, but overall it was quite fun! We did not end up winning (unsurprisingly) but I feel like I've achieved my own goals. Rowan and Nour are very driven builders who I enjoyed working with, and CodeNode is a great venue for this sort of thing. The next week, Rowan came over to my office to co-work and also dropped off some hackathon merch. I ended up passing on the merch to other people, as I felt a bit like I might be past my hackathon years now!