Two Hackathons in Vienna, and My Code Stays Home for Now

TL;DR

Two hackathons, one week apart, both at the House of Innovation Vienna (HOIV): HACK_002 this Saturday, Hack-Nation 7 on 3 October. For both, I'm assuming the project has to be built during the event. So beforehand I'm doing everything that isn't the solution. The repository doesn't exist until after the opening. Whether that's enough, I'll know on Sunday.

Last time, no sound

In March I went to my first hackathon, HACK_001 of the Telos Circle series, in the same building. There were four of us. The first idea got binned after 2.5 hours. And in the end our submission video had no sound.

We only noticed when we played it back. So down to the jury, ask for an extension, record the whole thing again.

On Saturday I'm back. Same series, same building, this time solo with my agents. Maybe someone joins at team formation, we'll see. The sound, at least, gets tested this time before it counts. And honestly, I'm really looking forward to it.

Two hackathons, and this time it's measured

HACK_002 runs on 26 and 27 September. Build after the opening on Saturday, submit early on Sunday morning. That's 19.5 hours, not 24.

Six tracks in two blocks, and every one of them states what gets measured, almost always against a baseline. Some want at least five test cases, others explicitly want to see failures.

At my first hackathon, one sentence on the benefit was enough. This time it has to be measured. I like that.

And a week later, straight into the next one: Hack-Nation 7 on 3 and 4 October, a global event with hubs in several cities plus online. One of them is the HOIV again, and that's where I'm registered.

Where the rules go quiet

Can you write code in advance? For HACK_002, nothing has been published on that. HACK_001 was clear. Its Devpost page said: "You can only submit a project that is built during this time period."

The HACK_002 event page allows third-party tools and datasets, as long as I can explain access, licensing and evaluation. Hack-Nation advertises "Build real AI products from scratch in 24 hours". Its originality rules live on the submission platform, which isn't open yet.

Where a rule is missing, I take the stricter reading: nothing of the project exists before the start. I'll still ask at the opening. The submission is a demo video plus a ZIP of the Git repository, so the jury sees the commit history. A first commit from the Wednesday before would speak for itself.

At BitGN it was the other way round

At the BitGN Personal Agent Challenge in April the dev benchmark was available in advance. From 4 April I spent a week building the agent; on the day it was three hours on site. The result: rank 15 of 222 worldwide, built solo, plus rank 1 at the Vienna hub. Back then, preparation meant code, and that week is what made the difference.

This time there is no such week.

Everything that isn't the solution

Held back until the start: the repository and the first commit, every line of project code including scripts, and the README. Code from my own projects doesn't come along at all. Everything else can happen beforehand.

Accounts first, because that's where the surprises are. One sign-up got stuck on a captcha. Another recognised my agent as a bot and blocked it, rightly so. I'd rather sort that out on Wednesday than on Saturday. The keys live in a local file outside any repository.

Then the research on rules, data sources and licences. Does everything that's recommended still hold? Not quite. One of the recommended APIs has changed its pricing model since. Another quietly ignores a filter if you only set half of it, and the response still looks plausible. Nothing turns red, it's just wrong.

The evaluation is fixed before there's a single number. As text: what I measure and against what. Basically a pre-registration. Set it after you've seen the numbers and you'll set it to fit. Same at a hackathon as in my day-to-day work, where I'd rather verify than type. I'll show what's in it after the event.

Still to do: the video template and a sound check. Record one take with sound, then actually play it back. Sounds trivial. It wasn't in March.

The pattern comes along, the code stays home

I run agents that work for me day to day. Over several generations a pattern has settled in. Consequential actions wait for my approval. Untrusted text is data to the agent and never an instruction.

That knowledge comes with me. After the start I rebuild the pattern in a fresh, MIT-licensed repository. I know roughly how long that takes, I've done it before. Public libraries only go in as declared dependencies and are listed in the README.

The same goes between the two events. What I learn at HACK_002 comes with me to Hack-Nation. The repository stays where it is.

Decide before it's three in the morning

Dropping the first idea after 2.5 hours was the right call in March. Just an expensive one. This time I do the dropping beforehand, on paper. A short list of ideas, scored in advance, each with the sentence that would refute it. The choice happens at the opening, once the sponsor challenges are on the table.

I've also decided in advance where I cut scope instead of building on. In the middle of the night I don't want to be negotiating that with myself. Which points those are, and whether I stuck to them, goes in the recap.

At Hack-Nation most challenges only arrive at the kickoff. So I'm bringing a rule that lets me choose in a few minutes.

Three questions for Sunday

Three questions come with me to the HOIV. Did I miss something during the night that I would have been allowed to prepare? Is 19.5 hours enough for a measured result with a baseline? And what does the opening say about prepared code?

Then comes the recap, with what held up and what didn't. If you're at the HOIV this weekend too, come say hi! And how I work with agents day to day is on the home page.