Bizarre tasks I shipped with AI

The place are all of the AI-generated tasks? It is a frequent query from AI skeptics: if LLMs are so good at writing code, the place is the tsunami of latest AI-generated apps, companies and video games?

I personally don’t discover this to be a lot of a paradox. Writing code is barely one of many bottlenecks concerned in really transport a brand new product, in any case. It’s additionally inconceivable to speak concerning the paid work I’ve performed with AI (you’ll merely must take my phrase that it’s elevated my productiveness). However one factor I can do is share a listing of private tasks I’ve constructed with AI within the final twelve months.

I undoubtedly wouldn’t have performed all of those by hand. I might need discovered the time to do one or two of them, however primarily based on my pre-AI monitor file they might in all probability have stayed within the “GitHub repo with a couple of commits” stage. This checklist is a form of existence proof: a bunch of bizarre tasks, helpful to at the least some folks, that might not have existed with out AI help.

Skifreedle

Most just lately I’ve constructed skifreedle.com, a daily-game model of the basic Home windows SkiFree recreation (i.e. “like Wordle, however for SkiFree”). The code for that’s right here.

I take pleasure in coding small net video games by hand, however undoubtedly wouldn’t have had the time to wire up all of the totally different SkiFree objects or construct neat options like a ghost of your quickest run. I additionally tried out plenty of totally different visible themes for the sport UI earlier than touchdown on one thing I appreciated. If I’d performed this by hand, I’d have solely had time to check out two or three totally different seems, as an alternative of fifteen or twenty.

I’m very pleased with how this turned out. I’ve been having fun with competing towards my brother to get higher instances, since each of us have plenty of nostalgia for the unique SkiFree recreation.

Autodeck

Final 12 months I constructed Autodeck! I wrote a weblog publish about this earlier than, however this got here from my accomplice wishing there was some technique to robotically generate Anki playing cards about random subjects she wished to find out about. It ended up being comparatively simple to arrange an countless feed of auto-generated spaced repetition playing cards:

I arrange Stripe funds for this one, extra as a result of I used to be apprehensive about somebody operating away with my Groq stability than as a result of I wished to make cash, however I used to be pleasantly stunned to see a bunch of individuals really use this. Over 5 hundred folks have tried it out, with sufficient paid subscribers to cowl inference and internet hosting.

I might need constructed this with out LLM help, however I nearly definitely wouldn’t have deployed it as a web site. The effort of establishing a database and Stripe would have simply been an excessive amount of work.

Infinite Wiki

I additionally constructed an AI-generated countless wiki. I wrote a weblog publish about this one as effectively. Like Autodeck, I used to be fascinated with the concept of non-chat interfaces for LLMs, and I assumed a wiki-based method the place you work together with the mannequin by clicking hyperlinks was fairly cool.

I discovered the laborious means that placing a LLM era name on the tip of an everyday hyperlink was a foul concept: scrapers would exhaust my inference finances rapidly. I ended up faking the no-article-exists-yet hyperlinks with JavaScript, which at the least thus far has defeated scrapers. Individuals nonetheless e mail me about Infinite Wiki, and there are over 280 thousand pages generated.

My unique aim was to see in case you might ultimately generate a web page for Neon Genesis Evangelion, beginning on the root web page and solely following hyperlinks (form of like wiki golf). I used to be profitable! You may learn the “Evangelion Anime” web page right here.

Nearly precisely a month after I launched Infinite Wiki, xAI launched Grokipedia. Clearly they didn’t plagiarize me. It is a very simple concept to have, and my website was not the primary infinite wiki (although I feel it was the primary one the place you needed to uncover new pages by clicking on hyperlinks). However it did take a few of the shine off.

VicFlora Offline

I constructed a PWA that caches the VicFlora plant identification database so it may very well be used with low or no web. This was extra of a utility venture for my accomplice, who likes vegetation and infrequently goes on subject journeys the place web is spotty.

I’d undoubtedly not have performed this with out LLMs. It was fairly troublesome to scrape the essential dichotomous key from the VicFlora web site: their API documentation was outdated, there have been a number of potential pathways for fetching knowledge (most of which weren’t practical), and the format of the info I did handle to fetch was laborious to parse. I feel I might have performed it, with sufficient effort, however it could have been a considerable quantity of labor.

I’m very pleased with how this turned out. It’s not good, but it surely’s practical, and I’ve even had the occasional Victorian botanist e mail me with bug reviews or function requests, so it’s clearly seeing just a little little bit of utilization.

Different tasks

I did a bunch of different stuff that doesn’t essentially rise to the extent of a “deployed venture”: my gh-standup GitHub CLI extension to robotically generate a standup report, which has simply over 100 stars, my (low high quality) picture geolocation benchmark, which I blogged about right here, or my talent for extracting options from open-source fashions.

There will not be a flood of AI-generated corporations (but), however at the least for me there’s been a flood of small, bizarre tasks that might not have existed with out important LLM help.