Real things everyday people got it to do, and how to try them yourself.
A random mix from the whole library. Shuffle for a new set.
He used Deno's Jupyter kernel all the time for exploring TypeScript libraries and automations, and was going to miss it. He pointed the AI at the problem and the result is cdp-kernel, Deno's Jupyter kernel extracted to run notebooks on any V8 runtime, now on GitHub.
He said he built his setup with AI and local models. He told the agent to use a GitHub tool to download his Garmin data automatically and store it in a local SQLite database, and the data includes sleep and HRV, not just workouts.
DeepSeek upgraded gnarly, exploitable PHP 7.x code to fully patched PHP 8.3 in about 15 minutes for a few dimes of tokens.
He uploads his own data and points the AI to specific sources and documents, with examples he made before to show the format he wants. The AI then builds the tables and graphics for his Education Rights Project, public pages on education rights and funding in US states, with policy briefs, slide decks and interactive widgets.
He used GPT-6 Astra to model a small weed holder and 3D print it. Talking to the computer and getting a real object back felt to him like a Star Trek replicator.
He is treasurer for a nonprofit food co-op. He gave the AI the co-op's financial statements, some notes on what he wanted to explain, and the logos, and 20 minutes later he had a ten slide deck and separate printouts for the annual meeting.
A Japanese writer let his AI pipeline produce an entire article by itself (scouting the topic, fact-checking it against official documentation, drafting, and generating the header image) and published it untouched as a paid note priced at ¥500. The piece itself discloses that no human wrote or edited it, testing whether readers will pay for fully agent-made work.
He had the AI translate his MaPollution app into Italian, and it took 30 minutes. He can now offer his free service to people who speak other languages.
A statistician had argued that class balancing helps mini batch gradient descent. He had the AI Astra design and run a simulation to test the claim, and the results partly corrected him: oversampling sped convergence but hurt probability accuracy, while importance weighted oversampling kept most of the speed and the accuracy. The code was cross checked with AI, and he posted the corrected conclusion as an answer.
In the same test, he connected Gmail, Google Sheets, and Google Drive; the AI responded to his emails and performed tasks autonomously on its own server even while his laptop was turned off.
Every idea is a real post from someone who tried it, summarized in our words, with a link to the original. App names are swapped for plain words.