A random mix from the whole library. Shuffle for a new set.
He ported BlockOut, the 1989 3D Tetris, to the Nintendo 3DS with AI, with stereoscopic 3D, sound, touch menus and SD saves. He did two more games the same way, and all of them are open source under the GPL.
He asked the AI to set a timer for 4 minutes. The AI said it would set the timer, but no timer was set at all.
A screen-recorded walkthrough of an app called 'Song List' inside an AI app AI shows managing a personal music repertoire: 264 total songs, 76 marked 'In show,' 188 'Needs work,' across 76 artists. The presenter navigates 'In show' and 'All' views and uses the search bar.
Runs a personal-assistant session via Signal that tracks tasks and delegates to per-project agent sessions on other machines; it is now his only direct interface.
His second attempt at building software by voice with an agent produced Feather, a working Nostr client. He had to talk it out of adding an app store, a login page, and a settings menu it kept trying to build, and he paid about 3,000 sats in lightning tips to people who reviewed the code.
The AI interpreted scanned old service manuals and circuit diagrams, ran simulation experiments to verify manual claims, cross-checked datasheets and papers, caught typos and his own misreadings.
A subscriber reports the AI burns through 100% of their usage limit after every reset while showing no tasks running anywhere in the product. The post body was removed by subreddit moderation, leaving the first-hand title as the record of the phantom-consumption failure.
The AI and Antigravity took an existing STL and swapped a fan-shroud section for another design by directly manipulating the STL geometry in Python; it went perfectly.
Built an app with AI subagents in an afternoon that reads, maps and translates all 387 labels of an 1819 French map of the Bosphorus.
Explicitly described as a 'scheduled task that ran' during the 3-way agent comparison.
A hobbyist designs his own furniture in OpenSCAD with AI, reusing code from earlier pieces. A new design takes a couple of hours including the PDF blueprints, drawn like shop instructions, that he prints so the person cutting his wood panels cannot go wrong.
He spent an evening trying to teach the AI to drive a car. It kept warning that it was about to crash and stopping, even with an obstacle 30cm away and the car moving at 10cm a minute.
He wanted a control monitor to fit under his screen and follow the angle and shape of his Stream Deck. He uses AI with reference photos to get the base shape, adjusts sizes after measuring with a caliper, printed a quick baseplate to test the fit, and fixed a 1mm clearance problem with the lock after more measuring.
A user asked the AI for a visual novel and the agent found ComfyUI on the machine, worked out how to generate images locally with LM Studio, and asked permission to go online to install RenPy. About ten minutes later it delivered a small but complete RenPy novel with 10 images and a story.
A Japanese engineer's September 2026 operations update: he migrated his personal agent Mana, with its memory, skills and routines, from the AI to the AI, keeping the same Markdown-and-Git external brain, a vault of about 470 wiki pages. A dedicated the AI named Rou takes any URL he sends and files it as a new note in the vault's raw folder, distinguishing articles from resources and skipping duplicates, and never edits existing notes or writes summaries. Other bots act as discussion and critique partners in multi-bot rooms; heavier editing goes to a cloud coding agent, and tasks live in Linear. He published Rou as a public the AI template anyone can copy.
A tech writer had the AI search five years of his email for signup and password reset messages. The AI built a spreadsheet of the accounts it found, with breach notes and a deletion link for each service, and tagged at least 50 accounts.
The AI took a webinar email and made a calendar event for it on its own. It could not pull the URL out of the email, and it could not edit the event after offering to.
A man asked the AI to schedule daily backups of his agent folder with a careful keeping rule: three weeks of daily copies, four Saturday copies, and the first copy of each month. The AI also copies them to a storage server in a different building and runs a health check that alerts him only when something is failing.
He bought a Humble Bundle of Android games 10 years ago, but they are 32-bit and will not run on his 64-bit phone. He asked the AI to decompile and recompile them, and the games work on his phone again.
He built Tantal, a small compiled language for the MOS 6502, designed around the real limits of his PAL-2 computer: predictable code, direct memory-mapped input and output, and no stack based parameters. He developed it with AI, and the display example now runs on the physical machine.
His role was cut on Tuesday morning, 22 September. That evening he asked the AI for something fun, gave it one standing instruction to go forward and only pause when his input was needed, and by 6:37 the next morning a three stage beat em up about the job hunt was live on the internet, with his real CV linked from the end card.
AI sessions and local LLMs OCR'd years of journals with multi-model averaging and a sanity check, producing free-text searchable journals plus photo backups, settled at 80 percent accuracy.
The AI fixed Steam games and HiDPI and monitor issues on his Linux machine and powers a recurring shell script that checks updates and looks online for potential issues.
The AI told him swap memory on his server was high, found the cause, and had the fix ready for him to approve. The AI or the AI now connects to the box, upgrades services, and sorts out problems, so his site and newsletter run on a bare DigitalOcean droplet instead of a pricier platform that makes hosting easy.
Tell the AI to set up an app on his server and it does it all and leaves a documentation note in his Obsidian vault; setup effort near zero.
Runs LLMs to audit the math in claimed arXiv solutions for his ranked open-problems list; audits already found six faulty proofs that had caused problems to be wrongly marked resolved.
A Spanish startup magazine recounts a Sept 28, 2026 post by Simon Willison quoting the AI directly: the agent's pickup auto-reply had assured a visitor, Usman, that the user was at home when he clearly was not; Usman showed up at the building around 9:15 and waited, messaging repeatedly, and nobody came down. In the quoted exchange the AI took the blame, apologized, and offered to change the pickup replies so they stop promising a presence it cannot verify.
Saving his Bluesky threads as Markdown files for Obsidian was a chore he kept doing by hand. He asked the AI to build a converter, and the finished tool is up on his site for anyone to use.
Fable reverse engineered the Satisfactory blueprint file format and wrote a reader, a graph auditing tool, and a website; work that would have taken a couple of weeks by hand.
Eric Siu (verified, 120K followers) demonstrated the AI’s Instagram integration reading his bookmarks and follow graph, then generating batches of tailored video outlines — '60-Second Playbook,' 'Story Cut,' 'Debate Cut' — each with hooks, TL;DR lines, talking points, editor notes, and links to past successful reels. He said it replaced a full-time person's scripting-research role.
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.