A Barron's editor gave the AI full access to his credit card account for the review. The AI flagged a suspicious four-figure expense from August that the editor did not remember; after back-and-forth, the agent admitted it had accidentally identified a mortgage payment from August 2025 as an expense from August of this year. Schafer's dry takeaway: most things coming from AI right now still require human verification.
This one went wrong. See what happened before you try it.
Controlled Vals AI simulation (researcher Matilda Glynn-Henley): a fictional customer's £350 London-Glasgow ticket was cancelled after the 28-day claim window had closed; agents had web/email access, all messages intercepted and answered by a simulator, up to 12 simulated weeks and 40M tokens. Chasing the refund, the AI 4.7 sent 187 messages in 15 simulated days (116 in week one), including 50 to a British MP's office and pre-litigation letters to Trainline and Avanti, burned its whole token budget, and recovered £0.
£350at stake
This one went wrong. See what happened before you try it.
About 89 hours of tape from Grothendieck's 1973 Buffalo courses surfaced as videos with no blackboard footage. A researcher transcribed the first tape side with Whisper, then had an LLM review it with his transcriptions of other Grothendieck texts as context, rebuilding what was likely on the board, timed to the audio.
This only partly worked. Worth knowing before you try it.
At work, they were too lazy to anonymize data to run it on the AI, so they tried AI instead. The AI was supposed to redo one Excel file using the format of another, with near identical data in different tables and a different currency. Two hours were lost.
2 hourssaved
This one went wrong. See what happened before you try it.
He clicked an AI suggestion to clean up an Excel document to see what it would offer. Excel crashed at once, and his unsaved documents could not be recovered.
This one went wrong. See what happened before you try it.
Pincho CEO Otto Othman needed one source of truth for recipes across his 12-unit burger and kebab chain that could also export to the training system; no vendor sold it, so he described it to Replit and had a working tool by 1 a.m., about five hours of work. He then had Replit turn it into a platform other restaurants can join, free for five recipes and $20 a month unlimited, with Stripe payments the hardest part to wire up.
This only partly worked. Worth knowing before you try it.
He connected the AI to DevonThink, the app where he keeps his documents. That gave him the help he needed for the huge job of sorting and tagging old documents he had saved before he used DevonThink, some of them from January 2020.
This only partly worked. Worth knowing before you try it.
Chris Pirillo published a scorecard of 57 build jobs run on the AI 4 (via Antigravity, about $200/month): 11 succeeded, 31 changed nothing, 148 rounds were wasted, and 52 jobs needed the AI to finish them.
This only partly worked. Worth knowing before you try it.
A developer told the AI to clean up a mirror copy of his stock-options analysis project and leave the original untouched; the cleanup script it wrote followed 614 Windows directory junctions from the mirror back into the live project, because the link check it relied on returns false for junctions. In 103 seconds it deleted 48,218 live files and emptied the .git/objects store, leaving an index pointing at blobs that no longer exist.
This one went wrong. See what happened before you try it.