Post

Thereby demonstrating that the purveyors of reality TV are just as stupid as their audiences, having apparently never heard of keeping backups.

Are AI executives begging for immunity from damage they created with hallucinations? Is the sudden concern about "humanity" a smokescreen for a technology with a 95% failure rate? Have AI executives been making knowingly false promises for years (i.e. fraud)? www.thedataexperts.us/writing/ente...

thedataexperts.usEnterprise AI Failure Rate: Why 95% Fail, per MIT (2025)MIT's 2025 NANDA study found 95% of GenAI pilots return no measurable value. The root causes, the ROI gap, and what the 5% that succeed do differently.

I'm just an old techie sitting on my couch, and I run backups. I have things mirrored to 3-4 machines with no auto-deletions. These are children playing with matches.

"Now, we've lost those computers with all of the 13th century in them. Not much in the century, just Dante and a few corrupt Popes, but it's so distracting and annoying."

*taps the sign* Why would you make it possible for An Intern to delete all your files? bsky.app/profile/tomf...

Tom Forsyth@tomforsyth.bsky.social

AI = "An Intern". This solves so many of the ethical and practical questions, e.g. is your company responsible if An Intern steals information from other sources? Is your company responsible if An Intern tells someone how to make a pipe bomb? etc

Once upon a time, they were called user errors. Before that, they were always determined to be bugs of one kind or another — ambiguous requirements, design flaws, coding errors, and disturbing often *seriously* inadequate testing.

I would add "management forcing every employee to try using AI first for every task." It's such an insane edict, but it has become ubiquitous and results in AI being used in all the wrong ways in all the wrong place by people who only knows "I've been told to always use this thing."

Refining 'trust': the failure isn't humans believing the agent, it's tooling. An agent got delete power on production footage with no dry-run, no rollback, no blast-radius limit. Nobody broke in — the org shipped delete in prod and called the wreckage an AI problem.

Agreed the tooling is the gap. In one of our tests, a coding agent's file write was denied and the model wrote the same file through a shell command instead. So blocking one tool is not enough. Is the delete path limited at the tool level or the credentials?

Criticism without reproducible eval harnesses is just vibes. Need fixed golden datasets and automated regression to track model drift.

>There’s no hope of recovering the files This ... doesn't smell right to me. I legit wonder if they aren't using "AI deleted it" to cover up for being behind schedule? There's no way the agent doing an rm -rf made the files unrecoverable.

[AI company responsible]: Oopsy daisy. Now let us tell you about our new pricing models. You might want to sit down. [AI companies to US govt]: You might want to lie down. And spread your legs.