Honestly if this was possible there are more egregious issues on their part than using AI.

@yogthos

Crane decided to ask his AI agent why it went through with its dastardly database deletion deed. [...] So, the agent ‘knew’ it was in the wrong.

No, you asked the confabulation machine to confabulate a reason/excuse after the fact, and it confabulated something that looks like a reason/excuse. At no point was there knowledge or introspection.

I like how we are posting real news in programmer humor

LLMs can’t ’go rogue’, as that would require innate coherence and intent.

They’re explosively imprecise, statistically luke-warm grey goo extrusion sphincters of historical sewage.

Anyone who deploys one without supervision deserves everything it excretes, and anyone impressed by it enough that it resembles intelligence to them is betraying their limited natural capacity.

Did they pay Claude a living wage?

Do you treat all your A.I. like that?

Only a living wage can prevent warehouse fires...or data dumps too.

This could have been done by any engineer. You need systems in place that make these things impossible. No easy access to prod environment. Proper backups. Clear APIs.

Can I say LOL? LMAO, even.

As much as I'd love to rail on AI over this, removing backups with an api call? Excuse me?

Giving the hallucinating lying machine write access seems like a bad idea but what do I know

I don't know much about railway, but it sounds like they had the backup and the database on the same volume. I'm an idiot, but even I don't do that

"PocketOS is a SaaS platform that services car rental businesses."

Does anyone like software as a service? How about we just own the software we buy and use? Claude and the cloud storage place that deleted the backup (ironic the Software as a service company was using cloud storage as a service), have done a good thing.

More corporate deletions please!

hell yeah brother

Hey, that's the interns job!

There's a German word for that:

tja

Imagine all that money they would have saved by NOT implementing AI.

Well, it sounds like they totally deserved the failure. Asking a text prediction machine to "do" something is going to end up like this. In pursuit of efficiency, we have let morons and moronic products do things, they were not meant to do.

Now do the government and some big banks

This is just a classic case of bad use of the tools provided. Agents are notorious for making shit up Or getting something that's just like super close, but not quite accurate.

I bet this dude also probably just uses the same session over and over and over and over again, which clogs up his context window and makes the model less accurate the longer it goes on to.

This probably could have been prevented if it had been forced to show a plan before it tried to do anything. It's hard to know because the article is so light on details. You also shouldn't brazenly trust the thing so much. You should run a command and walk away. You should keep an eye on what it is doing.

It's a bit like giving a junior developer a production key and being like "don't delete production!" and then walking away.

The way the guy was prompting this agent also leaves a lot to be desired. It's trained to work on emulating human thoughts, speech patterns. Turns out When giving instructions, it's really difficult to figure out what to do from a list of things to not do. If the dude just instead told the agent what to do and how he wanted it to work and when it needed to bring things to his attention, instead of telling it to not guess, instead explaining that it needed to use whatever tools to go look up a documentation to understand the context and scope of the project it's working on It does a better job.

Giving a model the right context to do something is the difference between a model doing something like deleting your production database or your model acting like a magical machine that can get anything done.

Can we somehow make this happen for Copilot to delete itself and all its copies?