ZWQbpkzl [none/use name]

the specs of an old piece of equipment and how to repair are suspect.

That’s the most dangerous thing you can ask an AI to do. At best its going to just do a google search for you and then mangle the results. If you ask an AI for the dimensions of a specific product it will get it wrong.

Someone should make an AI that just googles shit and then shifts through the SEO slop to find original sources.

Self hosted Claude? I think they’re self-hosting mistral.

That could be fixed by simply logging the prompt and code executed. Maybe also give each prompt/response a reference ID and demand that in tickets. The nightmare would be actually reading the code the AI generated.

You’re being silly. Clients can only prompt the AI and the AI has restricted read-only permissions on the database. Slap on a execution timeout to cover if the AI wrote an expensive query.

The real concern is the AI getting a query subtly wrong and giving the client bad info. That gets “covered” by some flimsy disclaimer.

Fortunately not that bad but the people who are using it do get praised while being a massive burden on everyone who has to review the code or worse, documents.

I did see one clever usage to basically replace most front end devs with AI:

• client asks LLM question about our data.
• LLM generates code to query our data and display charts
• client sees results

All runs with just mistral:8b. Very flexible solution compared to having front end devs constantly iterate on a UI monstrosity meant to serve every single clients needs. Of course this assumes the AI is writing the query correctly.