They are, at best, an unreliable tool you need to babysit. And because of that, you need to be able to do what you’re asking the tool to do so you can verify it actually did what you asked.
That’s is not at all how these tools are being marketed, or how many are using them. So the current result is 90% of usage is hype and slop.
Nah, just build a harness that validates the output of one model by running it through the same model again to check for hallucinations… And to make sure that second pass isn’t hallucinating, uh… run it through a model a third time to check the second isn’t hallucinating.
They are, at best, an unreliable tool you need to babysit. And because of that, you need to be able to do what you’re asking the tool to do so you can verify it actually did what you asked.
That’s is not at all how these tools are being marketed, or how many are using them. So the current result is 90% of usage is hype and slop.
Nah, just build a harness that validates the output of one model by running it through the same model again to check for hallucinations… And to make sure that second pass isn’t hallucinating, uh… run it through a model a third time to check the second isn’t hallucinating.
/s
Surely it will get better each time! Just like photocopying a copy!
Only by constantly re-photocopying something can something truly timeless ever be reified