The same way you control for hallucinations with say Medium articles or Stackoverflow answers. You verify the answer.

Being given some code and reviewing it is a lot quicker than writing that code. Copilot is great. Half the time it spits out the wrong answer but you can see what it was “thinking”.

Phind giving you an answer and references lets you quickly double check. Sometimes it hallucinates but the answer and references combo is much better than a Google search which in turn is better than nothing.

Of course I recommend people not use AI for everything. I will go straight to MDN for any WebAPI question and use —help as my first port of call on the command line. This is like your L2 cache as a developer. Using AI for everything is like swapping to disk for everything.

You use your critical faculties, very important things must be verified from other sources, unit, integration, numerical tests can evaluate code and results.

Sometimes you need to go back and forward a bit, I tell ChatGPT it’s wrong and the error message and then it spits out the correct result, sometimes I need an algorithm tweaked because it has assumed a wrong constraint on the problem, again just explain it clearly and unambiguously and it will make corrections, there’s only been maybe 1/40 problems I couldn’t get a correct answer on after (sometimes a lot!) of back and forwards.

I am not looking for a perfect oracle, I am looking for something to write 80% of the code and then I’ll fix it up. It’s still way faster this way, especially in domains I don’t know, E.g. I just learned CUDA with ChatGPTs help.

It’s not perfect, and neither am I, but it doesn’t have to be perfect to be useful, you can get to millions in revenue through 80% solutions

I don’t understand the obsession with hallucinations. People are completely wrong all the time; your co workers ‘hallucinate’ answers all day long. You verify and correct; what’s the difference?

Ask for sources. Ask if what it has already told me makes sense. Knowing that hallucinations happen and not to blindly trust.

Learn the limits of your tools.