3 ms·

> > You can't trust the AI's code! > Yeah, but you can't trust your own code! No, really, you can't trust the AI's code. It spits out code that references fun

> > You can't trust the AI's code! > Yeah, but you can't trust your own code! No, really, you can't trust the AI's code. It spits out code that references functions that don't exist. It spits out code that does something you didn't ask it to do, or doesn't do something that you did. It spits out code that does something vaguely related that the Stack Overflow guy it's cribbing from did, but you aren't doing. It spits out code that confuses your syntax for JS's halfway through. You can skim other people's code; you know what common failure modes are at their experience level, and you know where the complicated bits are that will house the bugs if there are any. You cannot skim the AI's code. Every word of it must be examined. You must be in full reviewer mode all the time, which is an unproductive state to be in when actually writing code, which makes specifically Copilot less useful. You cannot use it to replace a no-code tool, because you must understand the language it's emitting. I find that proponents shift between whether its ideal use-case would be Copilot or no-code; any flaws with the one approach get interpreted from the perspective of the other, where they can be dismissed.

LVB 4y ago
Having used Copilot for quite a while now, my feeling is that it simply is another (and for me faster) way to work. I've become used to its strengths, limitations and quirks, and find that even being in "reviewer mode" can still be quite a bit faster. There is a range of performance between stuff it reliably gets near perfect, and areas where it is so helpless that I just turn it off for a while. But I'm still in control, not blindly hitting tab. If I thought it was just fighting me or wasting my time I wouldn't be using it after all these months.
jgerrish 4y ago
> You cannot skim the AI's code. Every word of it must be examined. I understand what you're saying. But be careful of this argument, it's susceptible to safety counter arguments. "Well, this person thinks we shouldn't carefully review all code. Do we want him working on our super-critical FarmVille clone?" (That part is of course ironic. It changes if you're working on medical devices.) A closely related argument is that the cost of the back-and-forth with a code submitter who submits buggy code is drastically higher. The hand-holding and teaching and encouragement is very expensive. Maybe that's acceptable to the person who posted this article. But it's a possible avenue of resource exhaustion attack if you're not careful. I'm sure AI will improve, with loving prompters. But some of the discussions around it seem dishonest, which is troubling (not you of course).
- jgerrish 4y ago
  And I could see some social media companies incentivizing longer prompt sessions for engagement and eyeballs. Not necessarily OpenAI. Aka, the "Oopsy, I did a poo-poo! Dear User pays more attention to me when I smear it on the wall." mechanic. Sorry for the metaphor, just registering awareness.

LVB 4y ago

Having used Copilot for quite a while now, my feeling is that it simply is another (and for me faster) way to work. I've become used to its strengths, limitations and quirks, and find that even being in "reviewer mode" can still be quite a bit faster. There is a range of performance between stuff it reliably gets near perfect, and areas where it is so helpless that I just turn it off for a while. But I'm still in control, not blindly hitting tab. If I thought it was just fighting me or wasting my time I wouldn't be using it after all these months.

jgerrish 4y ago

> You cannot skim the AI's code. Every word of it must be examined. I understand what you're saying. But be careful of this argument, it's susceptible to safety counter arguments. "Well, this person thinks we shouldn't carefully review all code. Do we want him working on our super-critical FarmVille clone?" (That part is of course ironic. It changes if you're working on medical devices.) A closely related argument is that the cost of the back-and-forth with a code submitter who submits buggy code is drastically higher. The hand-holding and teaching and encouragement is very expensive. Maybe that's acceptable to the person who posted this article. But it's a possible avenue of resource exhaustion attack if you're not careful. I'm sure AI will improve, with loving prompters. But some of the discussions around it seem dishonest, which is troubling (not you of course).

jgerrish 4y ago
And I could see some social media companies incentivizing longer prompt sessions for engagement and eyeballs. Not necessarily OpenAI. Aka, the "Oopsy, I did a poo-poo! Dear User pays more attention to me when I smear it on the wall." mechanic. Sorry for the metaphor, just registering awareness.