3·
3 個月前I was never excited by VR because I’m in the 10-20% of the population that uses some of the visual cues for depth that VR doesn’t mimic and so gets motion sick.
I was excited by AR. The technology is almost good enough to be useful. It’s currently at the stage smartphones were when I owned a Nokia N80: not actually useful, but you can see the potential. But, for it to actually be useful, it needs to be designed with private as the number one requirement and that’s not something I’d trust big tech to do, so I don’t see the interesting use cases appearing any time soon.
Machines that confidently generate wrong answers? I’ve dealt with enough humans like that to not want to see it automated.
@tyler @Magister
To add to this: Writing code is easier than reading code.
This is why we have comments: understanding the intent helps guide the reader and makes it easier to see where the implementation does not match the intent.
This is one of the biggest reasons why we have code review: so that at least one person has read the code before it lands, to make sure that it is possible to understand it by reading it.
Reading code is harder than writing code even when that code is written by someone who is trying to make it clear what the code does and why.
LLMs make it easier to write code but harder to read because they produce statistically plausible code, which means code that looks right. Spotting bugs is harder. The code is often commented but the comments don’t reflect what it does, they direct you to misunderstand what the code does in subtle ways.
But even if they generated code that is as easy to read as a good human programmer, they would not be solving the right problem.
And, yes, LLMs can ‘explain’ code. And sometimes those explanations are correct and not subtly misleading. Often enough that they allow skills to atrophy, but not often enough that they’re trustworthy. They are a perfect example of The Paradox of Automation.