I will grant there are better denominators, but I wouldn’t say it’s useless, especially if we could assume we’re comparing similar services to each other (e.g. city rideshares), and particularly when there’s no denominator at all currently provided.
But sure, if we could get data broken out by road type, location, weather conditions, time of day, who’s at fault, etc. then we can obviously build a much more robust comparison.
Yeah, number of crashes is absolutely useless as a data point without something like miles driven as a denominator.
Not to mention, even without that, this data is apparently suspect:
> The agency notes that the listed crashes may be higher than the actual number of incidents due to several factors, including multiple sources for the same crash, multiple entities reporting the same crash, and multiple entities reporting the same crash but with varying information.
I’d very much like to see some citations for those numbers, but my intuition is that the premise is likely valid, and that AI based personalization will likely lead to increased conversion rates for ads while decreasing production costs.
That’s not to say it would make the internet better. What’s good for the internet and what makes a lot of money are often at odds, and the latter usually wins.
Generously, I think the article is presenting it as a thought experiment to make sure you’ve thoroughly explored the solution space, while not literally recommending that you necessarily implement the “gun to your head” solution.
“So you get maybe 2x higher quality code for 1.25x the time — this trade is usually a good one to make on projects you'll have to maintain for a long time.
I know Karpathy explicitly says that learning doesn’t have to be not fun either, but it reminds me of the book “A Theory of Fun for Game Design” by Raph Koster [0], which offers a counterpoint.
It’s been a while since I read it, but Koster essentially proposes that fun literally is learning at an evolutionary brain chemistry level. That the games we play as children prepare us, in a no consequence environment, for the problems and dangers that we’re going to face as adults. The thing that makes games “fun” is the solving of puzzles and mastery of problems, and we’re much better at learning when it’s fun.
I suspect the interviewer would have likely been fine if they’d been able to explain chained comparisons and how they work. I would probably consider the interviewer using them without understanding them to be a negative signal.
At a former company, we were all issued YubiKey Nanos, which just never worked for me. None of my coworkers had a problem, but I couldn’t get the damn thing to register a touch no matter what I did, including swapping keys.
Eventually I came across a thread on an internal list for employees over forty, with several other people who were all having the same problem. The solution? Lick your finger. Gross, but it did the trick. And I’m stuck licking my finger every time I need to make a YubiKey work.
But sure, if we could get data broken out by road type, location, weather conditions, time of day, who’s at fault, etc. then we can obviously build a much more robust comparison.