Hacker News Reader

Hacker News

Comment by hcks | Hacker News Reader

Parent Full thread

hcks

It’s just a proxy measure for IQ really. They could have handed out Ravens matrices and got better outcomes prediction.

View on HN

I assure you, they would if they could, but its illegal.

A quick search indicates that it is not illegal in the US as a whole, but maybe California?

https://en.wikipedia.org/wiki/Griggs_v._Duke_Power_Co.

What's illegal is IQ/ability tests having disparate impact on particular racial groups. You can only use them if you can justify how they're related to the job (which is much easier for algorithms stuff in SE roles).

You can also train for IQ tests.

People misunderstand the reason for this. Test takers can improve, but only up to a point, and only because they getting comfortable with the test format, not because they're getting any smarter. Solution: give the candidates a few practice tests first.

This extends to every test though. My ACT score jumped massively on my second try. The reason: I didn't realize how much little time they gave test-takers, so I ran out of time on some portions. The second time around, I rushed through as fast as possible.

You're certainly not getting smarter. It's the test format that is not very reliable.

Is memorization how we should measure IQ? If so then LLMs have already far surpassed humans.

A LLM would do pretty/very well in a leetcode interview

In a "have you memorized the code to solve this problem" interview? Probably.

In a "can you solve this new leetcode problem that just came out and is not in the training data"? Awful.

There are probably about 30 or so leetcode problems and then endless variations made for those 30 problems. The best people on leetcode practice and memorize.

If you want to see an example of a new problem check out https://news.ycombinator.com/item?id=37910297. You'd think if leetcode makes you good at solving problems then it should be doable any higher ranking leetcode programmer.

> There are probably about 30 or so leetcode problems and then endless variations made for those 30 problems. The best people on leetcode practice and memorize.

The comment I replied to claimed that "LLMs would do pretty well at leetcode interviews". OpenAIs own report [0] shows how GPT-4 fails miserably at medium/hard leetcode problems that are not in the training data.

[0] https://cdn.openai.com/papers/gpt-4.pdf

I wouldn't consider doing 31/41 easy problems, 21/80 of medium problems, and 3/45 of hard problems a failure. GPT-4 wasn't built to solve these problems but can still do 3 of 45 hard problems. Hell I don't know if you sat down 45 random programmers if 3 of them could solve those 3 problems GPT-4 was able to do, and nobody could solve them in the time it took GPT-4.

Hell I wish they would tbh. I have a decent IQ (140), but can't (or just unwilling to) grind leet code

An IQ score of 140 is considered near-genius or genius level and places you in the top 0.25% (99.75 percentile) of the population.

Actually 137 if test was accurate, I still don't do well with leet code whiteboarding though, shrugs.

Maybe you just need to find something you enjoy doing, programming might not be "it"?

Or do you do well solving the problems on your own and just have trouble when being interviewed?

The latter, take home projects I do well on (projects, not HackerRank quizzes)

That's a believable score around here. There are over 8 million Americans with that score (assuming OP is American) and this place probably has a disproportionate number of high-IQ people posting.

My IQ is 125 and I feel the same......I love solving problems but only if I feel like they are real world problems.

I did extremely well in algo classes in school.....but since then I have built actual applications that run businesses and haven't used algorithms at all.

Grinding leetcode feels like running on a hamster wheel for me.

Reply on news.ycombinator.com