Hacker News Reader

Comment by drooby | Hacker News Reader

The amount of compute required for R1 to determine that strawberry has 3 Rs is hilarious..

I'm not so convinced that the Nvidia panic is justified.

The Nvidia panic stems, in part, from the possibility that this is just the first of many potential significant optimization leaps.

Idgi, optimization leaps are virtually guaranteed.

I think this is a deeper question about the bounds of human desire... which seem virtually limitless. We seem to have an unlimited appetite for answering questions on the complexity of existence. Pair that with arms race issues, and you have an obvious need for massive compute regardless of how efficient the algorithms are.

That's fair but look at the assumptions currently built into NVDA's stock price. Hard to say for certain but to some it's priced as if it has a monopoly on all things AI for a decade. If you can spend $6K (and less in the future) on a full system that runs a model for you where does that leave the assumptions baked into that stock price? I'm too lazy to have come up with a model myself but off the cuff it seems like there might be some dislocation in a lot of assumptions. I dunno. I'm kind of a luddite because of the .com bubble and this has the same feel.

Jevons’ paradox seems relevant:

https://en.wikipedia.org/wiki/Jevons_paradox

> The amount of compute required for R1 to determine that strawberry has 3 Rs is hilarious..

How about the amount of compute to take a buggy hextree implementation I've been poking at for a few years and completely rewrite it into a fully functioning implementation (with test cases) without me having to write a single line of code? Well...other than me having to break out the printf debugger to help with tracking down some deep bugs that is.

I've been trying to get the original implementation to work correctly for so long I don't even remember what I wanted to use it for in the first place, I just mess with it a bit here and there when I have nothing better to do.

And that's just me, a half-assed self-taught junior woodchuck coder, playing around to see what all the hype is about.

I can only imagine that I'm giving them valuable training data as some of the bugs were very deep and took a whole lot of 'thinking' to track down the root cause. It does take a bit of prodding to get it to look in the right place but so far it has found and fixed them all. The last bug was an overflow in the tree iterator's stack it uses to track state across iterations that I was concerned would time out as it was thinking for a long, long time.

I'm not really one to defend the robots but this one is actually uesful.

--edit--

Oh... I guess that's a meme now. Nothing to see here, move along...

I wondered how I should feel about that thought process. However, I might have a similar thought process when inebriated.

I mean, to be fair, given the way tokenization works, this isn't that surprising.

I do agree that it's very funny though.

I think it’s like asking someone “how many times does your pen change direction when writing the word strawberry on paper”.

You need to think really hard to get to an answer, because that’s more fine grained than the way you usually think about words and letters.

Reply on news.ycombinator.com