Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Taste in the age of AI and LLMs

rajnandan.com

Post reply on HN
Print view View original

221–230 of 230 posts

Re: Taste in the age of AI and LLMs

#221

Post by AlexCoventry »

osti wrote: post #107
AlexCoventry wrote: post #78

Try using a coding agent to write an efficient GPU kernel. I guess they might get good at it soon, but they definitely aren't there yet.

I had a very complex cuda kernel and codex cli managed to improve the throughout 20x.

Hmm, I still have some nonrefundable API credits with OpenAI. Maybe I should try to use them for my kernel.

FWIW, this talk[1] from NVIDIA/Meta from March claims that coding agents can often write correct implementations of of CUDA kernels, but that they're usually dog slow, like 100x slower than a kernel optimized by a skilled human.

[1] https://www.nvidia.com/en-us/on-demand/session/gtc26-s81653/

Re: Taste in the age of AI and LLMs

#222

Post by e3df »

nate8bit wrote: post #37

> A practical loop for training taste Taste is cheap. Taste (or a rudimentary version of it at least) is something you start with at the beginning of your career. Taste is the thing that tells you "this is fucking cool", or "I don't know why but this just looks right". LLM's are not going to replicate that because it's not a human and taste isn't something you can make. Now - MAKING something that "looks right" is ha…

Taste comes from childhood imo. It is not something that magically appears in adulthood.

It was always there, albeit laying dormant until one does something that requires it.

Re: Taste in the age of AI and LLMs

#223

Post by solaire_oa »

jatins wrote: post #106

Title: Good Taste the Only Real Moat Left Followed by an entire AI generated fluff piece https://www.pangram.com/history/347cd632-809c-4775-b457-d9bc... Flagged

I got to the second section before I decided to scan how long it was, saw a wall of text, and decided that this article was low taste.

Moreover, the submitter of this article (probably not the author) spams ~4 submissions per day.

Re: Taste in the age of AI and LLMs

#224

Post by eru »

lamasery wrote: post #217
eru wrote: post #206

Earlier quoted context omitted.

Thanks for explaining. We don't get much of this 'uniform' here in Singapore. It's too hot for tech people to wear long trousers and long sleeves. I guess it makes sense of a joke. Though I don't think sense in fashion correlates much with sense in technology: the guy might just be wearing what everyone else is wearing, so that his wife doesn't bug him about it and he can concentrate on the hills he's actually prepar…

I don't think any wife has ever asked for this look :-) Someone just trying to minimally not-stand-out could do so way cheaper. The high cost is part of the point (for signaling). The other major standard look is rich-hipster: $300 Japanese jeans ("they bought the actual looms Levi's used in the '50s! It's more authentically Levi's than Levi's is now!"), $350-400 work-, jump-, or motorcycle-pattern leather boots, $15…

> I don't think any wife has ever asked for this look :-)

The wife wouldn't specifically ask, it's just that she doesn't bug him, when he's wearing something 'normal'. And normal is what everyone else is wearing.

> Someone just trying to minimally not-stand-out could do so way cheaper. The high cost is part of the point (for signaling).

Yes and No. You might have to spend so much just to blend in, depending on where you are. And it might not be all that much money, if you make enough.

Here in Singapore people (thankfully!) tend to dress down a lot. It's a big difference from eg Hong Kong. Flip-flops, shorts and t-shirt seem to be the norm.

I'm indulging a bit by wearing the same, but in wool. Eg from https://woolandprince.com/ but that's mostly because it deals with my copious sweat marginally better than cotton.

Btw, looking like a trope can be beneficial. Eg at tech conference, if you look like a techie you often get better treatment than when you look like some kind of influencer or (low level) manager. In normal life, signalling that you are well-off often gets you good treatment, too.

What's an REI?

Re: Taste in the age of AI and LLMs

#225

Post by nate8bit »

e3df wrote: post #222
nate8bit wrote: post #37

> A practical loop for training taste Taste is cheap. Taste (or a rudimentary version of it at least) is something you start with at the beginning of your career. Taste is the thing that tells you "this is fucking cool", or "I don't know why but this just looks right". LLM's are not going to replicate that because it's not a human and taste isn't something you can make. Now - MAKING something that "looks right" is ha…

Taste comes from childhood imo. It is not something that magically appears in adulthood. It was always there, albeit laying dormant until one does something that requires it.

I would... actually maybe agree with this. I'm not certain it can be taught!

Re: Taste in the age of AI and LLMs

#226

Post by edgyquant »

skydhash wrote: post #216
edgyquant wrote: post #214

Earlier quoted context omitted.

Regardless that is why most people are moving faster than writing by hand

Moving faster towards what? The biggest diff? The most tickets closed? Or a working software?

Working software/new features.

Re: Taste in the age of AI and LLMs

#227

Post by taude »

germinalphrase wrote: post #193
taude wrote: post #45

Earlier quoted context omitted.

I feel like you're pretty strongly agreeing that taste is important: " I'm finding that you have to have an extremely clear product vision..."" Clear production vision that you're building the right thing in the right way -- this involves a lot of taste to get right. Good PMs have this. Good enginers have this. Visionary leaders have this.... The execution of using AI to generate the code and other artifacts, is a ma…

“ I've looked at three non-engineer vibe-coded businesses in the past month, and can tell that without taste, they're building a pretty mediocre product at best.” Are you doing this altruistically for friends - or as a consultant?

Both a) to help a friend out and b) to help non-technical founders I've meet at some Meetups/AI events to launch their product. My short-term goal is to put together a checklist/cheatsheet for all the technical things someone needs to do to launch a business because it's not just having a webapp running on Vercel with Supabase. And if they do have an app, is it a complete mess or not.

I think the solo-founder hype is an overplayed unless the person has the right skills, and even worked at a tech company, and knows what they're getting into. Alerting and monitoring for example is one of like 30 things they should be aware of.

Re: Taste in the age of AI and LLMs

#228

Post by osti »

AlexCoventry wrote: post #221
osti wrote: post #107

Earlier quoted context omitted.

I had a very complex cuda kernel and codex cli managed to improve the throughout 20x.

Hmm, I still have some nonrefundable API credits with OpenAI. Maybe I should try to use them for my kernel. FWIW, this talk[1] from NVIDIA/Meta from March claims that coding agents can often write correct implementations of of CUDA kernels, but that they're usually dog slow , like 100x slower than a kernel optimized by a skilled human. [1] https://www.nvidia.com/en-us/on-demand/session/gtc26-s81653/

For me it was able to try out different architectures for perf improvement, then once it's settled on some good architectures, it can do lower level optimizations on them by profiling the code etc.

Re: Taste in the age of AI and LLMs

#229

Post by AlexCoventry »

osti wrote: post #228
AlexCoventry wrote: post #221

Earlier quoted context omitted.

Hmm, I still have some nonrefundable API credits with OpenAI. Maybe I should try to use them for my kernel. FWIW, this talk[1] from NVIDIA/Meta from March claims that coding agents can often write correct implementations of of CUDA kernels, but that they're usually dog slow , like 100x slower than a kernel optimized by a skilled human. [1] https://www.nvidia.com/en-us/on-demand/session/gtc26-s81653/

For me it was able to try out different architectures for perf improvement, then once it's settled on some good architectures, it can do lower level optimizations on them by profiling the code etc.

That's great. It seems like a large body of experts are having real problem with this, so maybe you should publish something about your methods, or start a business...

Re: Taste in the age of AI and LLMs

#230

Post by osti »

AlexCoventry wrote: post #229
osti wrote: post #228

Earlier quoted context omitted.

For me it was able to try out different architectures for perf improvement, then once it's settled on some good architectures, it can do lower level optimizations on them by profiling the code etc.

That's great. It seems like a large body of experts are having real problem with this, so maybe you should publish something about your methods, or start a business...

I can't vouch for whether or not it can beat human experts though because I'm no CUDA expert myself. The original CUDA code were human written and I first let codex adapt it to my specific use case. Then I basically let codex generate ideas and try the ideas out itself (I think it's a bit like Karpathy's autoresearch, except I was still doing manual prompting). And that was enough to get me 20x improvement.

I suspect when people said AI wrote non performant CUDA kernels it was beginning-mid last year and it's definitely vastly improved since back then. And the agent's ability to iteratively improve really impressed me.

Post reply on HN