Live data from Hacker News

I love LLMs, I hate hype

geohot.github.io

211–220 of 340 posts

Re: I love LLMs, I hate hype

#211
post #108

Thank you, I really needed to read a sane voice. The relentless hype-onslaught is not easy to cope with.

These comments are always interesting to me (genuinely! no shade) because it feels like we're on different sides of the Internet. Almost everything I see about LLMs is somewhere between uneasy acceptance, skepticism, and outright hate. Even those of us who admit they're impressively capable usually caveat it with "... but our jobs and society are cooked, aren't they?"

The "hype onslaught" is, like... Something I know exists, on some abstract level. In news reports about CEOs, or seeing screenshots of LinkedIn, or whatever. But it feels so far removed from a sentiment any real person holds.

Re: I love LLMs, I hate hype

#212
post #196

Earlier quoted context omitted.

Maybe he thinks the model will be smaller.

I doubt it can get < 100gb. Isn't fable like 10 trillion params?

The math doesn't math. Wearables have to be small and body temperature, which is not something that fable likes.

Re: I love LLMs, I hate hype

#213

Earlier quoted context omitted.

The fact that you're even saying this it is probably an admission that you do think it's making you dumber. Most people I know, who are honest with themselves, have admitted to me that they feel like it's making them dumber or "zombifying" them. This is also well studied already, https://arxiv.org/abs/2506.08872 LLMs are poison for the brain, I'm almost certain of it, at least when used in the way most people are usi…

It's pretty easy to generalize this, but it doesn't match my perception. People who are using llms to do things they could have already done, but faster, probably have atrophying skill sets. People who are using these tools to accomplish significantly more difficult or complex work than they used to are absolutely finding new ways to push themselves. The problems are just much bigger. The average Joe can easily vibe…

My perception is that this guy's response to "I forgot to write a readme" was "I should limit myself to tools that do it for me" instead of instilling discipline about documentation.

Re: I love LLMs, I hate hype

#214
post #39

I love LLMs too, but I am concerned about their cost. They are all still very subsidised. Is there any guarantee that I'll be able to run a Opus 4.8-level model on my personal computer before the big AI labs decide to hike up the prices?

> They are all still very subsidised. I think the opposite: I think the frontier labs have good margins on their inference unit costs. We can already see what it costs to run near frontier-size models. There are independent business pivoting to serving these models at reasonable prices and they're competing on OpenRouter for costs much lower than frontier labs. > Is there any guarantee that I'll be able to run a Opus…

I mean, Pfizer has a good margin on every pill they produce too.

Re: I love LLMs, I hate hype

#215

Earlier quoted context omitted.

>I predict MoE is a transitional technology While scaling laws hold (more weights = better), and time / financial costs are not trivial the incentives are in place to have MoE. MoE means you can have more weights without increasing the critical path of evaluating it. I am curious what you believe the problems with it that would cause people to prefer using less weights. I'm not following what you mean by MoE can't ha…

Weights are not created equal: while interpretability is a young field the prevailing view at the moment is that MLP (hence experts) in a mixture model are substantially where dense encoding of factual information resides, attention is even less easily interpreted but it should be uncontroversial that temporal/sequential modeling occurs here. So it's more consistent with available empirics to say that an architecture…

>the prevailing view at the moment is that MLP (hence experts) in a mixture model are substantially where dense encoding of factual information resides

Yes, because that's where all the parameters are. For reference in GLM 5.2 98% of the weights are for the experts.

>The "Mixture of Parrots" work is a good jumping off point

The paper shows increased performance on knowledge dependent task while having similar reasoning capabilities. This backs up what I was saying about how the weights unlock extra performance without increasing inference costs as much as a dense model would.

>model power allocated at this point or that will recover different performance profiles

While increasing the number of weights makes the model better, where those weights are does matter in how much better the model gets and also matter in regards to the cost of training / inference. Model design is a big set of trade offs and I see MoE as a useful tool that will survive in the trade off space.

>reasoning is left on the table

Even so, if there was 2 models with an equivalent amount of reasoning ability and priced the same would you rather pay for the one with narrow knowledge or wider knowledge.

>because of the paucity of modern dense models at the near frontier,

You don't need to be at the frontier to benefit from MoE. Even open source models that are behind the frontier, benefit from being able to host experts on different machines, and scale individual, commonly used experts separately from each other. On the other end with small models you are probably resource constrained so you want to maximize the tokens generated per second. This makes going for purely dense models niche like you are saying.

Re: I love LLMs, I hate hype

#216
post #87

This line: "this is my main argument against the valuation of frontier labs. It’s not that AI won’t create that much value, it’s that they won’t capture it." That is a very astute and concise way to explain everything about how the frontier labs are behaving and how they're trying to push more people to pay token rates for the best models. At the current subscription prices ($100 or $200 a month for a generous, thoug…

Who is going to end up capturing all this value being generated is going to be very interesting. Back in 1980, who’d have thought MS would capture the majority of the value from PCs over the next 3 decades, and not IBM?

I miss OS2

Re: I love LLMs, I hate hype

#218

Earlier quoted context omitted.

You will likely end up in maintenance hell soon. This will likely not be much easier with AI because coding is not the hard/annoying part, it's the fact that you need to dust off every little project every time a tiny fix is needed, and that's a lot of toil in the long run.

Maybe? I ran across an old pre-LLM project of mine recently, and past me was an asshole and didn't leave a readme for future me. Meanwhile post-LLM projects at least have a readme that the LLM generated for me or my agent to read and pick up context on. Being able to ask an agent what is this repo, what's going on here? Hey just make it do this, instead of toilsomely digging in and doing it tmmyself, seems to say tha…

Sometime a wrong and misleading README done more harm than good.

It is not that rare to see LLM waste hours on a wrong path because a misleading line in README. Even worse, they can't learn. Spawn a subagent and it repeat the same error again

Re: I love LLMs, I hate hype

#219

Earlier quoted context omitted.

5.6 Sol isn't Fable, but calling it unusable is a bit much.

It's unusable only because I can't set 5.6 to Extra thinking in the "Chat" tab, and the "Work" feature that does let me use Extra is totally gimped for coding tasks. I'd rather use Opus 4.8 Extra/Max instead of Sol High (look up the benchmarks, thinking level is everything ). It is quite literally unusable to me.

Oh so you used it but the results were sub par, not that you were unable to use it.

Re: I love LLMs, I hate hype

#220

Earlier quoted context omitted.

5.6 Sol isn't Fable, but calling it unusable is a bit much.

It's unusable only because I can't set 5.6 to Extra thinking in the "Chat" tab, and the "Work" feature that does let me use Extra is totally gimped for coding tasks. I'd rather use Opus 4.8 Extra/Max instead of Sol High (look up the benchmarks, thinking level is everything ). It is quite literally unusable to me.

> I'd rather use

Is different than

> quite literally unusable

Post reply on HN