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

Jevons paradox

en.wikipedia.org

Post reply on HN
Print view View original

51–60 of 165 posts

Re: Jevons paradox

#51

Post by llm_trw »

spamizbad wrote: post #31
llm_trw wrote: post #24

Earlier quoted context omitted.

There is nothing in the deepseek paper that suggests you can't use the order of magnitude in hardware costs you saved to just train models that are ten times as large.

But there are thresholds of commercial viability in all of this. DeepSeek's technology gets you over that line with less sophisticated hardware. There's already some pretty impressive work being done with folks using just a pair of M2 Ultras with r1 in a "home lab" context that goes way beyond what you could previously do with llama.

Yes, but the r1 model is shiny and new. I've been using it locally for a day and I'm already finding odd spots. One thing that I've seen it do a number of times is get the correct answer in the thinking tokens, then ignore it, generate thousands more tokens until it reaches it context length then give a wrong answer.

There are things, like doc strings for functions, that llama models give better results for.

Re: Jevons paradox

#52

Post by belter »

strobeflier wrote: post #25
belter wrote: post #11

Lots of NVDA holders and NVIDIA itself, trying to justify their valuations, but forgetting the Theory of Constraints will set limits on your Jevons Paradox. The only thing relevant, is the large sale that Nancy Pelosi did before the crash. Another incredible coincidence again of course, by one of he most successful hedge managers in history... https://www.newsweek.com/nancy-pelosi-sells-nvidia-stock-wee...

Sold last year. This is not really the place for misleading low-effort politics.

Sure...Nothing to see here, just a 84 years old lady innocently playing Bingo... :

"Nancy Pelosi Outperformed Nearly Every Hedge Fund In 2024" - https://finance.yahoo.com/news/nancy-pelosi-outperformed-nea...

Re: Jevons paradox

#54

Post by enragedcacti »

HeatrayEnjoyer wrote: post #34
StefanBatory wrote: post #30

I see it already mentioned in the page ;) Jevons Paradox and DeepSeek AI The Jevons Paradox, an economic principle stating that increases in efficiency often lead to higher overall consumption, has been observed in the context of DeepSeek AI. DeepSeek, a Chinese AI startup, recently introduced its R1 model, which achieves comparable performance to leading AI systems like OpenAI's ChatGPT while requiring significantly…

Deepseek is 651B, GPT models are suspected to be <200B. Where are they getting the idea that Deepseek is more efficient?

Its an MoE model so only 37B parameters are activated per token.

Re: Jevons paradox

#55

Post by isaacfrond »

pizza234 wrote: post #37

This is IMO a spefic case of the induced demand concept ( https://en.wikipedia.org/wiki/Induced_demand ).

Reddit explains the difference thus:

"Induced demand": The highway is expanded such that it only takes me 15 minutes to get downtown instead of 45, so I go downtown more.

Jevson paradox: The highway is expanded such that it only takes 15 minutes to get from the suburbs to my office instead of 45, so I move to the suburbs.

So one is movement along the demand curve, while the other is a movement of the demand curve.

https://www.reddit.com/r/badeconomics/comments/oo2s48/that_o...

Re: Jevons paradox

#56

Post by cjbgkagh »

lherron wrote: post #50
cjbgkagh wrote: post #36

The dose makes the poison, what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it. What if efficiency brought it down such that inference on CPUs that people already have is good enough. I see a large demand curve but not an infinite one. For example laptops are getting faster every year but most people I know are not clamoring for the latest laptop, same…

It’s a question of depreciation of the GPU vs time to deploy AI. If RL and synthetic data creation are all we need for self-improving AI, the large labs have an internal use case to utilize all available compute, regardless of efficiency gains. If they have to go external and deploy in the broader market, time to deploy would make the GPUs worthless before they are fully utilized.

Did they not already have an incentive to use all available compute?

Let’s assume you were buying the GPUs, how would you plan on paying for it? What’s the business case?

Re: Jevons paradox

#58

Post by spamizbad »

mminer237 wrote: post #33
casey2 wrote: post #22

The intuition behind this is that people who had a usecase for the tech but were locked out because they couldn't afford it, or didn't believe it was worth it usecase can be engineers or researchers. Some of these people may perform better than expected causing a gap in true market value which is filled when they scale up to their new size Related: Jenson's Paradox: The more you buy, The more you save

I guess my question about the current events is: Are people really locked out of LLMs due to price? It seems like everything already has AI in it and that virtually every end user hates it. I could see a shift to more local models or something, but not an increase. I feel like LLMs have largely been oversold as a solution in search of a problem. Or are people just applying this as everyone and their mothers are going…

Well there's two sides here.

Consumers of AI do have sensitivity to pricing. Many OpenAI customers "ration" their usage. I imagine lower costs open up new demand for these people.

On the Service side: it seems to have reduced the cost of operating a "commercially viable" (something people will actually pay money for) LLM. But even beyond that, "self hosted" models are also far more affordable now, which means models that target specific niches can be viable to build or buy.

However, this won't be an overnight phenomena. In the short-term, it will seem like demand drops, but in the long-term demand will go up. Big caveat: that demand may not be concentrated on the current incumbent players.

And finally, the elephant in the room: AI still needs to become more useful to reach its full potential. Easily a decade+ more needed here.

Re: Jevons paradox

#59

Post by llm_nerd »

This "paradox" is being cited everywhere, and while it might apply to AI, people are wrongly using it to justify why nvidia should have a $3.6T or greater valuation.

What happens to a holistic system doesn't necessarily apply to every player in that system.

nvidia's heights are largely because of a few super capitalized players dumping enormous amounts of money into data centre GPUs, letting nvidia enjoy outrageous profit margins selling rapidly obsoleted, rapidly iterating products to mega-money players who are all counting on this giving them such a differentiation that it's financially worthwhile. If AI is just some free thing that has little differentiation between the players -- commoditized, so to speak -- the market for $200,000 GPUs probably isn't going to be as big. People won't be rushing to build $500B buildouts.

Another argument I keep seeing is smaller models and edge compute. NVDA has almost no stake in that, and there is no reason to believe they're going to have any particular advantage if they tried harder there, any more than Intel could leverage PCs to Smartphones. The core operations of running a model are well known by all of the silicon makers, and they can all make massively scalable FMA matrix operators, etc.

Re: Jevons paradox

#60

Post by whatshisface »

kbolino wrote: post #42
mminer237 wrote: post #33

Earlier quoted context omitted.

I guess my question about the current events is: Are people really locked out of LLMs due to price? It seems like everything already has AI in it and that virtually every end user hates it. I could see a shift to more local models or something, but not an increase. I feel like LLMs have largely been oversold as a solution in search of a problem. Or are people just applying this as everyone and their mothers are going…

These are the four main problems with LLMs (and related technologies) as I see them: 1. You can't tune them to your needs; they have restraining bolts and the training data is a generic corpus 2. You don't own your interactions with them; your data transits a network and is processed by third-party servers 3. They waste an immense amount of power relative to the usefulness of their output 4. Their responses tend towa…

In many cases the solution to 4 is also to make them faster, because many tasks are checkable.
Post reply on HN