8 ms·
Jevons paradox
- davedx 2y agoI am right with you with a very large NVDA holding, friend! HODL!
- tbriudmepn 2y agoIf it feels very large it might be wise to diversify. Not for maximum $, but maximum zzzzz.
- wslh 2y agoEnergy is commonly used for this paradox.
- tbriudmepn 2y agoJevons Paradox Paradox: Markets fall on news that should make them rise, due to Jevons Paradox.
- marcosdumay 2y agoSomehow, nobody expects the Jevons' Paradox. Even when those same people talk about it every other day.
- YurgenJurgensen 2y agoThe Hofstadter-Jevon Corollary: Taking Jevon’s Paradox into account only makes it worse.
- bilbo0s 2y agoI made a comment yesterday that the whole NVDA thing seemed eerily similar to the AMZN thing in the first dot com crash. [0]. If the market can get sufficiently irrational, and we start to see that same like 80-90% markdown that we got with AMZN in the first crash, I mean, who knows? That's a golden opportunity. Not many generations get opportunities like that twice. I guess we'll see what happens with the stock price, but long term I'm pretty confident in this Jevon's Paradox effect on AI use. I'm praying for more negativity hype out among those who maybe don't know as much about tech as the HN crowd. Because those people collectively, can move stock prices. [0] https://news.ycombinator.com/item?id=42853937 https://news.ycombinator.com/item?id=42853937
- llm_trw 2y agoHaving read the paper from deepseek I seriously don't get how anyone can think "Oh now is the time to sell nvidia". Both at the tactical level and strategic level that paper means we will see _more_ nvidia gpus fly off the shelves. If anything it's showing why the h800 is a decent card and everyone in China should buy one (dozen thousands).
- bilbo0s 2y agoI had analyzed the paper as well, and gave some quick thoughts on a few of the low level things that stood out and would affect NVDA here: https://news.ycombinator.com/item?id=42856652 https://news.ycombinator.com/item?id=42856652 I have to agree. I'm not sure the architectural and engineering changes here don't put NVDA further ahead. Some stuff all gpu makers can do. Like quantize on transfer. But NVDA is so far ahead in other areas and DeepSeek leans heavily on that industry leading tech. At least at the low level. But I'm betting the average investor doesn't have the level of facility with ML to properly analyze the information that's come out. Heck, I'm betting most of the people working in AI are really just TF or PyTorch monkeys. The real AI experts have probably been hovered up by MS or FB or AMZN or OpenAI or what have you. So it feels like we could get lucky here and the AI related market could drop like a rock taking the good with the bad. Which I think would be awesome for most people in a position to wait for it to happen.
- llm_trw 2y agoI can't wait for the 2000 dot com crash for AI. What gets me about the deepseek work is that it's neither novel or theoretically difficult. I've been using a similar synthetic data pipeline to generate examples for sota bert embedding models. That low-hanging fruits like this are left unexplored by meta or alphabet - and if we're to be believed lead to crisis meetings for a week straight - makes one wonder wtf the engineers there are doing?
- spamizbad 2y agoThis is just a baseless conspiracy theory that I have, but I do wonder if they intentionally avoided certain avenues of research because it could erode the “moat” major tech firms have by dramatically reducing the capital costs required for training and inference. If you focus your research around work that mandates high-end hardware at scale you can lock out tons of potential incumbents
- aaron695 2y ago[dead]
- belter 2y agoLots 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-weeks-before-price-drop-costs-tech-stock-billions-2021725 https://www.newsweek.com/nancy-pelosi-sells-nvidia-stock-wee...
- strobeflier 2y agoSold last year. This is not really the place for misleading low-effort politics.
- mrbungie 2y agoSure, 31 December 2024 is by definition last year, but at the same time that was not even a month ago.
- strobeflier 2y agoCome on. Which do you think is more likely, that it was a normal EOY sale of a lucrative stock, or that they sold based on insider info obtained weeks before anyone else?
- throwmeme888 2y agoit was 2-3 days after deepseek paper release on Christmas Day. I read this paper too and drew the exact same conclusion (that nvda was due to decline) but I dont think its insider trading, just informed and reactive trading - not eoy profit trading either though
- strobeflier 2y agoFair. The existence of said paper makes this a normal trade based on publicly available information.
- mrweasel 2y agoThis has actually been on my mind lately and I've been looking for statistics, from Denmark, but haven't been able to find any going back far enough. My assumptions is that we in the late 1940s through the 1950s had achieved a reasonable standard of living, but perhaps not available to all. Even if we assume that everyone in a country has access to the same, high, late 1950s standard of living, but applied modern technology, like improved insulation, district heating, EVs, energy efficient light bulbs and appliances, then how much would that lower (or raise) our energy usage? More of the energy would surely be moved to electricity, from coal, gas and oil delivered directly to our home, but is it enough chance the environmental impact?
- TSiege 2y agoWell I can answer one question for you. There are studies that show that better insulation only briefly lowers the energy use. Once people adjust to the new insulation they end up heating or cooling their houses more since in their mind it is cheaper, but they end up using the same or more energy. https://www.cam.ac.uk/research/news/insulation-only-provides-short-term-reduction-in-household-gas-consumption-study-of-uk-housing https://www.cam.ac.uk/research/news/insulation-only-provides...
- mrweasel 2y agoI would not have guessed that that would be the answer. That's kinda crazy, but it also makes sense, every time I visit someone in a newly built home the temperature is always crazy high.
- harrison_clarke 2y agopresumably there are limits to that. once every room is the preferred temperature, i'm going to stop running the heater although maybe i'd get bored and run the heater and the AC at the same time
- nine_k 2y agoOften you can't because the heater and the AC is literally the same unit, the heat pump.
- _pete_ 2y agoThe best Jevons Paradox/induced demand explainer ever: https://www.youtube.com/watch?v=pCzCJzwrB_c https://www.youtube.com/watch?v=pCzCJzwrB_c
- casey2 2y agoThe 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
- mminer237 2y agoI 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 to start training competing models to carve out their piece of the market?
- sz4kerto 2y ago> Are people really locked out of LLMs due to price? Yes. For example Google has just made Gemini a standard part of Workspace, before that it cost $36 or something per month per user. That was too much for many SMBs to experiment with (you and I understand that the potential efficiency gains are way higher but for an SMB, paying 4x more for your office suite sounds bad).
- kbolino 2y agoThese 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 toward uncanny simulacra and hallucination Bringing the cost way down and making them trainable on consumer hardware solves or at least greatly alleviates problems 1-3. That just leaves problem 4, which might still be unsolvable and sink the whole endeavor, but at least can be focused on.
- scotty79 2y agoCurrent market panic looks like early railroad investors freaking out that someone made a faster train that goes over same tracks.
- kgwgk 2y agoIs that really the analogy you want to use? « In 1873, greed, speculation and overinvestment in railroads sparked a financial crisis that sank the U.S. into more than five years of misery » https://www.smithsonianmag.com/history/robber-baron-gamble-railroads-brought-down-bank-plunged-us-into-first-great-depression-panic-1873-180982877/ https://www.smithsonianmag.com/history/robber-baron-gamble-r...
- scotty79 2y agoAnd yet, in history, railroad is one of the few examples of bubbles that didn't ultimately popped. You can still lose money in the greed-fear cycle even if the road leads up to unimaginable heights.
- kgwgk 2y agoI guess if it « didn't ultimately popped » neither did the dotcom bubble - nor any housing bubble.
- scotty79 2y agoYes. Both dotcoms and housing still bring immense profits to investors.
- mrbungie 2y agoA bubble around a technology can pop, and yet the tech can be ultimately valuable in the long term. That does not mean it wasn't a bubble nor it didn't pop.
- Devasta 2y agoThe paragraph on deepseek reads like cope from an NVDA bagholder tbh.
- StefanBatory 2y agoI 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 less computational power and energy. This efficiency has made the technology more accessible and affordable, spurring wider adoption across industries.[30] As Microsoft CEO Satya Nadella noted, the increased accessibility and efficiency of AI tools like DeepSeek's R1 model could lead to a surge in applications and use cases. This expanded usage may result in higher energy consumption overall, despite individual efficiencies. For example, while DeepSeek's model uses fewer resources per task, the growing demand for AI-driven solutions—ranging from chatbots to complex reasoning models—could increase the total energy requirements of the global AI infrastructure.[31]
- HeatrayEnjoyer 2y agoDeepseek is 651B, GPT models are suspected to be <200B. Where are they getting the idea that Deepseek is more efficient?
- enragedcacti 2y agoIts an MoE model so only 37B parameters are activated per token.
- aurareturn 2y agoMoE models aren’t new. Rumors are that GPT4 was one.
- swyx 2y agoyea specifically 8x220B https://www.latent.space/p/geohot https://www.latent.space/p/geohot
- enragedcacti 2y ago
- cjbgkagh 2y agoThe 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 with phones, good enough is good enough and people would rather save the money. If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade.
- aurareturn 2y agonew method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it In my opinion, yes. If scaling laws continue at 1,000,000x, we will turn the whole planet (or Mercury) into a giant GPU.
- cjbgkagh 2y agoDo you have a use case for your opinion? On the assumption that we’d want super intelligence to do interesting things, consider that we currently have many very smart humans and society is not maximizing their use, quite the opposite.
- whatshisface 2y agoVery smart humans don't want to do random tasks, they want to study number theory.
- cjbgkagh 2y agoNot all, and many smart humans are prevented from doing what they want to do for all sorts of reasons. One reason being others feel threatened by them, but those same people will feel threatened by AI, probably more so.
- aurareturn 2y ago
- pizza234 2y agoThis is IMO a spefic case of the induced demand concept (https://en.wikipedia.org/wiki/Induced_demand https://en.wikipedia.org/wiki/Induced_demand).
- rawkintrevo 2y agoI was just coming here to say, this looks like the fancy way to say 'induced demand'. Why do you think this is a specific case of induced demand (as opposed to induced demand being a specific case of Jevons paradox, or the two being different words for the same thing?)
- BenFranklin100 2y agoAs an aside, the concept of Induced demand has been criticized. A more precise way to think about “induced demand” is that the demand for a product at a lower price point was preexisting. The transaction didn’t clear until the price actually became lower and this demand manifested itself. This seems like splitting a hair but it is a more consistent intellectual framework for understanding the dynamics of a particular market.
- simgt 2y agoI don't see how that is more precise. What is the definition of "demand" in that case? Anything that could ever happen at one point, or something that is actually wanted (needed?) by someone who's formulating the thought consciously?
- enragedcacti 2y agoInduced demand is just the name for realized latent demand, your more precise way to think about it is the concept*. It's only muddled because it is often invoked in complex markets where the costs are often not monetary, there are substantial externalities to consider, or the supply is centrally managed. Should a city spend $500,000 to install lighting in the park downtown? After all nobody uses it at night because its too dangerous! Induced demand is central to the answer but isn't a first thought for many because the problem space doesn't look like a traditional market. * there is the additional concept that the long term availability of a good or service at a lower cost changes consumer behavior in a durable way. Building a new train station realizes latent demand but also creates new demand as it causes denser housing to be constructed near it**. ** Whether you want to model that as just more realized latent demand or as a consequence that's technically distinct from induced demand is IMO where a lot of the confusion comes from. If you ask me this should probably have a distinct name as it is a long-term, interrelated process that doesn't map well to sliding around supply and demand curves. Generated demand is what Bloomberg calls it but that has insane overlap with a marketing technique.
- anshulbhide 2y agoBenedict Evans caught on early - https://www.ben-evans.com/benedictevans/2023/7/2/working-with-ai https://www.ben-evans.com/benedictevans/2023/7/2/working-wit...
- llm_nerd 2y agoThis "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.
- aurareturn 2y agoThis "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. People are arguing that DeepSeek is a bull case for Nvidia, not bear. Nvidia being worth $3.6T is a different topic. People think DeepSeek should have reinforced that valuation or even increased it, not lower it.
- steveBK123 2y agoGiven the likes of sama & co were asking for unfathomable levels of capital, chipmakers had backlogs, datacenter & power supply chain companies were getting bid to the moon, etc.. having a dramatic decrease in compute requirements might just bring things closer to a supply/demand balance. Where was Stargate gonna get $500B? How was Altman ever going to raise.. $7T? (~1/3 of the entire GDP of US) Does a 10x or even 40x decrease in compute cost just mean they can actually raise & spend the capital that their compute requires?
- aunty_helen 2y agoThe simplest explanation, Jevons paradox is cope from CEOs and retail investors trying to justify the share market. People seem to have forgotten that there’s a time element involved, not a perfectly elastic and instant demand. Thinking about it another way, if it turned out it was going to cost 30x more compute to train the next generation of models than we thought due to some scaling law, would the chip share prices be expected to rise, stay the same or go down? Probably rise. In that case, if it was 30x cheaper to do what we need today would we expect option a, b or c? Assuming that there would be less short / mid term sales, that impacts their bottom line and flattens off their (ie nvda) demand curve. Slower growth, lower share price. Slower growth, longer opportunity for competition. Therefore the speculative element of the chip leader’s share price’s should come down. I don’t have any long or short positions in the chip sector.
- suraci 2y agothe money must be happened, otherwise shit will happen it's like we are spending 500bn for x ai needs, but suddenly, the cost fell to 1/20, now we must convince everybody there'll be x20 needs, so that the 500bn will not go missing that is what Satya Nadella and Sam Altman are trying to achieve the real problem: 1. will there be 20x needs for ai? 2. how much ai needs are priced now? x? 5x? or, 20x?
- aurareturn 2y agoCompanies compete against each other - not themselves. It’s not like OpenAI set an internal goal of training 30x bigger and that’s it. No, OpenAI needs to compete. If training efficiency is 30x more, then Anthropic, Google, OpenAI, Mistral, Qwen, etc all get the benefit. It resets the baseline. They will not stop competing against each other just because there is a 30x efficiency improvement that everyone has access to.
- tobmlt 2y agoThis reminds me of something I remember from college but cannot ever seem to find: "The Jevon's diagram" by which I mean a two country system, with country containing two groups: highly educated and not very educated. The two countries have very different proportions of each group, and the analysis is around free immigration vs no immigration, and shows the harm done to the larger number of less uneducated in one of the countries. Anyway, I'm not going to do a good job of describing this thing, from 20 years ago in class, here, but if anybody knows offhand what I mean and can point me at an article or source on this, I'd be very appreciative! It's bugging me from way back! (where are my notes! Silly tobmlt, not taking notes in latex back then)
- typon 2y agoWith AGI/ASI, there will come a point where this paradox breaks down: once the ASI is self-sufficient and "agentic enough", people's direct use for it will actually reduce. You won't need to "micromanage" it - and the total amount of compute it requires might actually drop drastically if the ASI improves its own efficiency much better than humans can. At that point the question becomes one of alignment: will overall consumption drop because ASI subjugates or eliminates humanity or does it grow exponentially as we use it to explore the stars.
- aurareturn 2y agoWhy can’t ASI optimize its own efficiency and make better/more chips at the same time?
- HarHarVeryFunny 2y agoDoesn't really seem like a Paradox. There may only be 10 people willing to pay $100 for something, but 1000 willing to pay $10 for it, so lower price results in more revenue.
- aurareturn 2y agoThat’s a given. Another counter argument is that’s Nvidia will sell less high margin advanced GPUs that require expensive networking equipment. Problem with that argument is that every competitor gets the benefit of higher efficiency and the baseline will reset back to 0.
- HarHarVeryFunny 2y agoThe models will all get more efficient, but NVIDIA as the high cost provider of compute (as well as cloud services that use them) - 10x markup on H100s - may not benefit from increased demand. If the models turn out to be a commodity, then low cost providers will be the winners, likely running on chips like TPU and Trainium (Amazon/Anthropic) that don't have the 10x markup.
- aurareturn 2y agoLow cost providers will not be the winner because those will become commodity. It’s the leading model, the one closest to AGI that will make all the profit. Nvidia has products to serve lower end as well. There’s no reason why Nvidia can’t dominate in the commodity sector. And if Nvidia can’t keep their high margins, they will make it up with significantly more volume (Jevons Paradox in play).
- HarHarVeryFunny 2y agoI think there'll continue to be models at different levels of capability and price - and not at all obvious whether it'll be the Ford or Rolls Royce that'll make most money. My guess is that is that what passes as "AGI" will also become commoditized, maybe even free to consumers (funded by advertising, business referrals to restaurants, etc).
- ApolloFortyNine 2y ago>Energy-efficient lighting, such as LED bulbs, has become a popular solution to reduce energy consumption. However, it serves as a notable example of the Jevons Paradox. While LEDs significantly lower the cost of lighting per unit, this reduction often leads to increased usage. For instance, people may install more lights in their homes or businesses, use them for longer durations, or adopt decorative and outdoor lighting more extensively. As a result, the total energy consumption for lighting has not decreased significantly despite the efficiency improvements. This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat. So I took a look at the source, which doesn't actually have a source and just links to another white paper that claims light pollution increased, but is unclear about how much and due to what. Perhaps that whitepaper links to another whitepaper, but this is where I stopped. Kind of makes me question how often I blindly trust Wikipedia, though this point in particular is pretty minor.
- blagie 2y agoYour math is wrong: > This felt false to me, LED bulbs are generally 10x more efficient, so the average number of lights would have had to increase 10 fold, and this is ignoring the savings from not having AC cool those other 50 watts a bulb that went to heat. The first 5W also went into heat. Perhaps with the exception of light escaping through a window, *every* watt that goes into a lightbulb comes out as heat for your AC to cool. I would not be surprised if there were a 10x increase in use of lighting with LEDs. Not taking Wikipedia (or any source) on blind trust is always wise. More often than not, when you kick tires, you come back to folklore. It's fine for big things, but details are often wrong. Ditto for newspapers, for that matter.
- ApolloFortyNine 2y agoIf a 5w bulb and a 60w bulb output the same amount of light, those 55w are being wasted elsewhere (heat). I said 50 because I was afraid someone would go 'ACTUALLY LED's only save 50-55w', I didn't expect someone to argue that the 5w led bulb is also outputting heat. You are of course right, 2.5w of that 5w is likely heat being cooled by the AC. >I would not be surprised if there were a 10x increase in use of lighting with LEDs. Quite frankly I don't care either way, I just wanted a source. I'm happy you feel that way though.
- creamyhorror 2y agoAre people getting hung up on NVDA's fall? It feels like most aren't drawing direct logical links - the Jevons paradox doesn't apply directly to stock prices. I think the situation is fairly clear with some understanding of finance / equity markets. My analysis of the market and competitive situation: Deepseek's cheaper LLM services + providing open models for other hosts to provide => overall prices for using LLM services will fall due to competition (lower prices + more hosts entering the market); AI users won't pay so much for LLM services => the major LLM hosts/providers won't be able to project such high revenues or even have the funds to purchase as many GPUs (they will receive less capital investment to buy GPUs since revenues per dollar invested are lower; their stock valuations will fall) => corporate demand for Nvidia cards will fall (AND consumer demand will not rise as much to counteract it for multiple reasons) On the basis of this possible logic, portfolio managers and algorithms project lower growth/revenue for Nvidia and sell off its stock, setting off the usual chain reaction as other managers notice the downward price action and follow suit in order to stop further losses. Now, if the new hosts entering the market with lower prices attract significantly more users (instead of merely competing away margins) as we might expect by the paradox, then demand for GPUs might grow instead of shrink instead. If more portfolio managers think this is likely, then it's likely for NVDA's price to recover, and it would be wise to get back into the stock once the worst of the fall is over.
- crazygringo 2y agoNo, this isn't directly about Nvidia or stock prices at all -- it's about the emergence of much lower-cost AI models. So the way Jevon's paradox could apply is that if an AI model only costs 10% of the energy+GPU to train (or run), then we might not get 90% energy+GPU savings... we might actually wind up consuming more energy+GPU on the whole because the models become so much cheaper, we use them more than 10x as much. Obviously this will affect Nvidia's price, but this principle absolutely does not apply to stock prices directly.
- empath75 2y agoThe way the market reacted is completely illogical. If Nvidia announced a new chip that was 10x more effective at the same price, the market would be like WOW they're going to sell so many of those! But a software advance that achieves the same effect somehow causes the opposite reaction.
- khazhoux 2y agoAn example you may have learned about in school was Eli Whitney and the cotton gin. He thought his invention would reduce slavery by making it very cheap to clean the seeds out of picked cotton. Instead, this efficiency led to an explosion of demand for this lower-cost cotton.
- dvngnt_ 2y agonvidia stock holders favorite new term
- mrbungie 2y agoIs there literature about Jevons paradox invocations during gold rushes?
- Ekaros 2y agoBeen thinking about EVs. They are cheaper per kilometre driven. So would this not mean that the owners of them spend more overall than ICE owners spend. And would drive a lot more. As Jevons Paradox means the demand for driving will be higher thus overall money spend on driving will go up.
- omgwtfbyobbq 2y agoFYI, people mix up Jevon's and the rebound effect. https://en.m.wikipedia.org/wiki/Rebound_effect_(conservation) https://en.m.wikipedia.org/wiki/Rebound_effect_(conservation... The rebound effect is common (people have something more efficient and use it more, so the efficiency improvement is somewhat offset by more use), but Jevon's is relatively uncommon (people have something more efficient and the increase in use is large enough to offset the efficiency improvement and then some).