4 ms·

Greater efficiency of light bulbs has led to more light bulb use, not less. More efficient training of LLMs could just as likely lead to more chip use, not less

Greater efficiency of light bulbs has led to more light bulb use, not less. More efficient training of LLMs could just as likely lead to more chip use, not less.

gniv 2y ago
But did it not lead to net less electricity used for lighting?
- BobbyJo 2y ago
  I'd wager there is more light today than before LEDs were invented. I'd wager a lot in fact.
- divbzero 2y ago
  With more efficiency offset by more light bulbs, electricity used for lighting has been roughly flat since 2010: https://www.iea.org/data-and-statistics/charts/global-electricity-consumption-in-lighting-in-the-net-zero-scenario-2010-2030 https://www.iea.org/data-and-statistics/charts/global-electr... (For LLMs I wish that efficiency could lead to less electricity used for chips, but I think the best we can hope for is for electricity use to flatten out.)
  - gniv 2y ago
    From that graph it looks like residential went down quite a lot, which makes sense since that was the main usage of incandescents.
  - jes5199 2y ago
    we just have to get solar deployed sufficiently so that we no longer think of energy use as a waste of something precious
- jstummbillig 2y ago
  Maybe, at some point. From what I can tell there's are mostly two options: Either AI is and will be useless or it's severely undersupplied. People, even those deeply technical, where AI has the most impact right now, still widely argue about if AI even offers any value. Adoption is far from anything that is plausible, if (not when) it became clear that it does. If you land on "does not", given the investments so far, commercial entities would obviously be overvalued already and any investment goes to 0 over time. But if we land on "does", how could Nvidia not be anything other than undervalued right now? No matter what frontier model: I can look at my screen, LLM generated characters visibly appearing in chunks, depending on the model after initially waiting for 10-20 seconds, for even benign queries — because that is the best we can do right now. And that's while most people still argue if AI will actually do anything and humanity at large does not really use it, neither personally nor societally. If AI does in fact do something valuable and that something gets better, everyone will want it and there will be demand for lots of chips.
light_hue_1 2y ago
Across 300 years of lighting technology decreases in cost consistently led to much larger increases in light use every time. Until about the past 50 years or so when increased light use started to fall behind the drop in costs. https://www.researchgate.net/profile/Roger-Fouquet/publication/227349575_Seven_Centuries_of_Energy_Services_The_Price_and_Use_of_Light_in_the_United_Kingdom_1300-2000/links/0fcfd50fa79718f3c3000000/Seven-Centuries-of-Energy-Services-The-Price-and-Use-of-Light-in-the-United-Kingdom-1300-2000.pdf https://www.researchgate.net/profile/Roger-Fouquet/publicati... Cheaper models are good for industry. But the demand for better more reliable models is immense. The models won't stay cheaper.

Greater efficiency of light bulbs has led to more light bulb use, not less. More efficient training of LLMs could just as likely lead to more chip use, not less.

gniv 2y ago

But did it not lead to net less electricity used for lighting?

BobbyJo 2y ago
I'd wager there is more light today than before LEDs were invented. I'd wager a lot in fact.
divbzero 2y ago
With more efficiency offset by more light bulbs, electricity used for lighting has been roughly flat since 2010: https://www.iea.org/data-and-statistics/charts/global-electricity-consumption-in-lighting-in-the-net-zero-scenario-2010-2030 https://www.iea.org/data-and-statistics/charts/global-electr... (For LLMs I wish that efficiency could lead to less electricity used for chips, but I think the best we can hope for is for electricity use to flatten out.)
- gniv 2y ago
  From that graph it looks like residential went down quite a lot, which makes sense since that was the main usage of incandescents.
- jes5199 2y ago
  we just have to get solar deployed sufficiently so that we no longer think of energy use as a waste of something precious
jstummbillig 2y ago
Maybe, at some point. From what I can tell there's are mostly two options: Either AI is and will be useless or it's severely undersupplied. People, even those deeply technical, where AI has the most impact right now, still widely argue about if AI even offers any value. Adoption is far from anything that is plausible, if (not when) it became clear that it does. If you land on "does not", given the investments so far, commercial entities would obviously be overvalued already and any investment goes to 0 over time. But if we land on "does", how could Nvidia not be anything other than undervalued right now? No matter what frontier model: I can look at my screen, LLM generated characters visibly appearing in chunks, depending on the model after initially waiting for 10-20 seconds, for even benign queries — because that is the best we can do right now. And that's while most people still argue if AI will actually do anything and humanity at large does not really use it, neither personally nor societally. If AI does in fact do something valuable and that something gets better, everyone will want it and there will be demand for lots of chips.

light_hue_1 2y ago

Across 300 years of lighting technology decreases in cost consistently led to much larger increases in light use every time. Until about the past 50 years or so when increased light use started to fall behind the drop in costs. https://www.researchgate.net/profile/Roger-Fouquet/publication/227349575_Seven_Centuries_of_Energy_Services_The_Price_and_Use_of_Light_in_the_United_Kingdom_1300-2000/links/0fcfd50fa79718f3c3000000/Seven-Centuries-of-Energy-Services-The-Price-and-Use-of-Light-in-the-United-Kingdom-1300-2000.pdf https://www.researchgate.net/profile/Roger-Fouquet/publicati... Cheaper models are good for industry. But the demand for better more reliable models is immense. The models won't stay cheaper.