6 ms·
In order to see how much energy AI consumes, a model to know how much energy it saves in alternative non-AI tasks that would be replaced by AI. Without models,
by germandiago 3y ago
In order to see how much energy AI consumes, a model to know how much energy it saves in alternative non-AI tasks that would be replaced by AI.
Without models, even me taking my car somewhere could be bad news. But what is the alternative? More polluting or less?
- ori_b 3y agoHow much less have you been driving due to AI?
- pfdietz 3y agoYou have misinterpreted the argument.
- cuu508 3y agoI read the original comment a few times but could not make sense of it, the first sentence especially. Could you explain the argument?
- bruckie 3y agoThe argument is that you can't say activity A increases or decreases net energy demand without modeling whether there are activities B, C, D, ... whose energy demands will decrease if activity A happens. The implication is that AI may save more energy than it consumes. One made up possibility: maybe you can spend some energy on AI to optimize a supply chain, and that saves 50x as much energy every year after the supply chain is optimized. The driving example was just that: an example. Driving obviously consumes energy, but what if, for example, you're a home energy efficiency contractor, and you're driving to a job to insulate a house, and that's going to lead to a net decrease in energy much greater than the energy you consumed to get there.
- cuu508 3y agoThanks!
- tomoyoirl 3y agoOriginal poster seeks to articulate that the real cost of AI (whether dollars or energy budget) is best understood not by the price on the sticker alone: tasks not performed by AI might be performed another way, or other tasks might be performed in their stead. Those tasks also have costs. Economic analysis of this should take that into account, and deal with the marginal cost. Original poster laments the lack of models (not AI models) of the alternatives, models that would be used in any well structured attempt to assess this marginal cost. Poster then offers fragments of a parallel scenario (in which he drives his car) and in which models might also be used. Details of this scenario are not clear to me, but presumably there was some alternative? Unclear.
- zoklet-enjoyer 3y agoSimilar to Bitcoin vs traditional banking waste with all of their buildings and money wasted on executives and theft from customers etc
- endisneigh 3y agoExchanges have buildings and money also wasted on executives as well as theft from customers.
- freedomben 3y agoI think this actually reinforces GP's point somewhat. I could be misinterpreting, but I don't think they weren't defending exchanges, in fact I would guess they would criticize those as much as banks. They were defending bitcoin as a ledger system, which doesn't require any banks or exchanges to work. Exchanges in this system are optional and are a convenience. With traditional banks there is no way to do it without them.
- endisneigh 3y agoDoesn’t change the fact that most transactions occur through exchanges. You don’t need banks to barter either, after all.
- zoklet-enjoyer 3y agoI barely use centralized exchanges anymore. I use them to get dollars in and out, but everything else is on-chain DEXs on proof of stake networks, mostly Osmosis and Evmos
- bugglebeetle 3y agoMany machine learning tasks (e.g. NER) are far less efficient if one uses an LLM compared to, say, BERT models, so I’m not sure this is the case.
- bhickey 3y agoI'm not sure what you mean. BERT has transformer architecture and has a size comparable to GPT-2. Did you have some other model in mind? Or are you trying to draw a distinction between billion and million param models?