The costs are significant and growing but we should put some things into perspective to really tackle the problem efficiently. As an individual, heavy usage of these tools (something like 1000 images generated) is still roughly the same level of emissions as driving across town and generating text is pretty much negligible in all scenarios.
Where we really need to be concerned is video generation (which could easily blow current energy usage out of the water) and water usage in these massive data centers. However, most of the current research on the subject does a pretty poor job of separating water usage for “AI” and general usage. This is why the next step is enforcing transparency so we can get a picture of how things are shaping up as this technology develops.
All that said, there are some pretty low hanging fruit when it comes to improving efficiency. A lot of these models are essentially first-passes on a project and efficiency will improve simply as they start to target edge and local models. Similarly, these water cooling systems are predicated on some fairly wasteful ideas, namely that cool fresh water is abundant and does not warrant preservation. Simply factoring in that this is clearly no longer the case will go a long way towards reducing that usage.
To address the article a little more directly: it’s notable that the article begins with Sam Altman’s take on the subject. His feelings are based on two fundamentally flawed premises:
These models MUST get bigger for the improvements that their users DEMAND.
The only solution to any environmental criticism is FUSION. A technology that Altman has personally invested in.
2 is ridiculous just on the face of it, but I think folks may have a harder time understanding why 1 is problematic. It is true that OpenAIs business model essentializes the idea that these models can’t ever be run locally, but the incentive to use their cloud services are quickly diminishing as smaller, local models catch up. This cycle will likely continue until local models are good enough to serve the needs of the vast majority of people, especially as specialized hardware makes it’s way into more and more consumer devices.
The costs are significant and growing but we should put some things into perspective to really tackle the problem efficiently. As an individual, heavy usage of these tools (something like 1000 images generated) is still roughly the same level of emissions as driving across town and generating text is pretty much negligible in all scenarios.
Where we really need to be concerned is video generation (which could easily blow current energy usage out of the water) and water usage in these massive data centers. However, most of the current research on the subject does a pretty poor job of separating water usage for “AI” and general usage. This is why the next step is enforcing transparency so we can get a picture of how things are shaping up as this technology develops.
All that said, there are some pretty low hanging fruit when it comes to improving efficiency. A lot of these models are essentially first-passes on a project and efficiency will improve simply as they start to target edge and local models. Similarly, these water cooling systems are predicated on some fairly wasteful ideas, namely that cool fresh water is abundant and does not warrant preservation. Simply factoring in that this is clearly no longer the case will go a long way towards reducing that usage.
To address the article a little more directly: it’s notable that the article begins with Sam Altman’s take on the subject. His feelings are based on two fundamentally flawed premises:
2 is ridiculous just on the face of it, but I think folks may have a harder time understanding why 1 is problematic. It is true that OpenAIs business model essentializes the idea that these models can’t ever be run locally, but the incentive to use their cloud services are quickly diminishing as smaller, local models catch up. This cycle will likely continue until local models are good enough to serve the needs of the vast majority of people, especially as specialized hardware makes it’s way into more and more consumer devices.
Yeah I don’t think me using llama on my 10 years old cpu will do any harm :3