In context of benefit to society this should be a tremendously smaller amount as what we do to drive

For now, while we only have a few small models, sure.

But models are inefficient and always will be by the nature of being general purpose.

You don’t need a language model to tell you about your carbon footprint report. You could just use some of your human brain time to understand instead.

Not saying that AI is inferior always to human Intelligence, but it’s much like driving a car vs walking. One is much faster, but has a much higher impact on the environment

If the prices are cost, and the cost is electricity not hardware, and the electricity is $0.05/kWh, every 1000 tokens (750 words) of ChatGPT is $0.002 = 1/25 kWh = 40 Wh.

Average human reading speed varies by language but one average over many languages is 184±29 wpm, so those 750 words will take you about 4 minutes (4.07) to read.

If you're an American, your average electric consumption in that time is 1,387 W * 4.07 minutes ≈ 94 Wh.

But the human still exists.

(You compute your carbon footprint per minute by adding up all the things you do in a large period of time and dividing by the duration. This chatgpt footprint should be added to the numerator, it doesn't replace the whole thing)

But the human requires oxygen and water and lots of climate control.

If AI is useful as we think it is, then having huge compute farms in space would be the best long term goal.

Sure, I'm just saying ChatGPT is low energy even compared to the opportunity cost of reading its output.

You can run a ChatGPT-like model now using LLaMA + Alpaca on a regular CPU device in around 4GB of RAM. So if you want a low carbon ChatGPT-style model there are great options now.

To put it in context, ChatGPT would have to be forthcoming about their footprint. Which they are not.