When I AI code, it's more convenient for me to type in a prompt to change the name of a variable. This involves sending a request to the LLM provider, doing very expensive computations, and then doing a buggy job at it. Even though my IDE can do it for what 0.1% of the energy usage? Or even less? Try running an LLM locally on a CPU and you'll get a glimpse of how much energy it is using to do this simple task.
But coding with AI isn't that huge. What will become huge this year or next is MCP. It will bring "programming" to the masses, all of whom will do stupid queries like the above.
Consider this: I wrote an MCP server to fetch the weather forecast, and have separate tools to get a broad forecast, an hourly forecast, etc. I often want to check things like "I'm thinking of going to X tomorrow. It will be a bummer if it's cloudy. Which hours tomorrow have less than 50% cloud cover?" I could go to a weather web site, but that's more effort (lots of clicks to get to this detail). Way easier if I have a prompt ready.
OK - that doesn't sound too bad. Now let's say I want to do this check daily. What you have to realize is that with MCP the prompt above is as good as a program! It's trivial for an average non-programmer to write that prompt and put it as part of a cron job, and have the LLM email/text you when the weather hits a predefined criteria.
Consider emails. I sign up for deals from a retailer.[1] Now deals from them are a dime a dozen so I've been programmed to ignore those emails. But now with MCP, I can set a simple rule: Any email from that retailer goes to the LLM, and I've written a "program" that loosely describes what I think is a great deal, and let the LLM decide if it should notify me.
Everyone will do this - no programming required! That prompt + cron is the program.
Compared to traditional programming, this produces 100-1000x more CO2 emissions. And because there is no barrier to entry, easily 1000x more people will be doing programming than are doing now. So it's almost a millionfold in CO2 emissions for tasks like these.
When I AI code, it's more convenient for me to type in a prompt to change the name of a variable. This involves sending a request to the LLM provider, doing very expensive computations, and then doing a buggy job at it. Even though my IDE can do it for what 0.1% of the energy usage? Or even less? Try running an LLM locally on a CPU and you'll get a glimpse of how much energy it is using to do this simple task.
But coding with AI isn't that huge. What will become huge this year or next is MCP. It will bring "programming" to the masses, all of whom will do stupid queries like the above.
Consider this: I wrote an MCP server to fetch the weather forecast, and have separate tools to get a broad forecast, an hourly forecast, etc. I often want to check things like "I'm thinking of going to X tomorrow. It will be a bummer if it's cloudy. Which hours tomorrow have less than 50% cloud cover?" I could go to a weather web site, but that's more effort (lots of clicks to get to this detail). Way easier if I have a prompt ready.
OK - that doesn't sound too bad. Now let's say I want to do this check daily. What you have to realize is that with MCP the prompt above is as good as a program! It's trivial for an average non-programmer to write that prompt and put it as part of a cron job, and have the LLM email/text you when the weather hits a predefined criteria.
Consider emails. I sign up for deals from a retailer.[1] Now deals from them are a dime a dozen so I've been programmed to ignore those emails. But now with MCP, I can set a simple rule: Any email from that retailer goes to the LLM, and I've written a "program" that loosely describes what I think is a great deal, and let the LLM decide if it should notify me.
Everyone will do this - no programming required! That prompt + cron is the program.
Compared to traditional programming, this produces 100-1000x more CO2 emissions. And because there is no barrier to entry, easily 1000x more people will be doing programming than are doing now. So it's almost a millionfold in CO2 emissions for tasks like these.
[1] OK, I don't do it, but most people do.