Notes 📍

Sure seems like that times when you read the alarmist headlines. AI is going to be everywhere, it’s going to use up all the electricity, demand building millions of gas-guzzling power plants, put everybody out of work, break every security code, and give the everyman the information they need to build nuclear and biological bombs in their backyard.

I’ve been quite interested in AI, getting to the bottom of it, avoiding the hype. It seems like much of AI is just hype – all of the proposed AI data centers if built would increase the United States’ data center capacity by 15 fold. As if the world needed that kind of compute. Not to mention, whose paying for it and building all the electric plants to make it happen? Much of it is just speculation. People chasing money, on the possilbity of a need.

Likewise, much of the demand of the big US AI companies for cheap electricity, environmental exemptions, restrictions on the use of open models, seems primarily to be about incumbent US players. If you can’t build a better product, then you can try to get Congress to pass a law banning the use of your competition. But over time, the best product wins. If China is making better, more efficent models, then their open models will win. Americans have long had a very unhealthy obession with bigger is better.

I am a believer in small, efficent models are often better. I hate the idea of a few centralized companies controlling AI, I would rather have billions of small, task specific AI models running on individuals computers rather then some slow API calling distant know-it-all models on server farms hundreds or thousands of miles away. Most of the time, specific knowledge of a task beats encyclopedia knowledge. General reference has it’s place in this world, but most tasks are narrow and don’t require terabytes of data and knowledge to process.

Truth is things like Google Gemani are fun, and can help you think about problems and provide data in a way like an encyclopedia does. But they are generalists, and when you dig into them a bit and ask specific non-encylopedia type questions, they aren’t very good. They’ll try to churn out a probable answer and sometimes good answer, but at their roots, their little more then a giant encyclopedia that makes good predictions on what you want to hear. And I’m not sure we need 15 times more of their capacity. We need more efficiency in these models, something the Chinese are good at, but also we need more task-based, local models rather then these big encyclopedia remote models.