Energy for AI
Posted
Have you read Leopold Aschenbrenner’s argument for a ‘Manhattan Project’ to race to Artificial General Intelligence (AGI)? I admit to being swayed by some of the thinking. There’s certainly holes that can be picked in the arguments: here’s three that a good friend of mine called out:
- much of the data that an AGI might need for training is locked up in organizations and, even if access is given, is poorly formatted. Further, a lot of data out there is garbage
- no mention is made of quantum computing. That could upend much of the thinking
- any AGI will be concentrated in a few data centers with very visible power supply. In a race between nation states that’s relatively easy infrastructure to destroy
Leopold identifies energy as one of the key constraints to achieving AGI, and energy for AI was a key theme at the recent ENACT Majlis.
The Majlis agreed the answer to much greater energy demand for AI will be an ‘all of the above’ with gas, nuclear and renewables all part of the mix. I’m no expert on the supply side but I find myself doubting some of this demand thinking:
-
AI can be distributed: Apple Intelligence has been a bit of weak launch but the trend that our cell phones will run complex models and work can occur on the edge is only going to increase as models become more efficient and silicon more powerful.
-
AGI is likely further away than optimists like Aschenbrenner think it is. We’ve seen this trend in technology time and time again and particularly in AI. Further, there is some evidence that GenAI is not penetrating into companies as quickly as vendors expect as The Economist noted last week.
-
The biggest opportunity: we have lots of room for optimization. Physics shows us the hard bounds that we can convert power to motion or heat. Nature shows us that it takes 20W for NGI in the human brain. We are many orders of magnitude away from that right now with LLMs. If there is demand for AI then that creates massive market demand for efficient AI. Whether it’s binary neural nets, more efficient silicon, or knowledge graphs to ground models, there are many options to chase after.
There are many drivers of growth of energy demand. The vast majority of these are distributed with numerous market participants and regulators: EVs, developing economy growth, transition to heat pumps, decarbonization of emissions intensive industries to name but a few.
By contrast, AI is controlled and supplied by a few vendors who have massive incentive to keep their power consumption down and the deep pockets to drive that.
Will energy demand continue to grow? Of course. Will AI be a material part of that growth? I’m not so sure.