More Thoughts on AI

When it comes to the technology we call AI*, I feel I should very much put my cards on the table and admit that I am personally deeply skeptical about it. Humanity as a whole is devoting huge resources to the development of this new technology, but the decision to devote these huge resources is one that I and indeed most ordinary people were never consulted on.

Further to this, many of the use cases for AI technology that have emerged in the wild seem less than useful. Plagarism, for example has become a huge problem since it allows students to easily create essays with a simple prompt. AI also has been a huge boon to scammers allowing them to flesh out their companies with extensive AI generated websites and marketing materials. AI has in many ways turbocharged the age old internet problem of spam.

Then we get to the problem of power consumption, it's well worth a read of James O'Donnell and Casey Crownhart's excellent piece on the power usage of AI which covers so many aspects of AI and power consumption, some worrying, some less worrying. My main takeaways are:

• Power consumption for actual AI individual queries is actually quite low, for example a simple query with Llama 3.1 8B (8B = 8 billion parameters) uses about 114 joules (or about 1/30,000th of a kWH), for a model with more parameters like Llama 3.1 405B the cost is 6,706 joules (or about 1/500th of a kWH). AI images are actually energy less intensive than text though AI videos are quite a lot more energy intensive.
• While training AI's is expensive, the main power usage will ultimately come through usage of AIs
• Data centres are very carbon intensive since they need 24/7 power and thus it's hard to meet this requirement from purely renewable sources, data centre usage had been roughly flat for a long time as efficiences kept power usage down but the rise in AI has caused the power demand from data centres to surge.
• The big concern is not AI usage now, but the potential agentic future that the AI companies seem to be imagining, where we have tens, or potentially hundreds of interactions with an AI on a daily basis. That AI companies are proposing the building of nuclear power plants suggests they envison a future that will see much larger, more extensive use of AI.

That last point is the key one for me. AI as it exists at the moment is a loss making machine afloat by vast swathes of venture capital and at some point the venture capitalists will want returns on their investment. That return would seem to depend on some kind of mass agentic future where firing of tens or even hundreds of queries off to an AI, which would see a serious rise in power usage and accompanying CO2 emissions. Of particular concern to me are Microsoft, Apple and Google who could potentially use the monopoly power of their platforms lock us in to this future.

Ultimately, as I write and think through AI, I find myself hoping that the AI bubble bursts. I hope that the AI bubble fails and the agentic future the AI companies want us to embrace never comes to pass. I find myself hoping the push to expand AI fails and that ultimately our future with AI comes from salvaging the potentially useful elements from that failed experiment.

* It feels to me the term "AI" is more a marketing term than a descriptive one, a statistical model that generates outputs based on a prompt based on statistical probabilities derived from human created input certainly falls well short of the sci-fi vision of AI as a kind of artificial self-consiousness.