Much of the talk around AI is hyperbolic, speculative, and often misses the point. In our attempts to recreate our own intelligence using conventional computers we have ended up creating energy inefficient and over engineered systems that rely more on computation scale than “intelligence”. If we want intelligent machines as efficient as our own brains, then we must mimic how it computes. Following recent developments in architecture and learning, Jonathan Peters argues that the future of AI will be analogue.

The rise to prominence of artificial intelligence (AI) technology is arguably the most important technological change since the turn of the 21^st Century. So rapid have the technological advancements been that in a matter of a few years global understanding and perspectives on AI have changed from ignorance to a mixture of intrigue, curiosity, and fear of the unknown future consequences.

When most think of concerns regarding AI technology, the prospect of robot-like technology terrorising humans and taking over the planet will often come to mind. Whilst this is dramatic and highly unlikely in the near future, one concern people should be more informed about is the implications of the technology on the climate. An example of this naivety is clear when looking at the recent EU ‘artificial intelligence act’, dubbed the ‘first of its kind’ in the world, does little to address climate threats. The bill itself asks companies to self-report climate-related statistics, such as consumption, however there is no limits/punishments for extreme energy usage.

Statistics put into perspective just how serious the problem really is. Academic research predicts by 2027, AI alone will be responsible for the same amount of energy expenditure as countries such as the Netherlands, Croatia and Kenya.  Other research by the International Energy Agency forecasts energy use from data centres powering AI will be equal to that of Japan within two years.

Statistics so damning for the environment beg the question of why policy makers are not doing more now, especially with the Paris Climate Agreement in mind. But the answer is simple: there is a lack of knowledge on the subject. AI’s commercial profitability has brought secrecy from the largest players in the industry such as OpenAI, Google and Microsoft. This secrecy is competition driven, although sceptics would also suggest knowledge of such information could damage public interest in the technology, harming long-term business.

___

To cause meaningful change in the AI industry, the technology must be attractive... If not as profitable as current technology, large companies responsible for emissions will not adopt novel systems.

___

Reading this discussion, some people may argue that the power of the technology to change lives vindicates its extreme power consumption. Whilst the ability of AI to change lives is not disputed, the irony of praising the ability of current AI machines is that our brains are far more efficient AI computers.

SUGGESTED VIEWING Consciousness and ChatGPT With James Tartaglia

As an example, take chatbots: AI models designed to have human interactions online. In computing jargon, these often fall into the category of (large) language models (LLMs), and generative AI, which are notorious for requiring large networks to build successful models. GPT4 (just one LLM) has a daily power usage is equal to roughly 180,000 households in the USA, even after the more energy intense training phase is complete. Predictions suggest models will become much larger very quickly (Figure 1), which consequently will rapidly increase the technology’s energy consumption and carbon footprint.