AI will develop, and therefore will inevitably consume more energy, even material, than when it did not exist.
Atlantico: https://twitter.com/aympontier/status/1782795678520901948?s=43. Will this development help resolve the problems of computer chip shortages?
Pierre Beyssac: In part: this can potentially reduce pressure on the needs for graphics chips (GPUs), widely used for AI, by allowing their substitution by less demanding architectures.
We can already run, in use, “simple” AI on mobile phones, laptops or very light computers (Raspberry). Resource requirements have been greatly reduced over the past 18 months by the talent of numerous researchers and free software developers.
More generally, do you think that the results can allow us to move towards less resource-intensive technology?
Certainly: this trend will increase. We could see a double trend, lightweight AI sufficient for the most common needs, and with low resource consumption; and some heavy AI trained by big, deep-pocketed companies, which seem to be Mark Zuckerberg’s choice at Meta.
It is unclear what the limitations of current natural language processing AI technology are. There are probably some, which will prevent them from progressing as is “infinitely”. We should therefore sooner or later observe a stagnation in performance, which will put an end to the raw race for resources, while awaiting new algorithmic innovations to overcome the pitfall.
But isn’t such performance likely to be even more energy intensive?
AI will develop, and therefore will inevitably consume more energy, even material, than when it did not exist.
Optimizations make it possible to reduce the impact, but by making the technology more resource-efficient, conversely promote the development of uses, therefore a re-increase in impact. It is common to speak of a “rebound effect”. At worst, this can even go so far as to swallow up the initial optimization gain. We then speak of the “Jevons paradox”, named after the economist who theorized it.
But, when it comes to new technologies, the growth in uses is largely attributable to the reduction of the digital divide, an objective defended by all. Optimizations therefore remain beneficial.
Finally, we must not neglect impact transfers and positive externalities. For example, if AI allows me to spend only 15 minutes instead of 2 hours on my computer, by assisting me in my task, I save energy.
