Earth down toilet

The AI Arms Race

Conclusions first:

  • Individual liberty and quality of life in the 21st century hinges on us not being manipulated en masse by AI (we are already losing this race).
  • Massive investments in infrastructure to produce commercial AGI systems are wasteful at best in all but the most utopian scenarios.

You should:

  • Carefully manage your information diet (limit time on your phone - 3H per week being manipulated is more than enough)
  • Build human-centric software applications / share them with the author so I can promote them

Where are we now?

There are two main threads in current AI development:

  1. Many existing succesful applications, many causing harm
  2. New applications being developed, for better and worse
  3. Race for AGI/ASI

Recommender systems powering addictive online propaganda platforms (TikTok, Instagram, YouTube, Netflix) already cause tremendous harm in terms of mental harm, wasted time, lost human connection, and lost sleep [1]. This is not offset by their benefits, even though these benefits are real and significant.

I have been building practical AI applications for 15 years, and the recent progress in GenAI has sped up the development of real applications, good and bad. Good examples: AI systems to speed up and improve healthcare delivery, automate paperwork, and provide 1-1 tutoring. Bad examples: Deepfake scam callers, deepfake revenge porn, targeted political propaganda, target marketing campaigns, all of LinkedIn.

GenAI progress has made building compelling prototypes trivially easy, feeding into the hype around AI while still being generally disappointing with the quality and reliability of end results. Applications powered by the current generation of AI will be (very) economically disruptive, but not in the catyclisimic way people worry about with AI.

An artificial general intelligence (AGI) or artificial super-intelligence (ASI) system would surpass humans in mental skill, but would be fully owned by people, unconscious, and amoral. The purpose of persuing this technology is to make human intelligence irrelevant.

Some people point to a utopian future brought about by AGI. Their actual intentions seem to range from shameless overpromising to raise investment capital, to owning AGI and accumulating more power for themselves, to transcending humanity - very creepy.

The AGI Race

Datacenters are expected to double their share of US power consumption in the next 5 years [2].

As a frequent user of AI APIs and products, I can confidently say that there is currently no shortage of inference capacity - querying state-of-the-art genAI models is fast and cheap today, more than enough to support every working AI application. In other words, there is no real justification for scaling up AI datacenters to build practical applications, this is being done entirely to pursue a scale-centric AGI arms race.

It’s unclear whether continued expansion of compute capacity is the best way to achieve AGI - I believe that algorithmic progress is likely to be more effective (and certainly less of an environmental disaster).

One thing that is certain is that quickly expanding our energy consumption to train huge AI models will come at a large environmental cost (for example, [3]).

References

  1. “Netflix CEO Reed Hastings: Sleep Is Our Competition” 11-06-2017, Rina Raphael
    https://www.fastcompany.com/40491939/netflix-ceo-reed-hastings-sleep-is-our-competition

  2. “AI is poised to drive 160% increase in data center power demand”, Goldman Sachs
    https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand?utm_source=chatgpt.com

  3. “We went to the town Elon Musk is Poisoning”, More Perfect Union, https://www.youtube.com/watch?v=3VJT2JeDCyw