Superhuman intelligence already exists
And how it has affected the game of chess tells us something
People worry the new, generative AI models are going to steal jobs. We hoped for a world where the machines did the menial physical labour, freeing us to create art and pursue our passions. Now it looks more like the vision of commercial AI is that they will handle creating art, poetry and doing research, leaving us stuck with the menial physical labour.
Don’t panic. I come from the future with a message of hope.
Not, of course, literally from the future, but from a domain where superhuman intelligence has been the reality for decades.
In 1997 Chess world-champion Garry Kasparov lost to the IBM supercomputer Deep Blue, auguring in the age in which Chess computers dominate human players. Now, the chess app on my phone can play at grandmaster level, and the better chess engines effortlessly outpace the best human players.
Chess uses a rating system called Elo which calibrates all players according to their wins and losses against other players. A beginner might have an Elo of around 1000, and strong club player 1800. At the time he lost to Deep Blue Gary Kasparov had an Elo of 2795. The highest human Elo of all time is 2882, achieved by current highest rated player Magnus Carlsen. Now, the best AI computers have ratings above 3600.
From the perspective of an ordinary human, this is like playing chess with God.
§
The effects of Kasparov’s defeat on the chess world were immediate.
Globally, people stopped playing chess. Clubs folded and a bitter argument broke out over whether to allow computers to play in the world tournament against humans, or whether to have the world tournament entirely played by computer with no humans involved at all. After all, why watch humans when from this point on the best chess games would be played by machines, not humans?
Just kidding, the opposite of this happened. Chess has enjoyed a massive boom in popularity in the last few decades. More people play chess than ever before. There has been a large increase in younger players, and particularly in younger female players. The world (human v human) chess tournament is better attended (in person) and more watched (online) than ever before. Across the street from me a local bar has just started a chess and vinyl night on Wednesday evenings (admittedly this last one is anecdotal).
Why didn’t everyone quit when they saw that the machines could do it better than them? Why does anyone still want to watch games from the inferior human players?
The answers to these questions have important lessons for all the other domains where AI may outperform humans (or perhaps just substitute in a mediocre way that is good enough).
The lesson is that sheer performance is not what gives something meaning. Chess is popular because we like playing it. It means something because of the effort champions put in to honing their skill, because we know and value the individuals in pursuit of that goal. Something in us wants to see the best - human - chess player in the world win the championship, and that is why the game still has an audience. And something in us wants to develop our own skills, and that’s why we haven’t abandoned the game to the machines.
AI can mimic human performance, sometimes exceeding it. It will have astounding successes, and weird failures, but regardless it does not value things, nor intend things. This is where the meaning derives from, and it’s why we still go to see live music despite having limitless access to artificial (recorded) music, and why we’ll still read novels regardless of whether a computer can write them or not - we’re always going to want to hear communication about what it is like to be a particular human in a particular place.
§
EDIT: realised this post really should include the image which comes up when the Lichess servers are down. Credited to “gia”.
§
The advent of commercial-grade superintelligence in chess has, however, changed the game. When it became possible for anyone to play against AI the skill level of chess played increased. One of the things which make chess a great game to research, as well as play, is that you can use the chess engines to estimate the quality of *each* move made. Strittmatter et al (2020) took data from 125 years of professional chess tournaments and looked at the proportion of moves which matched the “optimal” move according to the chess engine Stockfish (an industry standard). Here’s part of their results: the share of optimal moves according to player birth year.
Two things are clear: standards in chess have been steadily increasing, and they start to accelerate for players born after ~1980. This is the age when these players were young enough to still be learning when chess superintelligence was available to every home.
§
The example of chess tells us a few things about how the new AI will change us.
It shows that ultimately the effects will be downstream of what we value. The problem with the current models is that the values are coming from large corporations with minimal accountability. To me that seems a bigger issue than theoretical speculation about what AI might eventually be able to do (or not).
If we replace human jobs it will be because of value choices we make collectively. That we still play chess, and gather to watch chess tournaments, suggests - just as we still pay to see live music - that meaningful human activities will stay meaningful, regardless of whether a machine can “do” them or not.
Further reading
This is good on the economics of how AI will affect jobs: Automation and the value of expertise (4 Jul 2025). David Autor from MIT argues that true automation is much harder than assistance, and that the economic effects of assistance depend on how much expertise is left in the "bundle" of tasks that make up an occupation. If the machine does the unskilled part, you can do more (and maybe charge more); if a machine does the skilled part, more people can do your job (and you will maybe have to charge less). So a cash register does a skilled part of the cashier work (adding up), meaning that more people can work shop counters, and hence leading to a decreasing economic power (salary) of that work. The prediction would be the opposite for a job where the most expert, hardest to achieve, part was the human interaction and the adding up was a drag on the comparative advantage of the cashier.
This is good on the so-called Alignment Problem and how to make sure AI models follow the "chain of command" of values built into them.
ACX: Deliberative Alignment, And The Spec (12 Feb 2025)
"I worry we’re barrelling towards a world where either the executive branch or company leadership is on top; you can decide which of those is the good vs. bad ending. But I find it encouraging that people thinking about third options."
Catch up
This was the fifth in a mini-series on concepts for thinking about AI.
…And finally
A woman seated, c.1645
Govert Flinck
(British Museum)
(via @jdmccafferty)
END
Comments? Feedback? 1. d4 d5 2. c4 ? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online
Point well made, thank you. Humans strive; they don't just perform. High performing machines don't render our efforts moot any more than the existence of Olympic champions and professional athletes I will never match in their domains make it pointless for me to do those activities as well as I can. Pieces of the framing from the olden days of humanistic psychology is still relevant: winning and breaking records is often tacitly assumed to be the goal of doing something, but a more valuable framing may be that winning is just one among various means to the more flexible, broader, longer time frame, and more human end of striving and growing. The chess expertise example seems very powerful to me.