Great piece Matt, this sort of thing is clearly on the way, and these simple economic simulations (games) are exactly why Google DeepMind has been focusing on go, chess, Starcraft etc, games with reasonable degrees of freedom to try things and see which optimise the outcome. The two keys as I see it are (1) the mapping from simulation to world (we probably need simulations more in line with the complexity of those underpinning weather forecasts) and (2) educating people (eg at school) by letting them play with such simulations (as games) to understand how tweaking them can change outcomes and how they outperform intuitive and 'obvious' human variable settings. Ultimately with oversight, so that the settings don't get set to something just obviously crazy, like taxing a certain group at 100% because it enhances the outcome.
Thanks for this comment Matt. I agree that gamified simulations are laying the groundwork for something much more expansive and significant (say, AI economists). There is a huge gap between the complexity of current simulations and what is needed, but reinforcement learning AI continues to show an enormous amount of promise In terms of the settings, I suspect there would be a number of parameters set in the speculative world I set out in the piece - potentially giving an AI a spectrum of possible tax/economic policies, rather than a blank slate. That would also help achieve some predictability and consistency between elections, and avoid an AI lurching hugely between policies depending on what party gets to set its priorities. Or, as you say, some governance or high-level oversight.
Great piece Matt, this sort of thing is clearly on the way, and these simple economic simulations (games) are exactly why Google DeepMind has been focusing on go, chess, Starcraft etc, games with reasonable degrees of freedom to try things and see which optimise the outcome. The two keys as I see it are (1) the mapping from simulation to world (we probably need simulations more in line with the complexity of those underpinning weather forecasts) and (2) educating people (eg at school) by letting them play with such simulations (as games) to understand how tweaking them can change outcomes and how they outperform intuitive and 'obvious' human variable settings. Ultimately with oversight, so that the settings don't get set to something just obviously crazy, like taxing a certain group at 100% because it enhances the outcome.
Thanks for this comment Matt. I agree that gamified simulations are laying the groundwork for something much more expansive and significant (say, AI economists). There is a huge gap between the complexity of current simulations and what is needed, but reinforcement learning AI continues to show an enormous amount of promise In terms of the settings, I suspect there would be a number of parameters set in the speculative world I set out in the piece - potentially giving an AI a spectrum of possible tax/economic policies, rather than a blank slate. That would also help achieve some predictability and consistency between elections, and avoid an AI lurching hugely between policies depending on what party gets to set its priorities. Or, as you say, some governance or high-level oversight.