Re: “researchers at Stanford and Google created a simulated mini-town called Smallville, and asked the 25 AI inhabitants to organise a birthday party”, I’m way less optimistic than you are.
This is basically what’s called “agent based modeling” in complex adaptive systems circles. For example, SFI’s Complexity Explorer has offered courses using NetLogo programming for years. While simulations can layer in more complex rules and numbers of agents, the main shortcoming is that it’s nothing like the predictive modeling most people expect.
Initial conditions and the non-linear, emergent, and self-organizing behaviors can create very different outcomes with each simulation run. The best predictive analyses thus come down to Monte Carlo-level mass simulations, hoping to draw out some select aggregation of scenarios happening more probabilistically than others. Thus it’s less like predicting the weather and more like the ever-flawed AI simulations of who will win the World Cup - the latter of which is frequently outdone by octopuses and armadillos.
I sincerely doubt this is the anticipated outcome you’re excited about. But there’s still magic in the uncertainty despite our best analytical efforts.
Thanks for reading Swag! I agree that social simulations of the kind we're talking about can never be prediction machines. Human behaviour is far too complex, and subject to random chance, for that. But perhaps we'll be able to run thousands or even millions of parallel simulations, and develop a better picture of the balance of probabilities when it comes to certain outcomes. Also, there will be lots of use cases for these simulations beyond just telling us 'what is going to happen next'. We could, for example, use them to study the transmission of ideas through a social collective, or the behaviour of markets under different conditions.
There's probably a bit we could do with Monte Carlo simulations and seeing what the categorized future scenarios might look like in aggregate. But unfortunately that's a better press release than it is a useful outcome, IMO.
Since simulations are always missing data, especially for new novel situations (e.g., COVID), I have very low confidence that we'll ever get them to tell us that a WHO proclamation of a global pandemic will have even one scenario showing a global run on loo roll. But retrospectively we could develop some insights on the gaps as you say.
Re: “researchers at Stanford and Google created a simulated mini-town called Smallville, and asked the 25 AI inhabitants to organise a birthday party”, I’m way less optimistic than you are.
This is basically what’s called “agent based modeling” in complex adaptive systems circles. For example, SFI’s Complexity Explorer has offered courses using NetLogo programming for years. While simulations can layer in more complex rules and numbers of agents, the main shortcoming is that it’s nothing like the predictive modeling most people expect.
Initial conditions and the non-linear, emergent, and self-organizing behaviors can create very different outcomes with each simulation run. The best predictive analyses thus come down to Monte Carlo-level mass simulations, hoping to draw out some select aggregation of scenarios happening more probabilistically than others. Thus it’s less like predicting the weather and more like the ever-flawed AI simulations of who will win the World Cup - the latter of which is frequently outdone by octopuses and armadillos.
I sincerely doubt this is the anticipated outcome you’re excited about. But there’s still magic in the uncertainty despite our best analytical efforts.
Thanks for reading Swag! I agree that social simulations of the kind we're talking about can never be prediction machines. Human behaviour is far too complex, and subject to random chance, for that. But perhaps we'll be able to run thousands or even millions of parallel simulations, and develop a better picture of the balance of probabilities when it comes to certain outcomes. Also, there will be lots of use cases for these simulations beyond just telling us 'what is going to happen next'. We could, for example, use them to study the transmission of ideas through a social collective, or the behaviour of markets under different conditions.
There's probably a bit we could do with Monte Carlo simulations and seeing what the categorized future scenarios might look like in aggregate. But unfortunately that's a better press release than it is a useful outcome, IMO.
Since simulations are always missing data, especially for new novel situations (e.g., COVID), I have very low confidence that we'll ever get them to tell us that a WHO proclamation of a global pandemic will have even one scenario showing a global run on loo roll. But retrospectively we could develop some insights on the gaps as you say.