The race to collect the data and make it useful for predicting our fate
[Apologies for the late arrival, I was getting over a COLD (not COVID!), which I suppose is just a sign that nature is healing. Yay.]
Why aren't we better at modeling Earth?
This has been on my mind recently — and not just because of COVID, although that's certainly a big piece of it. But with things like climate change and politics flooding us with predictions and probabilities and forecasting — you'd think we would be a lot better at it by now.
Shouldn't it be the case that as we see improvements in technologies, sensors, and computing power — we would be seeing significant advances in modeling itself?
A couple of recent stories sparked this for me.
Simulating Model UN
First, the UN is backing a new simulation tool, the Policy Priority Inference, that uses a combination of agent-based modeling to compare outcomes across spending priorities. In theory, this will make it easier for governments to prioritize various scenarios. (If you're a super nerd on this, there's an excellent white paper and link to their GitHub repo). It looks like this particular approach is to essentially plug-in as much government data as possible, including traditional measures of development indicators, but also things like policy goals, trends, rule of law, etc. In this way, the model hopes to simulate a more realistic situation where official government data would not always be trustworthy. The model can then track progress in a more objective and identifiable way.
Predicting with Satellites
Second, a fascinating tool from a team at Stanford looking just at satellite data to estimate patterns of growth, poverty, and development over time. The model they developed was trained by looking at satellite imagery for 200,000 villages, compared with collected data about those communities for at least ten years. The imagery would show things like buildings, roads, agriculture, etc. The result is a type of pattern recognition that could at least get you a ballpark estimate for an area for which you have no data (or no recent data, as communities grow and shrink). We can do similar analyses for ocean pollution as well.
Learning from Video Games
Yes, reading about agent-based modeling scenario planning tools and playing video games are essentially the same.
Actually, this is a little bit true. Suppose you reach back to the origins of Sim City, the video game that defined the genre of simulations. In that case, you'll find the creator in 1984, Will Wright, actually borrowing heavily from a 1969 book entitled Urban Dynamics by Jay Wright Forrester. Forrester had created a sort of blueprint for modeling equations for the growth and decline of cities — which were controversial even at the time for leaning a little too heavily on hands-off approaches to business and innovation.
Modeling for Granted
None of these areas have it all mapped out — from government policy data and trends to satellite sensor data to sophisticated game mechanics and modeling. And that was very clear as we struggled to understand the evolution and expanse of COVID-19 over the last two years.
It's just — tough! And it's no surprise that not a single model or forecasting approach has become a household tool (though full disclaimer that there are many models out there that are being used by health authorities even if they aren't household names!)
There's so much data we would need to have, and even the data we do have is typically not of a quality necessary to provide a meaningful seed for modeling.
Interestingly, one area that has seen steady improvements has been weather forecasting. The weather includes an enormous amount of data. There are sensors worldwide collecting raw data, and there are time-series data that have been continually collected, sometimes stretching back over more than a hundred years. And there are many organizations, agencies, and people with vested interests in working models, forecasts, and accurate predictions.
It's a remarkable system.
GPT-3 as a Model
When I think of the volume of data that we'd need to see meaningful improvements in more meaningful Earth simulations, I think of what the team behind GPT-3 used to get their language model to its current state. GPT-3 is a natural language processing model built using 175 billion language parameters (and around $12M for a single training run that is great, but not perfect, at generating believable human speech.
What GPT-3 has done has got to be probably the minimum level necessary for believable Earth simulations involving so many entangled, overlapping, co-dependent variables of human and environmental interactions.
Just Getting Warmed Up
All of this is on my mind a bit anyhow, as I've been tinkering with an impressive COVID modeler that uses network analysis to simulate how policy planners can think about planning at tiny scales of human interactions. In this case, looking at the classroom level (think how elementary school students stay in one class all day with the same peers versus high school students who mix in different classrooms throughout the day). And even then, it's at a small scale, but the math gets very complicated very fast!
I recently learned of a new effort from EU scientists to create a digital "Earth 2" to do what we've been talking about here.
Let's hope they learn the fitting lessons from Sim City in the first place.
What I've Been Reading
How Shein beat Amazon at its own game — and reinvented fast fashion - Rest of World — As someone who knows nothing about fast fashion (or fashion), I thought this article does an excellent job of framing Amazon, manufacturing, supply chains, etc. in an accessible way.
Autocracy Is Winning - The Atlantic — Sigh. Yeah, I mean, yeah.
How the FBI Discovered a Real-Life Indiana Jones in, of All Places, Rural Indiana | Vanity Fair — Wow. This story is pretty incredible — and that one that I just have to assume could not possibly happen were things 40 years later, much less today?
On "Succession," Jeremy Strong Doesn't Get the Joke | The New Yorker — I mean, c'mon. This is still incredible.
Thanks for reading,
GoodTinker is a weekly email from Gabriel Krieshok about technology, design, and social impact. If you've enjoyed this edition, please consider forwarding it to a friend. If you're reading it for the first time, consider subscribing (it's free!).