A Retirement Journal: Heavenly predictions-AI@Advent

My regular tennis game last Thursday was rained out. More accurately, we called off play even though there was no rain falling during the scheduled time of our game. It had sprinkled a little a few hours before and we called it off because we “assumed” the courts would be wet. Rain in the Bay Area, particularly at the start of the rainy season, can be spotty. Some years the rain is spotty all through the season. Last year was an exception as atmospheric river after atmospheric river slammed into our section of the California coast. However, the first rains in the Bay Area are like snow in Atlanta or Dallas. Everything comes to a halt!

It is always tricky with the atmospheric rivers that produce the bulk of the 15 inches of rain that is Palo Alto’s yearly average. Because the path of these rivers of waters can vary and often miss the Central California Coast and points south, the standard deviation on the average is huge. Having only 5 inches of precipitation one year can be just as likely as getting 15 inches, can be just as likely as getting 25 inches. The last happened last year and dormant lakes came to life.

My friend Mike Rodgers runs a Saturday tennis group. His group has the benefit of a resident meteorologist in Mike, though that is only one of his skills and not the source of his livelihood. When Mike sends out his e-mails regarding play for an upcoming Saturday during the rainy season, his group receives a detailed weather forecast. Mike is like other specialty weather forecasters for agricultural enterprises or airlines. These meteorologists take forecasts from an organization like the National Oceanic and Atmospheric Administration (NOAA) and tailor it for their specific audience. ?Mike’s Peers Park Dogfight Tennis Doubles group gets forecast that explain the likely strength and trajectory of an atmospheric river that might be bearing down on Peers Park in Palo Alto.

NOAA and its counterpart organization the European Centre for Medium Range Weather Forecasts (ECMWF) have broken our globe into over a million surface grids. These grids span 0.25 degrees of latitude and longitude which translates to 28kmX28Km at the equator. Then spherical cube blocks of the same dimension are layered on top of the earth surface cube to reach up into the troposphere. ?NOAA issues forecasts every 6 and ECMWF every 12 hours that span the forecast range from hourly to 10 days. Both organizations use massive clusters of the world’s biggest supercomputers to crunch numbers using physics’ fluid dynamics equations. Because of the complexity of the equations and the number of spherical cube blocks covering the earth to a fine degree of granularity, the amount of compute power required is HUGE. Each calculation takes hours gating the frequency of the updates. Big weather services like AccuWeather and The Weather Channel, as well as meteorologists like Mike Rodgers, work off ?the forecast models like GFS or HRES offered by NOAA and ECMWF respectively.

A few weeks ago, Google’s Deep Mind unit unveiled an AI forecasting system called GraphCast. Using a graph neural networking method of machine learning, GraphCast is proving to be better than the ECMWF forecast, including for one-off events such as hurricanes and atmospheric rivers. The comparison between GraphCast and HRES for hurricanes and atmospheric rivers is shown above. The amazing thing is that GraphCast machine learning algorithm does not need the physics equations. It infers them based on training data since the 70s from both NOAA and ECMWF. To this learning model, one merely inputs two initial conditions- current and current minus 6hours. The model can run in under a minute on something like my MacBook contrasted to the hours on supercomputer clusters for the big deterministic models. Amateur forecasters like Mike will soon be able to run their own Peers Park weather models!

DeepMind’s GraphCast is indeed a super-intelligent AI though not of the potentially sentient kind. The latter is causing concerns of the type I have discussed earlier and that spawned the recent drama at OpenAI. During this Christian season of Advent, I have been ruminating on the comparison and contrast between AI and God, which I call Real Intelligence (RI). The advent of AI effectively began a year ago with the introduction of ChatGPT. I have been putting my concept of RI through the same paces as I am for AI or specifically Artificial General Intelligence (AGI). Assuming RI truly exists, “it” is indeed super-intelligent and sentient. RI like AI can also create hallucinations.

Looking to the heavens has always been the realm of the gods. Like GraphCast, RI, if real, can for sure predict, if not cause weather. ?The Old Testament story of Elijah summoning RI to predict or cause the arrival of rain to break a multi-year drought in ancient Israel is a case in point. Elijah’s “forecast” was accurate to the minute. ?In the Christmas narrative, RI predicted, or caused, a comet or other celestial body to lead wise men known as The Magi to the Christ Child. Maybe, instead of a celestial body prediction, ?RI caused the Magi to hallucinate to arrive at the birthplace of the Christ Child- same effect.

I realize my hubris in putting God or RI through the same paces as AGI! That said, the exercise has driven me to think of both AI and RI in different and unique ways.? Next week is the 4th and final week of the Christian Advent. I will offer my concluding thoughts on RI for this Advent Season- again assuming RI exists.

How about YOU? Are you thinking of RI as you think of AI, whether this Christmas season or otherwise? Whether the Christian concept of RI or other?

Jake

PS: Liz Magill, who I had written about earlier, resigned last Saturday as President of Penn. I will comment more on this in the New Year.