While Scott Van Pelt opens his version of "SportsCenter" on ESPN with "The Best Thing I Saw Today," I can't limit myself to just one thing here. I've seen a whole bunch of smart things over the past week that I'd like to share as a sort of grab bag.
Several relate to AI -- even though we all feel like we're inundated with news about the field. The one that could be most important for insurance concerns what may be a breakthrough in how we forecast major storms -- three models based on a new approach were extremely accurate in forecasting the path of Hurricane Lee in mid-September, beginning when the storm was thousands of miles from North America.
As the Washington Post reported, "The models are orders of magnitude faster and cheaper to operate than conventional, government-run weather models. While AI models don’t yet provide all the capabilities needed for operational forecasting, their emergence portends a potential sea change in how weather forecasts are made."
Let's have a look.
The new models take me back to the late 1980s and into the 1990s, when "expert systems" were the cutting edge in AI. The basic idea was that you find an expert and build a system around that person's (or those persons') expertise. You ask or observe how that stock trader, or plant operator or whomever made decisions in as many circumstances as you could imagine, then build software that would make those same decisions in those same circumstances. The problem was that you could never imagine all the circumstances or draw all the expertise out of the person, who often had developed gut instincts over time that they'd act on in, say, a stock market crash but that they wouldn't know how to articulate ahead of time. So the limits of so-called rules-based systems became clear.
The breakthrough came when computing power became so plentiful that AIs could be turned loose to simulate a nearly unlimited set of possibilities and to see what responses were optimal. This simulation approach is what has led, for instance, to the AIs that have defeated the world's best Go players. The initial AI was a bit of a hybrid -- it built on a base of human expertise. That was then surpassed by an AI that was simply given the rules of the game and learned by playing billions of games against itself. That, in turn, has now been surpassed by an AI that started without even being given the rules as it started its simulations.
Go players talk with reverence about Move 37 in game two of a series that an AI, AlphaGo, played against Lee Sedol, a top-ranked player, in a series that the AI won 4-1. The AI's move went against all the precepts of Go that have been taught for centuries but, analysis now shows, was brilliant.
The switch from expert systems to endless simulations and machine learning is, very roughly speaking, the change that may be beginning with weather forecasting for major storms.
At the moment, the two main approaches -- one developed in Europe, one in the U.S. -- operate based on data and knowledge that has been collected and developed over decades and that has been turned into extremely elaborate models. A supercomputer needs perhaps an hour (and a lot of electricity) to conduct trillions of calculations and turn those formulas into a forecast about the path and intensity of a hurricane.
The new models -- produced by Google, Microsoft, Nvidia, Huawei and a number of startups -- start with the extensive data on weather conditions that are collected for the supercomputer-based models but ignore all the formulas that the supercomputers then use to generate forecasts. The new models are built on deep analysis of decades of prior weather data and, based on the patterns discerned, can produce a forecast on a desktop computer in a minute, or even seconds.
In the case of Hurricane Lee, the new models accurately predicted on Sept. 10 that it would make landfall in Nova Scotia six days later and were ahead of the established models in suggesting that the hurricane might travel close enough to Cape Cod to produce severe weather.
The hope is that the speed and low cost of operating the new models will also allow for what are called ensemble forecasts. The new models could be used to generate a whole series of forecasts based on slight variations in the weather data it's fed -- which can be imprecise -- and generate a range of forecasts that would provide a more robust look at how a storm might behave.
"Ensemble forecasts from conventional models can miss extreme events, such as excessive rainfall or heat, because they are limited to about 50 simulations due to the time and cost of generating them," the Washington Post article said. "AI could enable the generation of much larger ensembles in as little as a few minutes, potentially leading to more useful forecasts and risk assessments for emergency managers, the general public and numerous industries.
"'Our hypothesis is we can easily now scale up with AI models to thousands or tens of thousands of ensemble members,' Anima Anandkumar, senior director of AI Research at Nvidia, said in an interview."
Given the normal trajectory of technology, I can imagine this new sort of model moving from hurricanes to other sorts of severe storms, including tornados and derechos, which have historically been less predictable. In time, I could even imagine these models being used to warn of the sorts of severe thunderstorms that dumped 25 inches of rain on Ft. Lauderdale in April and that hit the New York City area with as much as 10 inches of rain last week. This sort of severe storm is a relatively new phenomenon, apparently related to high temperatures in nearby ocean waters, but AI could well recognize the signs that what looks like a routine thunderstorm could actually last for many hours.
One storm obviously doesn't prove anything about the new models, but they're off to a good start, and AI tends to improve rapidly once it gets its arms -- brain? -- around something. Any improvement in forecasting does, of course, increase the odds that people can protect themselves and their property, assisted by their insurers.
That explanation took longer than I expected, but let's still get to the other smart things that caught my eye in the past week:
--The clearest example I've yet seen of ROI from generative AI, from an interview in Fortune with Erik Brynjolfsson:
"In a study that colleagues and I conducted, a company with a call center did a phased rollout of a large language model—generative AI—that gave suggestions to some of the workers [as they responded to callers], but not to others. So we got a kind of controlled experiment. The people who had access to the technology were dramatically more productive. It was about 14% on average, but the least experienced workers were about 35% more productive within just a couple of months: a big, big change.... Customer satisfaction dramatically improved.... The employees seemed happy. They were less likely to quit—much less turnover."
--A smart framework for thinking about investments in generative AI, from my longtime friend and colleague Tim Andrews. He says we are still in the Institutional phase of his i3 model, characterized "by the need to go to large institutions for access." Next will come the Individual phase, when "the technology is affordable but not cheap and generally requires some expertise to install and maintain it." Finally, there will be the Invisible stage, where "the technology disappears from view and becomes embedded in just about everything possible....Interest in the technology itself wanes, except when broken or missing.... A lost internet connection is painful because of the interruption of a video call or streaming movie, not because of the underlying internet technology."
--A mind-boggling stat on the homeowners' insurance crisis, suggesting that we have a long way to go before it's resolved:
"First Street estimates that 39 million U.S. homes are insured at artificially suppressed prices compared with the risk they actually face."
--Five myths about customer loyalty, from Jon Picoult, one of my go-to's on the topic, including:
"Myth #1: Satisfied customers are loyal customers.
"Satisfied customers defect all the time. In a widely cited customer experience study, Gartner found that 20% of customers who said they were satisfied with a particular company also said that they planned to shift their business to another provider. This is why customer satisfaction is really a one-way ticket to the business graveyard. To cultivate true, long-term loyalty, businesses must do more than just satisfy customers – they need to impress them, thereby cultivating the repurchase and referral behavior that is the lifeblood of any thriving company."
Cheers,
Paul