When writing about technology that is rapidly evolving, there are two obvious failure modes. Looks like a failure mode This is the graphFamously created by Auke Hoekstra In 2017 And updated every year, shows how the International Energy Agency has repeatedly underestimated future growth in solar power:
Solar installations have grown by about 25 percent annually, but for more than a decade, the IEA has underestimated them, often predicting that they will level off and stabilize or even decline going forward. This is no small flaw; Vastly underestimated advances in solar capacity dramatically change the picture of climate change mitigation and energy production.
Early in Covid-19 you saw some examples of the same failure mode. It is very easy to underestimate exponential growth, especially in its early stages. “Why should we fear something that hasn’t killed people here in this country?” An epidemiologist argued the LA Times in late January 2020. The flu was a “much bigger threat,” the Washington Post wrote a day later, compared to how many people there were How many people are currently infected with covid-19 are infected with the flu.
I hardly want to point out that this analysis misses the point badly. Of course, there weren’t many people with Covid-19 in the US in late January. The concern was that, given the way viruses work, that number was going to rise exponentially – and indeed it did.
So it’s a failure mode: Repeatedly ignoring and emphasizing an indicator will equalize any minute now, resulting in the dramatic loss of one of the most important technological developments of the century, or causing people to not think about a pandemic that shuts down the whole thing. Just a few weeks later the world.
Other failure modes, of course, are:
Some things, like the initial spread of Covid-19 or the increase in solar capacity to date, prove to be exponential curves and are best understood by thinking about their doubling time. But most of the cases are not.
Most of the time, like a child, you’re not looking at the early stages of exponential growth but simply at … normal growth, which can’t be pushed too far without ridiculous (and wrong) results. And even if you’re looking at an exponential curve, at some point it will go out of level.
With Covid-19, it was easy enough to assume that, at worst, it evened out when the entire population was exposed (and in practice, it was usually much less than that, as people changed their behavior in response to being overwhelmed). increased local hospital and morbidity rates). At what point will the solar power level end? This is hardly an easy question to answer, but the IEA seems to have done a pretty poor job of answering it; They would have been better off just drawing a straight trend line.
There is no substitute for hard work
I think about this a lot in the context of AI, where I see people posting dualities of these two concepts. as far as Making AI systems bigger has made them better Covering a wide range of tasks from coding to drawing. Microsoft and Google are betting that trend will continue and Spending a lot of money On the next generation of Frontier models. Many skeptics claim that, instead, Scale would be level off convenience – or already doing so.
The people who most vehemently defend the scale of returns argue that their critics are playing the IEA game — repeatedly predicting “it’s going to break even any minute” while the trend lines just go up and up. Their critics accuse them of resorting to dumb oversimplifications that current trends will continue, hardly more serious than “my baby weighs 7.5 trillion pounds.”
Who is right? I’ve come to increasingly believe that when you’re considering these questions there’s no substitute for digging deep into the weeds.
To expect solar production to grow, we Need to study How we make solar panels and understand the source of ongoing downward costs.
There were ways to predict how Covid-19 would play out Guess how contagious the virus was Extrapolate the probabilities of successful global control to and from the earliest available outbreak data.
In no case can you replace broad thinking about trend lines – it all comes down to the facts on the ground. These things are not impossible to guess. But it’s impossible to be lazy and fix these things. Detailed topics; The superficial similarities are misleading.
As for AI, building large models will quickly create AI systems that can do what humans can do – or that’s a lot of hyped-up nonsense – that can’t be answered by drawing trend lines. Nor can it be answered by mocking the trend line.
Frankly, we don’t even have a good enough measure of general reasoning ability to describe the growth of AI capabilities in terms of trend lines. Insiders at the labs that build the most advanced AI systems say that as they make the models bigger and more expensive, they see constant, big improvements in what those models can do. If you’re not an insider in the lab, it can be difficult to evaluate these claims — and I certainly find it frustrating to examine papers that tend to overhype their results, trying to figure out which results are real and real.
But there is no shortcut to doing that. Although there are some questions we can answer from first principles, this is not one of them. I hope that enjoying our batting charts back and forth doesn’t obscure how serious work needs to be done to get these questions right.
A version of this story was originally publishedFuture perfectNewsletterRegister here!
