AI Developments... and Predictions

 I've been watching the AI developments... and the bloggers writing about the AI "foom".

• Scott Alexander on the feeling of being on the verge of AI Apocalypse.
• Scott Alexander on the new Janus-faced AI simulators, a weird new breed of alien intelligence we didn't expect.
• Erik Hoel on why we shouldn't be cavalier about the potential impact of super-intelligent AI.  (Item 4 of 9 on the page).

But in my opinion, the new AI chatbots may replace Google's current search, but they will stop far short of replacing anything but the most mundane human jobs. 

Its so hilarious that people are worried about artificial intelligence when we haven't even made a robot that can flip hamburgers yet!  Wake up and smell the grease:  the real world is complicated and messy.  We can't even replace the jobs that everyone, even the people currently working them, want replaced.  I don't think we're going to be replacing professionals anytime soon.

Just look at driverless cars.  Driving is a great example of something you'd think computers would be good at.... if you'd never driven a car in the real world!  Computers are great at driving cars in video games, but the real world is messy.  That's the Real World problem.

The second problem is even bigger: the messiness of the human social world!   Even if driverless cars were better than human drivers, that's not good enough.  We expect computers to be perfect.  If a company sells me a car and I wreck it, I'm responsible.  But if the company sells me a driverless car that crashes, the company is responsible.    

For every task that matters we want a human to be responsible.  Even if they solve the Real World problem with bigger and better AIs, the Responsibility problem will always prevent human replacement.  Imagine if my boss could replace me with a computer: he would then be 100% liable for any mistakes the computer made!  Is that the kind of liability any Pointy Haired Boss wants??  As long as he has real humans working for him, there's always someone to share the responsibility.  

Example: imagine something that's even simpler than driving a car: flying a passenger airplane!  I'm sure autopilot could fly an airplane as well as a human.  But would anyone want to fly on an airplane with no human pilot?  We'll always have the human pilots, even if the autopilot does more and more of the routine work.

Humans are social creatures.  The humans who think AI will rule the world must be living inside a computer world.  Its interesting that, as more and more of our world is computerized, we hear more and more from the people living inside the computer world.  But its not the real world!

In These Cheatgrass-Infested Hills

Matthew Miller, The Nature Conservancy writer/editor, has written a beautiful article about coming to grips with beautiful yet non-native landscapes.

In my opinion, there are too many conservationists lost in a dream of pure "native" nature, unable to see the flawed-but-still-beautiful world around them.  I empathize with his struggle to learn to love degraded places, even the hard-to-love places that are infested with cheatgrass.  

Trying to widen our circle of appreciation to include even nasty invasives helps us appreciate the natural in the unnatural: the native pollinators that use invasive wildflowers, the native birds that nest in invasive trees. 

To quote Princess Mononoke, our task now is "to see with eyes unclouded by hate."

Climate Prediction Skill

The US Climate Prediction Center issues forecasts beyond the normal National Weather Service's 10-14 day window.  They provide weekly and monthly forecasts out to 3 months.  Given the timeframe and the fact that their forecasts cover the entire contintental US, its not surprising that the forecasts are often wrong.  But how wrong?  And is their skill improving over time?

I analyzed their 3 month temperature and precipitation forecast skill using data provided on their "Gridded Seasonal Verifications" webpage.  

Note that skill is measured on a scale from -50 to 100, where -50 would be a forecast that was exactly wrong in every area, 0 would be a prediction that did no better than chance, and 100 is a prediction that was exactly right in every area.  

They provide data starting in 1995.  Since that time in the mid 1990's, linear trendlines show that their forecast skill has slightly improved for both Temperature and Precipitation.  Precipitation skill started out lower, but has almost doubled (from 10 to 20) while Temperature skill started higher but has not increased as much (from 22 to 28).

However, the last 10 years have not been as successful:

Since 2012, neither Precipitation nor Temperature skill have increased.  In fact, mean temperature forecast skill has decreased markedly since 2018.  Before that, Temperature skill had been doing quite well in the period 2014-2018.  It is not clear what changed in 2018.  A similar transition may be happening with Precipitation, where the period 2019-2022 saw consistently good predictions, but since the beginning of 2023 the forecast skill has fallen off a cliff.

With increased use of machine learning, it seems likely that long-term weather forecast skill should increase.  However, complex chaotic weather patterns are most impactful to climate predictions in the 1-3 month time frame, so this area of weather/climate prediction may continue to have lower than hoped for success.