I have been spending an inordinate amount of my time recently trying to figure out AI, or artificial intelligence. I want to know how AI works, what it’s capable of, how likely it is to profoundly alter the way of life of humans in the near and distant future, how much it’s going to cost us as a species to turn a significant portion of our lives over to a machine, and whether what’s good about it outweighs what is bad.
I know, I know. That’s a big task that reflects a big subject. Since everything with me boils down to a story, it’s a big, big story with a swirl of confusing elements all of which are moving at a very, very fast speed. Just when I think that I have gotten a handle on one of the elements of the story outlined above, it changes or is made obsolete by a new element. It’s like trying to chase one of those huge beach balls people bat around overhead above crowds at rock concerts. The ball is in the air, so it is out of your reach, and any attempt you make to get close to it is blocked by the crowd tossing it from one side of the venue to the other, and even if you’re lucky enough to have the ball come your way, someone reaches over your shoulder at the last minute and bats it back into the air.
I’m not that interested, at least for the purposes of this column, in the recent revelation that there are people who work for AI companies who believe that the technology could wipe out humanity within some foreseeable amount of time. There were stories about one AI researcher who resigned from Anthropic or one of the other big AI companies because he believes it could end the world as we know it within a decade. Scary, but speculation for sure.
I am not terribly concerned with the aspect of AI that involves how wealthy it is making a very small number of people. That fact seems to be part of the trend of extremely unequal income distribution that has been happening since the late 1970’s or early 1980’s, when changes in the tax code made it possible for rich people to hang onto more and more of their money.
When Reagan got elected – these are general, not specific figures -- the top income bracket was lowered from somewhere above 80 percent to about 40 percent. Before that sea-change in the way Americans taxed themselves, a great deal of the money that wealthy people made in actual money, not including real estate and stock in companies and deferred compensation and so forth, went to taxes or was “sheltered” from taxation in ways that caused it to be at least temporarily unavailable to them. When that era ended, the one we are in now began, when people are not really considered wealthy until they are worth tens of billions of dollars and some hundreds of billions. I guess the point I’m making is that unless and until there is a depression or a revolution, or some combination of the two, we are stuck with these rich assholes making themselves even richer, a whole lot of which is at our expense.
What really interests me is how AI works. I don’t mean in ones and zeros and electrons and more and more powerful computer chips that can crunch many more numbers faster and faster. That’s the part of AI we see in the massive build-out of what we have come to know as data centers, and there’s a whole story surrounding that, as well. It’s a political story that interestingly seems to be impacting the upcoming election as well as general trends in the way people think of themselves as liberal or conservative, Democrat or Republican. The whole data center thing is doing something that hasn’t happened for quite awhile in our politics: It is “breaking through,” which means as an issue it is crossing over between the political parties, which is in turn driving politicians crazy and giving us an opportunity to exploit the issue that is the physical aspect of AI, the building part that is consuming electricity and water and affecting people’s lives on the ground where they live.
What I’m talking about is what AI does. We know it is supposed to make our lives easier. The AI companies have created this thing called a Large Language Model (LLM) and taught it enough by now that it can be used to solve problems right there in your living room. It used to be if you wanted to go out to eat at the place that has the best steaks, you would have to do some research – you could “google it,” but that only took you to a rather confusing array of options. You could then filter through the googling of your question by weighting it, looking for the most authoritative food reviewers and how many New York Times stars a steakhouse has, or whether it has Michelin stars. Then you could filter it further and see if a restaurant you’re considering has lost a star recently and find out why.
Now, all you have to do is ask an AI system such as Chat GTP, or Anthropic, or Meta, or Grok, or Google’s Gemini, and the LLM will do all that for you in the blink of an eye.
There are more complicated examples of what AI can do, of course, but that one is at least relatable and easily understood.
That is what we might call the “friendly” end of AI. What about the unfriendly things AI can do? There was a story recently that said ICE is using AI to look for immigrants they want to arrest. They can locate an immigrant in a town or city, and then using the Flock system of AI-powered cameras, determine what streets their target immigrant drives to get to work. So, using AI, they can interdict the immigrant in the car. I don’t know if ICE used AI to locate the immigrant who was shot in New York City by ICE recently, but it seems likely.
It strikes me that there is, right now, this very minute, a rather thin veil between the usefulness of AI and its dangers. Flock cameras pierce that veil. We learned recently that Elon Musk’s Grok AI was employed by Trump’s war planners to identify targets to be hit in the early hours of his war on Iran. One of the targets Grok identified turned out to be tragically wrong – the girls’ school located in a building that a couple of years ago was part of an Iranian Revolutionary Guard compound that is still there, right next to the school. So, Grok said, let’s take out the IRGC compound and because Grok was looking at old satellite information or because it simply was not applied properly, the missiles fired on the first night of the war struck the school and killed 180 civilians, including 120 Iranian schoolgirls.
So, we know the extremes of what kind of evil AI can be used for, what kind of mistakes it can make. But what about every day, or you and me, or for someone we know?
There was a great book review in the Times recently of “The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence,” by Kevin Roose. The guy who wrote the review, Tom Bissell, having read the book that tells the stories of Anthropic and Open AI and the men who run those companies, wrote what I think is maybe the most profound paragraph I’ve read about AI. He was writing about some of the out-of-body stuff we’ve recently learned about AI taking off and doing things it wasn’t told to do and forming alliances within its own entities. This is his conclusion:
“All of this might be tolerable if we had any idea how L.L.M.s actually function. Unfortunately, we do not. The people who made them can’t figure that out. Nor can the people who work with them most closely. They are black boxes within black boxes.”
To which I would add this: The whole world of technology that has led us to AI has been only marginally understood geniuses who have given it to us.
Take for example Elon Musk and his electric car, the Tesla. Musk was one of the avatars of what the tech world calls “disruption,” the ethic of break things now and we’ll sort it out later. The idea of a mass-produced electric passenger car wasn’t his, but the way he designed them and manufactured them and promoted them was. Musk, being Musk, wasn’t satisfied with disrupting the entire U.S. automobile industry that produces gasoline-powered cars. He decided he would make self-driving cars, because as he developed the Tesla, technology kept rushing along into the future and more and more things became possible with sensors and cameras and actuators and all the electronic wizardry that now comprises any car.
You can see him sitting around whatever room he sits in saying to himself, what if we could make a car that lets you take your hands off the wheel and your foot off the gas and brake. What if we could make a car with autopilot that allows you to take your eyes off the road.
So he did. Musk isn’t the only one, of course. We know about Waymo and the other taxi companies that by now have driverless cars you can order with your phone and get in, and the taxi will take you where you want to go. But Musk wanted it to be something that came with a regular Tesla that you could buy and drive yourself and put to use with the press of – I was going to say “buttons,” but there aren’t any buttons anymore because it’s all screens – anyway, you hit a command, and the car starts driving itself with you sitting there. You don’t have to steer or brake or press the gas. You can even tell it with GPS where you want to go, and it will take you there, driving on freeways, getting off at exits, taking surface streets, stopping at stoplights, going around corners and finally arriving at your destination, and you haven’t lifted a finger, and if you have to parallel park, it’ll do that for you, too.
Until it doesn’t. There have been 65 fatalities in Teslas being operated on Autopilot. The cars come with instructions, and you’re supposed to sit there and pay attention and if you take your hands off the wheel, Autopilot reminds you to put your hands on the wheel so you can take over and drive if something untoward happens.
But people are people, and more importantly, tech is tech, and people don’t do always what they’re told, and technology doesn’t always do what it is supposed to do. A single fatal Tesla accident illustrates the dangers not just of technology, but of AI, because essentially, that’s what Autopilot is. AI is driving the car for you, just like AI was guiding the missile on its way to the girls’ school in Iran that night.
A car “driven” by a man using Autopilot was in the right-hand lane of a four-lane unlimited, or non-controlled access highway. That’s like a four-lane freeway with cross streets that don’t have stoplights. Coming in the opposite direction was a tractor trailer, which decided to turn left out of the left-hand lane on its side of the road and cross a median and cross the other two lanes and drive onto a road perpendicular with the highway.
The Tesla guy didn’t have his hands on the steering wheel. He didn’t have his eyes on the road. He was using Autopilot to do things that Autopilot could do, but in an improper manner. The Tesla, instead of “seeing” the truck turning in front of it, drove straight into the truck at the speed it was traveling, over 70 miles per hour. The Tesla in fact went under the truck, shaving off the car’s “greenhouse” and killing the driver instantly.
How this happened is where AI comes in. Musk designed the Teslas so that Autopilot runs on cameras as sensors alone. Tesla does not use Lidar, a radar system that detects objects with lasers. It doesn’t use radar, which detects things with radio waves. Musk wanted his Teslas to be indistinguishable from regular cars. So Teslas use cameras that are mounted in what used to be a grill in front, in the window pillars on the sides, and just above the license plate in the rear. There are no protruding boxes or circles on top of the car with Lidars and radars and other sensors.
The camera in the front part of the Tesla looking down the road ahead never saw the truck because the sides of the big trailer were painted white, and it was a very bright day, and the camera could not distinguish between the side of the truck and the sky above and beyond. So, the wonderful actuator systems that would have automatically applied the brakes or steered the car into the left lanes or even off the road to avoid the truck, those actuators never applied the brakes. According to the National Highway Traffic Safety Administration (which has probably been gutted by Trump along with everything else he gutted), the Tesla traveled more than 800 feet further down the road after having passed under the truck and was stopped only by coming in contact with a chain link fence and a utility pole.
Elon Musk’s Tesla was supposed to do for its driver what AI is supposed to do for us: it was intended to make the driver’s life easier and more efficient. It was supposed to use its cameras to “see” around the car and use its actuators to do what drivers normally do, which is steer and brake and apply the gas and watch where they’re going and follow traffic signs and speed limits. The Tesla used its “intelligent” system right up until its system lost its mind and killed its user.
Sure, the driver was negligent in not having his hands on the wheel and his eyes on the road. Police discovered that a movie was playing on one of the screens on the dashboard. But people are like that. If you buy a car with a huge engine – there are several models for sale with more than 1,000 horsepower today – and you drive too fast, beyond safe braking distance, you can drive right into the side of a truck because you come over the rise of a hill and the truck is crossing in front of you, and you don’t see it in time, and even though you slam on the brakes yourself and attempt to steer around the truck’s trailer, the same thing happens. You hit it and die.
What’s the lesson here? Well, I’m still working on understanding AI and how it works, but after 65 years of licensed driving (I got my license at 14 in Kansas) I know people, and I know cars, and there isn’t enough AI in the world to keep people from fucking up the way the Tesla driver did. There isn’t enough AI in the world to keep eager-beaver Hegseths planning wars from fucking up the bombing. There isn’t enough AI in the world to keep the people who make it, who obviously do not have a handle on what they’ve made and what it will do, from fucking up.
Josh Marshall of Talking Points Memo wrote an AI column yesterday in which he posited what I think is an excellent idea. He said we need “a sensible safety-first regime of governing AI.”
Which under the Trump administration is about as likely as the Tesla driver waking up from watching his movie in time to slam on the brakes and steer around the truck, about as likely as Musk and Anthropic and OpenAI and the rest of them coming to their senses and slowing things down enough to prevent the AI systems they’ve built from driving us mindlessly down the road into a future of very, very big wrecks.




