AI's Appetite for Energy Is Going to Be Huge
Also: A humanoid-robot milestone, new robot-learning techniques and a reminder to beware "enshittification"
1. AI’s Energy Demands
As you've probably noticed, many expect that we'll soon be living with AI baked into everything, from grammar checkers to toasters to shoes (to say nothing of all the robots). That's going to require a lot of electricity — so much, says OpenAI's Sam Altman, that this supposed AI utopia can't be realized until humanity makes an epochal breakthrough on energy. A few more solar panels and wind farms won't cut it. We'll need something more on the order of practical nuclear fusion. "There's no way to get there without a breakthrough," he said last week.
2. Humanoid Robot Trial at BMW
Robots have been everywhere in car manufacturing for 50 years, but there's always been plenty of work left for humans. Our hands are much better at fine-motor tasks (for instance, snaking a bunch of electrical cables through an awkwardly shaped little aperture). And our bodies fit naturally into spaces designed for us (like a warehouse whose shelves have to be stacked with boxes).
So it's significant that BMW just signed a contract to test a human-sized, human-shaped robot at a plant in Spartanburg, South Carolina. The robots will be supplied by Figure, one of a number of humanoid-robot companies that have begun manufacturing real humanoids for real work in the past couple of years. Another, Agility, has a couple of contracts to test its humanoid in warehouse work.
Nothing is inevitable
Take note, then — but don't fall into the trap of thinking these kind of robots are inevitable.
As Steve Crowe reports here, this is a preliminary test, with one robot. It's going to move boxes around until BMW works out some other possible uses. Likely those will involve chores where its two five-fingered hands can come into play. Maybe it will be great at a whole bunch of jobs that people now do. But maybe it won't. Big companies nowadays test robots for all kinds of tasks, and sometimes the robots flunk out.
One thing, though, seems safe to say: After 100 years of science fiction and a decade or so of prototypes and lab work, 2024 looks likely to be the year when real human-shaped robots start appearing in places where us civilians roam.
How Innovation Can Make Things Worse
"Enshittification" is the American Dialect Society's Word of the Year for 2023. The term is Cory Doctorow's coinage to describe a situation in which a product or service is made worse over time, often for everyone involved (producers and consumers alike) — but to the profit of people who own the company.
One example Doctorow cites is the Chamberlain MyQ brand of garage-door opener, which many buyers had integrated into "smart home" software, like the Home Assistant open-source application.
The company disabled its API (the software bridge that lets Home Assistant and other apps plug into the door opener), forcing owners of its product to use its own app, which, he writes, is riddled with ads.
Why am I mentioning this here? Because digital tech offers tools that make enshittification possible — after all, without networking, a company wouldn't be able to disable a feature on products people had already bought.
Moreover, AI will make further enshittification possible. For instance, some day soon a CEO will fire almost everyone in the Media Relations department and replace them with a high school student who runs ChatGPT queries.
This will be worse for the employees (their jobs are gone) but also worse for consumers of Media Department's products, which won't be as good as what humans can produce. (ChatGPT is great at making stuff up out of whole cloth, but it falters when it has to collect and compare facts and figures, as one must to write a corporate press release.)
As AI improves and gets deployed in more and more places, we should beware of it being used for such purposes. Avoiding enshittification should be a goal of AI regulation and antitrust efforts, as Doctorow explains here.
4. Making Robots More Capable — Google DeepMind
Google DeepMind announced a suite of enhancements that make robots able to understand people's desires, and to explore the world in ways that help them perform their missions. To give the robots information they need about the world, they use large "foundation models" — AIs like ChatGPT, which can produce human-like sentences, and DALL-E, which can invent images — to give the robots a feel for the world, and for what people might want. Explanations here.
5. Making Robots More Capable — Stanford
The Google project is about leveraging lots of pictures and words to tell robots what to do. Another way is to simply do the thing you want done, then have the robot do what you did. That is what is happening in this project from Stanford: A human acts a little like a puppeteer, and the robot then takes over, doing on its own what it once did when the human was driving. That's the way the robot in this demo has gotten the hang of cooking a shrimp and rinsing out the pan, among other tasks.