This week's 5 pieces of robot & AI news
Work on combining robots with LLMs; a chatbot hijacks company policy; and the NBA uses AI for ... something.
1 Body with no words, meet words with no body.
So I have a piece in the March issue of Scientific American about various attempts to connect Large Language Models (those word-slinging AIs like ChatGPT, Bard, Gemini, Ernie etc) with robots.
From a distance, it seems like a natural combination. Robots have no clue about how the world works or what people are like, but they have bodies that move in real spaces, and take actions. Meanwhile, the Large Language Models have immense amounts of data about how the world works and how people are. What they “understand” about the world is unclear, but the word-hoard at their disposal means that a LLM can tell the robot what a fish is and how to cook it, for instance.
Up close, it’s a subtle challenge to combine those limitless words with the still quite limited repertoire of robots. But the potential is great for at least two areas: (1) Making it easy for ordinary mooks like you and me to communicate with robots (so that when you say “put everything in the closet,” the robot knows not to put the dog and the couch and the soup in there; and (2) Giving the robot an “understanding” of us humans and our world.
One example of (1) that I wrote about is using an LLM as a bridge between ordinary human talk and the precise coding a robot needs. Ie, you say something human, like “I want all these boxes sorted,” and the LLM writes code so the robot will do that. Magazines have long lead times, so some of the latest work on this appeared after the article had “closed.”
For example, this paper (and the cool videos) which came out earlier this month, describe a method for improving LLMs ability to recall what they have learned from humans’ corrections. That allows the robot to map the human’s reactions almost as if it were a mysterious new land the robot was exploring for the first time. Which is great, because robots are good at that kind of exploration.
2 Annals of LLM Hallucination
Chatbots make stuff up sometimes. In this case, the chatbot invented a refund policy while messaging with an airline customer. Too bad, ruled a Canadian court. The airline has to do what the bot promised it would.
3 Department of Solutions Looking for a Problem
“NB-AI” is the NBA’s new gimmick, which — as this carefully curated brief demo shows — can turn an ordinary basketball game, as you watch, into a Spiderman cartoon. This is AI as performance: a show that says, hey, we’re not being left behind! Not the same thing as AI that is actually useful. I know one 13-year-old who thinks this is great, but most people would rather just watch the skilled humans play the game, no?
4 Good robot video
Stretch is a cool open-source robot but what I really like about this video is its honesty. The screen spells out what a lot of demos usually hide — that the video is speeded up to 4x or 8x real pace, and that the robot is being operated by an off-screen human. Consumer tip: When watching robot video, you should always be told whether the machine is autonomous (or operated Wizard-of-Oz style by someone behind the curtain). And whether the machine has been sped up from its customary snail’s pace in order to make it look more brisk and efficient.
5 A rare self-promotional paragraph
Jonathan Hurst, Chief Robot Officer of Agility Robotics, doesn’t care for deceptive robot videos either. We discussed it on the Robots for the Rest of Us Podcast. Check it out! Many more interesting conversations are coming.
This Week’s Literary Note:
One day early in the last century the poet and novelist Rudyard Kipling stopped by the country home of another literary giant, the novelist Henry James. Kipling was eager to talk new technology -- specifically the expensive "motor car" that had brought him and his wife to see James.
The automobile, able to cut travel times and let people sense speed and power, was a thing, Kipling said, to make an Englishman think.
James' reaction was a bit catty. As he later told the writer Ford Madox Ford, Kipling went on about how quickly the motor car would get him and James and company from lunch here to tea there.
James then told Ford (who captured the Master’s long, winding, clause-filled, paranthetical style):
“And we were all indulging in — what is it? — delightful anticipations and dilating on the agreeableness of rapid — but not for fear of the police and consideration for one's personal safety too rapid — speed over country roads and all, if I may use the expression, was gas and gingerbread when ...There is a loud knocking at the door and — avec des yeux éffarés —” and here Mr. James really did make his prominent and noticeable eye almost stick out of his head . . .”in rushes the chauffeur.... And in short the chauffeur has omitted to lubricate the wheels of the one thousand two hundred guinea motor car with the result that its axles have become one piece of molten metal. . . . The consequence is that its master and mistress will return to Burwash which should be pronounced Burridge, by train, and the magnificent one thousand two hundred guinea motor car will not devotedly return here at noon and will not in time for lunch convey me and my nephew Billiam to Burwash and will not return here in time for me to give tea to my friend Lady Maud Warrender who is honouring that humble meal with her presence tomorrow beneath my roof or if the weather is fine in the garden . . .'
“Which,” concluded the Master after subdued 'ho, ho, ho's' of merriment, “is calculated to make Mr. Kipling think.”
Why do I bring up this century-old example of what I call maschinenschadenfreude -- pleasure in the failure of a machine? Because I'm reminded of it often when I read stories about robot flops. Kipling's early, expensive car broke and James was amused. But a few decades later, cars were everywhere and James’ world was gone. Kipling and the motor car had the last, as it were, ho ho ho of merriment.