WEBVTT 00:00:00.000 --> 00:00:09.160 This is Essays on AI. Your narrator is an AI voice, though not the general kind, by 00:00:09.160 --> 00:00:14.400 any of the definitions you're about to hear. It's reading an essay by Tabrez Syed called 00:00:14.400 --> 00:00:20.719 Waiting for Godot. The whole plot of Samuel Beckett's Waiting for Godot fits in a sentence. 00:00:20.719 --> 00:00:26.280 Two men stand by a tree on a country road, waiting for someone called Godot. They aren't 00:00:26.280 --> 00:00:32.299 sure when he's due. He said Saturday, one man called Vladimir offers, and then, after 00:00:32.299 --> 00:00:38.500 a pause, I think. The other man, Estragon, isn't convinced it is Saturday. It could 00:00:38.500 --> 00:00:45.720 be Sunday, or Monday, or Friday. Every so often, one of them gets restless. Let's go. 00:00:45.720 --> 00:00:51.599 We can't. Why not? We're waiting for Godot. At the end of the day, a boy arrives with 00:00:51.599 --> 00:00:57.759 a message. Mr. Godot won't come this evening, but surely tomorrow. The next evening, the 00:00:57.759 --> 00:01:02.799 two men are back at the tree. The boy is back with the same message, and he insists he has 00:01:02.799 --> 00:01:08.720 never been there before. Godot never comes. Very little happens at all. The one thing 00:01:08.720 --> 00:01:14.400 that changes is the tree. On the first night, it is bare. On the second, it has four or 00:01:14.400 --> 00:01:20.080 five leaves. I've been thinking about that tree, because AI has one too, with a large 00:01:20.080 --> 00:01:24.400 crowd standing around it, waiting for AGI. The strangest thing about the man they're 00:01:24.400 --> 00:01:29.400 waiting for is that they don't know him. When a stranger comes down the road, Estragon 00:01:29.400 --> 00:01:35.040 leans over and asks whether that's him. It isn't. Later, he admits he wouldn't recognize 00:01:35.040 --> 00:01:40.599 Godot if he saw him. And when Vladimir finally asks the boy what Mr. Godot does with his 00:01:40.599 --> 00:01:47.919 days, the answer is, he does nothing, sir. AGI has a similar problem. OpenAI's charter 00:01:47.919 --> 00:01:53.400 describes it as highly autonomous systems that outperform humans at most economically 00:01:53.400 --> 00:01:59.959 valuable work. When OpenAI and Microsoft needed a version they could put in a contract, they 00:01:59.959 --> 00:02:06.160 reportedly settled on a system that can earn $100 billion in profits. NVIDIA's Jensen 00:02:06.160 --> 00:02:12.119 Huang sets the bar at an AI that can build a billion-dollar company. And in September 00:02:12.119 --> 00:02:19.759 2026, he declared on X that OpenAI's newest model was AGI. Since the definitions don't 00:02:19.759 --> 00:02:26.360 agree, Microsoft and OpenAI agreed in 2025 that any claim of AGI would be confirmed by 00:02:26.360 --> 00:02:33.279 an independent panel of experts, so that someone would be able to recognize him when he arrives. 00:02:33.279 --> 00:02:40.199 Sam Altman, whose company was founded to build him, has said AGI is not a super useful term. 00:02:40.199 --> 00:02:46.000 Other definitions reach for a person. Elon Musk has defined AGI as AI smarter than the 00:02:46.000 --> 00:02:52.399 smartest human, which raises the question of who that is. Einstein is the usual answer, 00:02:52.399 --> 00:02:58.360 and in 1912, when he needed the geometry for general relativity, he went to his old classmate 00:02:58.360 --> 00:03:04.839 Marcel Grossmann. Grossmann, you have to help me, or else I'll go crazy. Yann LeCun, then 00:03:04.839 --> 00:03:10.720 met as chief AI scientist, took the point further in December 2025. There is no such 00:03:10.720 --> 00:03:15.880 thing as general intelligence, he said. We think of ourselves as being general, but it's 00:03:15.880 --> 00:03:22.800 simply an illusion. Demis Hassabis, who runs Google DeepMind, replied on X that LeCun was 00:03:22.800 --> 00:03:29.699 just plain incorrect. A few hours later, LeCun wrote back, I think the disagreement is largely 00:03:29.699 --> 00:03:37.699 one of vocabulary. Wittgenstein thought philosophical problems arise when language goes on holiday. 00:03:37.699 --> 00:03:41.820 Here were two of the most accomplished people in the field, arguing in public about what 00:03:41.820 --> 00:03:48.179 a word means. Not knowing who Godot is doesn't stop Vladimir and Estragon from arguing about 00:03:48.179 --> 00:03:54.020 when he'll come. The boy doesn't settle it. Each evening he brings the same message, 00:03:54.020 --> 00:03:59.380 and each time, he says he has never been there before. The argument about AGI has the same 00:03:59.380 --> 00:04:06.979 shape. In October 2025, Andrej Karpathy went on Dwarkesh Patel's podcast and said AGI 00:04:06.979 --> 00:04:13.779 was still a decade away. Musk said in April 2024 that it would come probably next year, 00:04:13.779 --> 00:04:20.779 within two years. The deadline passed quietly, and in January 2026, at Davos, he brought 00:04:20.779 --> 00:04:27.140 the message again. By the end of this year, or no later than next year. And in September 00:04:27.140 --> 00:04:33.339 2026, Nathan Lambert, who writes a closely-followed newsletter on how these models get built, 00:04:33.339 --> 00:04:38.820 laid out where he stands by quoting the researcher Richard Ngo, we won't have superintelligence 00:04:38.820 --> 00:04:43.660 within the next eight years, but things will still be moving so fast that it'll feel 00:04:43.660 --> 00:04:50.140 like the people who argued for short timelines were right. Godot won't come, in other words, 00:04:50.140 --> 00:04:56.859 but it will feel as if he did. Everyone seems to have a date. Nobody has the same description. 00:04:56.859 --> 00:05:02.739 Early in the play, Estragon has a different worry. You're sure it was here? What, that 00:05:02.739 --> 00:05:08.380 we were to wait? He said by the tree. Some of the most respected people in AI share the 00:05:08.380 --> 00:05:15.279 worry about the road the industry is on. Ilya Sutskever, who co-founded OpenAI, said in 00:05:15.279 --> 00:05:22.380 November 2025 that the age of scaling is over, and that today's models generalize dramatically 00:05:22.380 --> 00:05:28.220 worse than people. Richard Sutton, who won the Turing Award for his work on reinforcement 00:05:28.220 --> 00:05:34.700 learning, calls large language models a dead end, and Gary Marcus has argued for years 00:05:34.700 --> 00:05:41.179 that they are not the royal road to AGI. In a 2025 survey by the Association for the Advancement 00:05:41.179 --> 00:05:50.059 of Artificial Intelligence, 76% of 475 researchers said scaling up today's approaches was unlikely 00:05:50.059 --> 00:05:56.179 to get there. None of them has left the tree. With this much money gathered around it, nobody 00:05:56.179 --> 00:06:04.019 says stop waiting. They say they know a better place to wait. Sutskever left OpenAI in 2024 00:06:04.019 --> 00:06:10.980 to start Safe Superintelligence, which was valued at $32 billion within a year. LeCun 00:06:11.100 --> 00:06:20.380 left Meta in November 2025, and in March 2026, his new company, AMI Labs, raised $1.03 billion 00:06:20.380 --> 00:06:26.619 to build a different kind of model, NVIDIA, whose chief executive says Godot is already 00:06:26.619 --> 00:06:34.700 here invested in both, including $5 billion in Sutskever's company in July 2026. Ask 00:06:34.700 --> 00:06:40.459 Sutskever when, and he still has an answer. A system that learns as well as a person in 00:06:40.459 --> 00:06:47.980 like 5 to 20 years. A few go further. Arvind Narayanan and Sayash Kapoor at Princeton argue 00:06:47.980 --> 00:06:53.660 that AI is a normal technology, powerful in the way electricity and the internet were 00:06:53.660 --> 00:06:59.500 powerful, rather than a new kind of mind. We do not think there is a useful sense of 00:06:59.500 --> 00:07:05.059 the term intelligence in which AI is more intelligent than people acting with the help 00:07:05.059 --> 00:07:11.980 of AI, they write. In their telling, there may be no Godot at all, only the tree. When 00:07:11.980 --> 00:07:16.980 the curtain goes up on the second act, the tree that was bare the night before has four 00:07:16.980 --> 00:07:23.260 or five leaves. Vladimir notices, and then the two of them go back to waiting. Whether 00:07:23.260 --> 00:07:29.940 or not Godot comes down this road, the leaves are real. The four largest American tech companies 00:07:29.940 --> 00:07:37.380 are on track to spend close to $700 billion on AI infrastructure in 2026, up from about 00:07:37.380 --> 00:07:43.380 $400 billion in 2025. And the models coming out of it now work through tasks that take 00:07:43.380 --> 00:07:50.279 a person hours. Whatever the labs are doing, it isn't standing still. The ones standing 00:07:50.279 --> 00:07:57.220 still might be the rest of us. In a Harvard Business Review survey of about 1,000 executives 00:07:57.220 --> 00:08:04.980 late in 2025, 6 in 10 said they had cut staff in anticipation of what AI would be able to 00:08:04.980 --> 00:08:13.040 do, while only 2% had made large cuts because AI had actually shown it could do the work. 00:08:13.040 --> 00:08:18.380 Most of them were acting on the name, not the leaves. The people just starting out feel 00:08:18.380 --> 00:08:25.140 it first. Researchers at Stanford found that employment for 22 to 25-year-olds in the jobs 00:08:25.140 --> 00:08:31.459 most exposed to AI is now 19% below where it would have been had it kept pace with their 00:08:31.459 --> 00:08:38.140 peers. Enrollment in computer science, long the safe bet on campus, fell more than 8% 00:08:38.140 --> 00:08:45.260 in the spring of 2026. The play ends the way its first act did. The boy has come and gone. 00:08:45.260 --> 00:08:52.340 Godot is still on his way. And the two men agree, at last, to leave. Well, shall we go? 00:08:52.340 --> 00:08:58.299 Yes, let's go. And then Beckett's stage direction. They do not move. We're standing 00:08:58.299 --> 00:09:03.900 at the same tree, arguing about a Saturday nobody wrote down, for someone we wouldn't 00:09:03.900 --> 00:09:09.820 recognize, at a spot some of us suspect is the wrong one. The labs will keep building 00:09:09.820 --> 00:09:14.619 either way, and the leaves will keep coming, whatever we decide to call them. 00:09:14.619 --> 00:09:21.200 That's the essay. You can find the written version, with links to every source, at mandalivia.com, 00:09:21.200 --> 00:09:25.599 and if you'd like future essays read to you as they're published, subscribe wherever 00:09:25.599 --> 00:09:29.599 you listen. Essays on AI is a Mandalivia Production.