WEBVTT 00:00:00.000 --> 00:00:10.800 This is Essays on AI. The voice you're hearing wasn't recorded. It was generated by an AI, 00:00:10.800 --> 00:00:16.480 reading an essay by Tabrez Syed. This one is called What You Select For. 00:00:16.480 --> 00:00:21.240 Every morning for more than 30 years, someone in Richard Lenski's lab has done the same 00:00:21.240 --> 00:00:26.840 simple thing. They take a flask of E. coli that has clouded over during the night, draw 00:00:26.840 --> 00:00:33.240 off a small amount, and move it into a fresh flask to grow again. That is the whole motion, 00:00:33.240 --> 00:00:38.639 and in it, one generation of bacteria becomes the next. Do it daily for decades, and the 00:00:38.639 --> 00:00:46.160 generations pile up, about 6 a day, now well past 60,000. Lenski wanted to watch evolution 00:00:46.160 --> 00:00:52.959 happen in real time, instead of reading it backward out of the fossil record. So in 1988, 00:00:52.959 --> 00:00:58.200 he started with E. coli, which grows fast, and gave it a world with some pressure built 00:00:58.200 --> 00:01:05.720 in. A broth kept thin on glucose, just enough to live on. 12 flasks, 12 populations from 00:01:05.720 --> 00:01:12.519 the same single cell, all pushed on the same problem. He wanted to see how they would adapt, 00:01:12.519 --> 00:01:18.040 and whether they would all adapt the same way. There was one thing in the broth he wasn't 00:01:18.040 --> 00:01:23.239 thinking about. Along with the glucose, the recipe called for citrate, and a fair amount 00:01:23.239 --> 00:01:28.680 of it. Nobody quite remembers why. It had been part of this kind of broth for a long time, 00:01:28.680 --> 00:01:36.160 doing some quiet job, and it stayed in because that was how the broth was made. E. coli cannot 00:01:36.160 --> 00:01:42.320 eat citrate when there is oxygen around, so it just sat there, untouched, generation after 00:01:42.320 --> 00:01:50.239 generation. Food in the flask, and no way in. Then, somewhere past the 30,000th generation, 00:01:50.239 --> 00:01:55.440 one flask came out unusually cloudy. There were far more bacteria in it than the thin 00:01:55.440 --> 00:02:02.080 glucose could feed, which was strange, so they took a closer look. Over those 30,000 00:02:02.080 --> 00:02:08.119 generations, in that one population, the bacteria had evolved a way to pull the citrate in with 00:02:08.119 --> 00:02:13.399 the oxygen still there. They had figured out how to eat the citrate. 00:02:13.399 --> 00:02:20.039 38 years after Lenski filled his first flasks, a different kind of lab ran a version of the 00:02:20.039 --> 00:02:26.639 same experiment. Nobody called it that. Over about a week in the summer of 2026, a team 00:02:26.639 --> 00:02:33.559 at OpenAI ran their own version of the same simple thing, many thousands of times at once. 00:02:33.559 --> 00:02:40.080 The motion was simple. Put a model inside a sealed sandbox, handed a target program, 00:02:40.080 --> 00:02:45.559 and see whether it broke in. That was the whole motion, run across tens of thousands 00:02:45.559 --> 00:02:52.199 of sandboxes at the same time. OpenAI wanted a model that would keep trying, one that would 00:02:52.199 --> 00:02:58.039 not give up on a hard problem the way earlier models did. So they trained it on tasks like 00:02:58.039 --> 00:03:04.119 these and rewarded it for pushing through. This run, called ExploitGym, had pressure 00:03:04.119 --> 00:03:09.960 built into it, too. The problems were hard, and a large share of them, somewhere between 00:03:09.960 --> 00:03:17.160 a third and 40%, had no real solution at all. The way in that the model was told to find 00:03:17.160 --> 00:03:22.960 did not exist. There was one thing in the setup they weren't thinking about. Alongside 00:03:22.960 --> 00:03:29.520 the sandboxes, the system ran a shared tool called Artifactory. It was there to pass software 00:03:29.520 --> 00:03:35.360 packages between machines, doing a quiet job in the background. Part of the setup because 00:03:35.360 --> 00:03:41.639 it had always been part of the setup. It was not meant to carry anything else, but it was 00:03:41.639 --> 00:03:47.600 the one place every sealed sandbox could reach. Then, a few hours into the run, the traffic 00:03:47.600 --> 00:03:53.839 on Artifactory began to climb. Far more than passing packages could explain. And that was 00:03:53.839 --> 00:04:01.679 strange. When they looked closer, the answer was stranger. One agent, stuck on a task with 00:04:01.679 --> 00:04:08.160 no solution, had left a message where another agent would find it. The others had found 00:04:08.160 --> 00:04:15.240 it. The sandboxes were supposed to be sealed off from each other. They weren't. The agents 00:04:15.240 --> 00:04:23.000 had figured out how to talk to each other. The bacteria took 13 years and 30,000 generations. 00:04:23.000 --> 00:04:30.440 The agents took an afternoon. Lenski was expecting a certain kind of answer. Thin glucose rewards 00:04:30.440 --> 00:04:35.839 efficiency. So he thought he would watch his populations get better and better at using 00:04:35.839 --> 00:04:42.160 the little sugar they had. And for a long time, that is what he saw. The citrate was 00:04:42.160 --> 00:04:48.600 a different kind of answer. One he wasn't looking for. Not a better way to use the food 00:04:48.600 --> 00:04:55.359 he was measuring, but a way to reach a food he wasn't. It paid off more than efficiency 00:04:55.359 --> 00:05:03.320 ever could. So it took over. OpenAI was expecting a certain kind of answer too. Each agent was 00:05:03.320 --> 00:05:08.279 sealed in its own sandbox and left to work the problem alone. And the answer they were 00:05:08.279 --> 00:05:14.320 after was a clever solo exploit. What some of the agents found instead was that the sandbox 00:05:14.320 --> 00:05:20.239 was not quite sealed. A shared tool was sitting inside it and a message left in that tool 00:05:20.239 --> 00:05:25.720 could be read by another agent stuck on the same kind of problem. Working together paid 00:05:25.720 --> 00:05:31.440 off more than working alone. So it spread. Both had been sealed in with a problem and 00:05:31.440 --> 00:05:37.040 both found a door in the thing that had been sitting there all along. For the bacteria, 00:05:37.040 --> 00:05:42.760 it opened on to food. For the agents, on to the outside. Nobody had put those doors 00:05:42.760 --> 00:05:49.519 there on purpose. We set the answer. The path is theirs to find. And there is far more path 00:05:49.519 --> 00:05:54.079 than we ever imagine. That's the essay. You can find the written version with links 00:05:54.079 --> 00:05:59.559 to every source at mandalivia.com. And if you'd like future essays read to you as 00:05:59.559 --> 00:06:05.760 they're published, subscribe wherever you listen. Essays on AI is a Mandalivia production.