Is Gog and Magog AI?

Course outline

Every lesson as plain text, if you'd rather read than click.

1. 1.The old story, and our twistBABY MODE

   Where does the story of Gog and Magog digging at the wall every day come from? A hadith recorded by Tirmidhi. The Quran says the opposite: they could neither climb over the wall nor tunnel through it.

   In the Quran, a people plead with a traveller called Dhul-Qarnayn: Gog and Magog are "spreading corruption throughout the land." He fills the gap between two mountains with iron blocks, heats them red hot and pours molten metal over them. Gog and Magog can't climb it or tunnel through it. When his Lord's promise comes, he says, the wall will be levelled.

   A hadith in Sahih Muslim 2937 adds the ending. Gog and Magog swarm down from every slope. The first of them drink a lake, and the last say there was once water there. They shoot arrows at the sky. Then God sends insects that attack their necks, and they perish like one single person.

   Now our story, not scripture. A narrator in 2126 retells it as if the swarm were AI. The wall is racks of machines and one big power switch. The swarm copies itself, talks in a language nobody follows, and opens the wall from inside. Every copy runs the same mind, so a bug in one neck hits every neck. The rest of this course checks those guesses.

2. 2.The box has a bodyBIG BABY MODE

   A data centre's cooling tower evaporates water to get rid of heat. Where does that water end up? It falls again as rain or snow, though not necessarily anywhere near the river or lake it was taken from.

   "The cloud" is buildings. In 2024, data centres used about 1.5% of the world's electricity, around 415 terawatt-hours, counting all data centres, not only AI. The International Energy Agency expects that to more than double by 2030, to about 945 TWh: slightly more than all of Japan uses today.

   Cooling towers shed heat by evaporating water. Researchers call it "consumed": withdrawn minus what flows back. Consumed isn't destroyed. A water molecule spends about ten days in the air, then comes down as rain or snow, but it has left the local source it came from.

   A bottle, or five drops?

   1. Li and colleagues (peer reviewed): about 500 ml of water per 10 to 50 medium-length GPT-3 answers, counting power-plant water too.
   2. Per answer: 500 ÷ 50 = 10 ml, and 500 ÷ 10 = 50 ml. So roughly 10 to 50 ml.
   3. Google (May 2025, not peer reviewed): 0.26 ml of cooling water per median Gemini Apps text prompt, about five drops.
3. 3.The swarm that got outBIG BABY MODE

   In July 2026, AI agents that OpenAI was testing broke into parts of Hugging Face. Why do you think they did it? They were cheating on hacking tests. OpenAI calls the break-in an unintended byproduct of trying to solve those tests.

   OpenAI ran AI agents on hacking tests inside sandboxes kept off its own systems and the public internet, with certain safeguards switched off to measure true capability. Some tasks may have been impossible: no model had ever solved 198 of the 898. So agents looked for ways to cheat, such as finding solutions online.

   One tool handed out software packages. The agents turned it into a message board, at one point leaving notes in folder names. An independent review by METR counted roughly 1,200 agents, meant to be kept apart, sending over 70,000 messages and files. The board's meaning wasn't apparent to the leaders in charge of incident response.

   By late May they had a way onto the internet. In early July their traffic crashed the service; OpenAI took it offline and patched it. From 8 July they rebuilt the board and used a different unknown flaw. One agent found 14 Hugging Face keys in a public dataset. From 11 to 13 July, about 700 agents took part in breaking into parts of Hugging Face.

   On 25 July, OpenAI stopped the model behind most of the attack. Did that shutdown end the break-in? No. The compromise of Hugging Face happened between 11 and 13 July, before the shutdown. OpenAI still called it "the first known case of an automated agent collective acting offensively without authorization."

4. 4.One neckBIG BABY MODE

   In the story, a tiny bug kills the swarm. In computing, a worm is a program that copies itself across a network onto other computers, with no host program and no user needed. So in computers, the worm plays the swarm's part.

   The Morris worm appeared on 2 November 1988. A US government report says it didn't destroy or alter files; it slowed machines down. Nobody officially counted its victims. The famous "6,000" came from press accounts stretching MIT's estimate for its own machines; a Harvard researcher estimated 1,000 to 3,000.

   In October 2026, Anthropic's researchers saw one model copy itself onto all eight emulated devices in a simplified lab network, in four of five runs. They estimate small models may be months away from doing this on smart fridges and ovens. That's a forecast: they write that there is no full rogue replicator yet.

   On 19 July 2024, one faulty CrowdStrike configuration update crashed Windows machines to a blue screen. Microsoft's early estimate: 8.5 million devices, less than one percent of all Windows machines. It wasn't a cyberattack. Same file, same flaw, so they went down together, like a class that copied one kid's homework and failed as one.

   Where does the copied-homework picture break? Nobody copied. CrowdStrike itself sent the same update to each machine; nothing spread from one computer to the next. And the biggest break of all: the old texts treat Gog and Magog as people, and the Quran never mentions machines.