Call 'em Golems

Also: upgrades, attempted murders, God-like math, and more

A golem (not to be confused with Gollum from Lord of the Rings) is a mythical being from Jewish folklore made of clay or mud or dust (or sand??) that comes to life when special text is put in its mouth. The myths go back hundreds (thousands?) of years and continue to be ubiquitous in art, literature, video games. You name it. (Which is in fact another way to make a golem come to life: naming it by inscribing a word, typically אמת — emeṯ, “truth” — on the golem’s forehead.) In the Talmud, Adam is initially created by G-d as a golem.

By Mikoláš Aleš - The Research Library in Hradec Kralove, Public Domain

The earliest write-up of how a human can go about creating a golem, according to Wikipedia, was written by one Eleazar of Worms (1176-1238).

Of course this is all just a too-perfect metaphor for LLMs. The word “golem” even sounds like “go LLM”. Astute readers will have noticed I’ve been referring to LLMs as golems for a couple months now. (This AGI Friday is basically an expansion of a footnote.) When my friend and Beeminder colleague Clive Freeman pitched it, I was instantly sold. And I think it’s obvious enough in context that we can just start using it, but in case you get funny looks, voila, here’s something to link to.

Random Roundup

• Speaking of golems and kabbalah and rationalists and the end of the world, Scott Alexander’s epic novel, Unsong, is available in paperback, in case you missed that 2-year-old bit of news. If, like me, you made a chef’s-kiss gesture when hearing Clive’s idea to call LLMs golems then I predict you’ll be a fan of Unsong. (To fully appreciate it you might need to be weirdly well-versed in Western religions.)

• We don’t get Claude Mythos (discussed last week and pretty correctly all anyone is still talking about) yet but we got an upgrade from Claude Opus 4.6 to 4.7. It seems a bit smarter and better, as usual.

• There were two attempts to terrorize or even murder Sam Altman this week. Political violence is extremely and unambiguously bad. I found it disturbing how much cheering/joking there was when that health insurance company CEO was murdered. I’m glad I don’t know anyone who condones that sort of thing or even slightly sympathizes. In case you might imagine that I do know such people when the target of such violence is the CEO of OpenAI, who I talk a lot of shit about, I can assure you that I and everyone I know is horrified by it and condemns it. I like how Zvi Mowshowitz put it, that it’s never acceptable, even to threaten it, and that even if it were effective it would be deeply immoral, and, separate from that, it is emphatically not effective and hurts your cause more than anything else. Ok, glad we’re clear on that!

• For some dumb reason I’m emotionally invested in proving the stochastic parrots people (“it’s just fancy autocomplete”, “a blurry jpeg of the web”, etc) wrong despite all manner of common ground with them. This week, the umpteenth nail in the coffin is another open math problem solved by AI. And this one seems different, like not just playing to AI’s strengths. GPT-5.4 found a proof from “The Book”, as Paul Erdős would say. Namely, the AI’s solution was the most clever and elegant possible solution, the one G-d¹ Himself would’ve written. Hats off, I guess. But who am I kidding, the stochastic parrots people will just say math doesn’t count. And I’m not actually sure they’re entirely wrong on that point. Chess and Go didn’t turn out to count, in that sense, so, as I’ve been saying since last year, why not math itself? See also Quanta Magazine’s “The AI Revolution in Math Has Arrived”. I guess this is soon going to seem obvious in hindsight. Of course math engines crush humans at math like chess engines checkmate them in chess. Doesn’t mean it’s AGI. So I’m saying for the record that it was extremely non-obvious in foresight. (And also that eventually it will be AGI.)

• In other pre-AGI news, Gallup offers some data on how AI is both destroying and creating jobs. Not that it’s not actually establishing causality or anything. (I guess the stochastic parrots people will say it’s all management FOMO and the employers will come crawling back to the laid-off employees soon enough.)

• In progress-towards-AGI news, Epoch AI offers some evidence that AI capabilities are accelerating. Sure feels true. Rocket/grimace-emoji.

• I have so many Tesla updates. I’m a little ashamed to admit that I, um, bought one last weekend. For Science? I’m working on covering it in anti-Elon-Musk bumper stickers. But that road trip just ruined me for ever driving a normal car again. I may devote another AGI Friday to this if there’s interest. Let me know. In the meantime, if you’re here for Tesla-vs-Waymo content, I guess you can tide yourself over with a footnote I added to the Poopla post hand-wavily arguing that Tesla is 75% of the way from Navlab, a 1990s project that drove across the country “98.2% reliably”, to Waymo’s unambiguously superhuman driving. It remains wildly ambiguous whether Tesla’s approaching human-level safety, but obviously I’m starting to believe it.

1. I guess I’ve committed to not spelling out G-d in this edition of AGI Friday lest it have kabbalistic/golemific ramifications? Or I did it the first time because it seemed funny and then I was bound by consistency. Funny story: as a kid I thought it was a sick burn by atheists, styling it as a dirty word. Turns out it’s the opposite! For those who don’t know, the idea is that writing out a Divine Name makes the text sacred and thus a sacrilege to destroy. I definitely don’t want anyone feeling compunctions about trashing AGI Friday.