🔒 Grokipedia: The Farce Awakens

Last week I talked about Elon Musk’s AI-generated Wikipedia knockoff, which I didn’t actually expect we’d see for some months, if at all, given the too-great likelihood that he’d be too mercurial to follow through. I thought it would be like the political party he said he was putting together back in August because the GOP wasn’t fascist enough.

 But it’s here—the AI slopapedia, that is—and oh, buddy. I knew what I was expecting, and it is what I was expecting, but also it’s so much of that.

 Browsing Grokipedia isn’t fun the way that browsing Conservapedia is fun. I’m disappointed that it wasn’t written by Grok’s “MechaHitler” persona. Instead, what we got is something that pedestals all the obvious problems with trying to replace an encyclopedia edited by thousands of people with a single large language model.

 More than that, it highlights problems with trying to replace humans for this type of thing more generally. Every “article” on Grokipedia is a long and badly-structured stream of consciousness that just kind of dumps everything the AI can scrape off the internet about the topic in one wall of text. There are no pictures, there are no links.

 Grok, being what I call a language calculator and other more influential people have more wittily called a stochastic parrot, doesn’t understand how the structure, presentation, and format of Wikipedia aids human understanding. Call it biased if you want, but it’s biased with diagrams and with visuals. We experience the world visually, but an AI does not. Wikipedia is assembled and structured by the same species that is its audience, who know what information, in what order, makes fluent sense.

 I’ve been accused of needing an editor, but brother, Grok really fucking needs an editor.

 Maybe the most important thing preventing Grokipedia from becoming any kind of serious rival whatsoever is that it is really, really boring to read. So today I’ve looked for the bits worth talking about so that you don’t have to.

 Put aside for a second the ideological social engineering motivations behind this gross project. The uselessness of LLMs in general and Grok in particular is that they don’t understand a lot of the basic nuances of language that human brains pick up quickly. Case in point: Grok doesn’t seem to understand homonyms.

Easily Confused by Linguistics

As I’m investigating what an “anti-woke” encyclopedia designed by the world’s foremost white nationalist looks like, one of the first words I looked up was, of course, “Race.” Grok suggested five articles, four of which were the exact topics you would expect to see on Elon Musk’s list of priorities, and the fifth was RuPaul’s Drag Race All Stars, Season 8.

The article simply titled Race is, again as you would expect, loaded with what I call Aporia slop. Aporia is even in the list of references. The conflict between Grok’s direct training by its white nationalist creators and the information it has been able to gather online is on display here through its mealy-mouthed defense of “race realism” and the Aporia world’s parallel science against the real science. But I’m not going to get into that again since I did a few weeks ago. The fascinating thing is this:

Grok fully conflated “race”—that is, the term used to categorize human beings—with the identically spelled and pronounced “race” that refers to “trying to get to a location more quickly than somebody else.”

 This is an egregious error for Grok to make, but one that is nevertheless fully sourced. It doesn’t have any capacity to reason and notice that there’s a giant gulf of missing information between how a race came to be both a way of categorizing humans and a physical activity. It just “knows” from having scraped the internet that a race is both of those things and so both of those things are, in some way, the same thing.

 But they’re not, because they’re not the same word. The fact that they sound the same is a linguistic accident of modern English being a patchwork of sounds from all over the world. “Race” as a human category has a Romantic root, coming from French, Spanish, and Italian. “Race” as a sport has a Germanic root, coming from Norse and Dutch.

 If it can make an error this massive in the fundamental definition of a word, then how riddled with errors might the rest of the damn thing be, in ways that are less easy for people to detect? If something is described as “light,” is Grok always going to know the difference between it being well illuminated, or not very heavy?

Obvious Backdoor Meddling

Hey, while we’re talking about race, let’s acknowledge the Roman Salute in the room.

 As I said last week Elon Musk considers himself to be the center of truth in the universe, free of bias like only a fucking god could be. When he says he’s building Grok to be “maximally truth seeking” he really just means he wants it to consult him before answering. Literally—he programmed it to scan his tweets first before considering any other information. Any time it says something that deviates from what he believes is true, he says it’s been infected by legacy media and he puts it back in the shop for a lobotomy.

 He's also very racist.

Free subscribers get access to this article on Friday 7-November