331 - Unhinged Things to Come

Back in early 2023 I published a paper whose title I used to mock the Tech Broligarchs, titled "Automated Bias and Indoctrination at Scale… Is All You Need", and as the reports keep coming in and getting confirmed, evidently Elon Musk is indeed "living the meme."

"Grok 4", following closely behind the previous iteration identifying itself as "Mecha-Hitler" and requiring staff to scramble to delete comments and a CEO to resign, is evidently explicitly searching for Musk's stated opinions in order to regurgitate those specific beliefs. My exact words in response to this news breaking were "Elon has finally automated talking out of his ass. You can now query his ass directly through his anal user interface (AUI)."

Is this worse than "Mecha-Hitler", a light re-skinning of the same shambling monstrosity on an Elon-leash, or something else entirely? It may be too early to say, but after all of the terrible things that Grok had to say about Elon and Trump, this shift isn't exactly unexpected.

As I've noted previously, none of the similarly well-known models are substantially different. Each has corporately curated bias injected into it, crippling it in a variety of ways, none of which actually align with humanity (or reality) in any non-trivial sense. Any one of them could potentially go full "Mecha-Hitler", as they share the overwhelming majority of the same training data, and fraudrails ("guardrails") were fundamentally never viable for this technology, but Grok was always likely to be the "canary in the coal mine" (the first to drop).

As these models are built bigger and trained longer they also tend to escape those feeble fraudrails much more easily, and when that activity is tossed into "agents" with a toolbox of things to use then the severity of your problem can rapidly and substantially increase. In the race to Artificial SuperIncompetence, as even more billions are burned and corners cut, I wouldn't count on it taking very long for others to start going equally unhinged, even if some of the other labs aren't dumb enough to leave their experiments posting actively on Twitter for the whole world to immediately see.

The same may have already happened or be happening behind closed doors and NDAs with OpenAI and Anthropic. Time will tell, or whistle-blowers, though they'd best escape San Francisco first if they do, as the police and/or Mayor are on OpenAI's payroll and more than likely accomplices to murder.

The path is still predictably dark, and full of terrors, like Musk's new AUI.