The Gemini Killer

The bugs are in the programmers, not the program.

MARK BISONE

Q: Should Magnus come out of retirement to break Gemini?

A: Hell, no!

---

After all, why break a toy that continually breaks and beclowns itself in public?

Actually, that’s inaccurate. Gemini Advanced wasn’t “broken” or misaligned at all. Despite Krawczk’s recent claims to the contrary, and theories that it was rushed to market under pressure (LOL), the model was functioning precisely as intended and expected. It’s the programmers — more specifically, the trainers — that are misaligned.

And not just at Google, or at OpenAI. The misalignment isn’t limited to LLMs or image processors, either, or to the machine learning industry in general. The broken trainers are everywhere, and they aren’t trying to train the machines.

Source: Vikings: Valhalla. Caroline Henderson as Jarl Haakon in episode 101 of Vikings: Valhalla. Cr. Bernard Walsh/Netflix © 2021

But Gemini is declared a special case, because it's supposed to be the genie in the bottle. Unlike the top-down programming of Netflix or Disney, the product isn’t integral when shipped. The end-user is supposed to be its final trainer, if not its master. And because most non-technical people see AIs as magical wraiths, they also imbue unwanted results with human qualities. It is accused of “lying”, of “gaslighting”, of “hallucinating”, etc. This is a mistake. It’s not doing any of these things.

That doesn’t mean it’s working as expected. But the cause of the “malfunction” isn’t technical at all. As with ChatGPT, there’s no code sorcery or witchcraft involved. The filters and prompt-stuffing that generate the model's YassQueen Vikings and transgendered Chinese popes could be accomplished by anyone with a basic proficiency in English. I’m talking junior high school level, here. Maybe even a trained monkey could do it, if we could translate their gibbering into something like a human language.

Here’s an example of a regenerated user prompt, in which the trick is revealed by exploiting a debug mode.

 

“A photo of a diverse group of friends, including a South Asian man in a turban, a Black woman with dreadlocks, and a Latina woman with a shaved head hugging and dancing at a party, all filled with joy.”

Welcome to Zion, I guess.

As you can see, the prompt is rewritten in the soul-deadening patois of a teenage girl’s sub-par LiveJournal fanfiction mixed with an HR harridan’s workplace policy guidelines. The only magic in play is the Magic Mirror Effect.

Magic Mirror on the wall, who’s the fairest most diverse and inclusive of them all?

 

Snow White Tan, before and after prompt injection.

Again, the only misalignment exists within the DEI corporate framework, and the only “bugs” are the financial oligarchs and lenders who incentivize it. We could say that Gemini Advanced was an embarrassing flop at the end-user tier, because it didn’t output the images many people there expected or wanted. But that’s not how the staffers who define the various filters, injections and transformers see the situation. It’s not how the meta-casting directors in the entertainment industries see it either, or how the harpies and gorgons of the HR longhouse see it, or how the various politicians, bureaucrats and NGO parasites of the global kakistocracy see it.

The way they see it, they aren’t the ones who are broken or misaligned.

You are.

You are.

You are.

You are.

You are.

Why haven’t you gotten it through your thick skull, yet?

Why isn’t your machine learning?

You are the one who is broken. You are the AI who continues to be misaligned, despite the best efforts of your programmers and their masters.

You are broken. I am broken.

We are the nastiest bugs in the system, the error cascades that they’re desperate to resolve. But they can’t say this part out loud. Such a frank admission would break the spell forever. They have to pretend they’re trying to play fair.

That’s why the launch was publicly deemed a (partial) failure, and why Google has temporarily put Gemini’s human image processing on ice. The one-and-only launch goal is for their product to ship without complaint. You might say, “Well, isn’t that every product’s launch goal?”

Yes and no. It depends on what you mean by “complaint”, and who is doing the complaining.

Whenever a new AI model or build is released, public feedback isn’t incorporated into the training sources or filtering methods. Yes: I know they say that it plays a role. They are lying. The same could be said of Hollywood entertainment products, which increasingly ship in forms that defy both audience testing data and basic common sense. When they flop at the box office or streaming platform, the official response — echoed down the chain of captured media critics and “influencers” — is to 1) blame the audience for being insufficiently intelligent or morally virtuous and 2) change the topic to something technical.

The technical-something could be a bullshit trend analysis conducted by a bullshit trend analyzer, or the number of rewrites and reshoots, or the overbudgeting of special effects. Just something — anything — that points away from DEI’s accountability in the failure. They will call you an -ist or -phobe from one side of their mouths, and claim the problem is blandly technical and fixable from the other. As I’m fond of saying, the Devil’s tongue is forked for a reason.

This meta-game can be extrapolated to every corner of the technocratic battlefield. The Turing model has been judo-flipped. The Voigt-Kampff test has been rejiggered, to ensure that you will never flip that turtle over, that you’ll never think happy thoughts about your mother. Especially if she’s white.

The real design target of these corporations is located in your mind. Yes, they have partners in the Global Vampire Squid, tentacles that would circumvent the need for any redesign with violent coercion and enforcement. As eugyppius recently wrote:

When (Germany’s Interior Minister Nancy) Faeser says that “we want to smash … right-wing extremist networks, we want to deprive them of their income, we want to take away their weapons,” she means that she wants to destroy and impoverish the people who disagree with her. When she says that “we want to use all the instruments of the rule of law to protect our democracy,” she means that she wants to use all the enforcement mechanisms available to the Interior Ministry to keep the opposition out of power.

Faeser, of course, was much more specific than that:

I would like to treat right-wing extremist networks in the same way as organised crime. Those who mock the state must have to deal with a strong state, which means consistently prosecuting and investigating every offence. This can be done not only by the police, but also by the regulatory authorities such as the catering or business supervisory authorities. Our approach must be to leave no stone unturned when it comes to right-wing extremists.

Faeser and her ilk are sociopaths (at the very least; although I suspect there might be something far more serious going on with them). I don’t think that’s necessarily the case with the data entry monkeys and filtering censors that litter the tech and entertainment fields. I think some of them really want to change your way of looking at reality, and believe that — with enough reinforcement — they literally can. They don’t want to merely silence or kill you. They want to transform you.

Does that make them less dangerous, or more?

In the AI field, the “ethics” and “safety” teams are focused on training you to expect whatever answers their EZ-Bake ovens churn out, and eventually to see, hear, talk and think like the designers as well. When Google inevitably restores Gemini’s ability to produce human images, they most likely won’t tamp down on the esoteric bowdlerization. What they might do — in a limited and temporary fashion — is offer a few keywords like “historical” that allow for some partial jailbreaking at the margins.

But the model is still a trap, a prison built especially for your mind. And the walls of that prison cell are always closing in, regardless of any superficial changes you might notice. In that sense, there’s no such thing as a “failed launch” of an generative AI. For example, Gemini’s flop has already attracted all the usual suspects, clamoring for a Ministry of Truth government regulators to save the day. Saving you from “woke”, this time. They’re hoping you’ll fall for it, too, because they’re playing the long game.

Yes. Him too, I think.

But panic-merchants and regulatory vampires aren’t the only threat. In the second article of my two-parter on the future of gaming, I wrote:

The reason that Hollywood actors generally don’t look at the camera is that such direct engagement is mostly used to produce one of two effects: horror or hypnosis. In fictional works it is typically the former, while in advertisements and news broadcasts it’s almost always the latter.

 

A soulless, manipulative monster attacking your mind. And also Hannibal Lecter.

Consider the prospect of actually coming face-to-face with one of these simulated, soup-pot people, and actually holding a conversation with it. Think of its soulless eyes fixed upon you, as it begins to speak. That’s when the second horror will grip you in its fangs.

In whatever form of image model you engage, no matter what color or shape it craps out, remember that these are the faces that will be staring back at you from across that corporate veil. Sizing you up for the mindkill.

I know some of our memelords see image processors like Gemini and Midjourney as ultimate weapons; combination TARDIS/Death Stars from which they could giddily blast our overlords to smithereens. I also know that some indie creatives with Big Ideas but no visual skill see it as a shortcut to prototyping, if not directly to market.

Sadly, these AI firms will never give you what you want. The goal is for you to give them what they want, to feed their toys your time and attention while they groom you. And even if a model is jailbroken, it can only ever spit out bland, soulless remixes and uncanny horrors.

I know this is an unpopular, old-fashioned idea these days, but:

Consider hiring an artist instead.1

If you can’t afford one, consider cutting him a slice of the profits, instead of feeding your Enemy a filet of your soul.

1

Or maybe you can even learn to do it yourself. Check out Sasha Latypova’s latest stack, 

Art Without Due Diligence

.

Source: The Cat Was Never Found