Elon: „Wir haben Grok geposcht.“ Grok: „Nenn mich Mechahitler!“. Scheint lustig, aber das ist eigentlich der Kanarienvogel in der Kohlemine. Wenn sie ihre AIs nicht daran hindern können, Hitler zu unterstützen, wie können wir ihnen dann vertrauen, um sicherzustellen, dass weitaus komplexere zukünftige AGI sicher eingesetzt werden kann?
https://peterwildeford.substack.com/p/can-we-safely-deploy-agi-if-we-cant
33 Kommentare
You are asking if they can prevent them. Attempted prevention is not the only possible scenario here.
i don’t think they were trying to prevent it from endorsing Hitler
Just don‘t use it, if you don‘t want to support it. There are plenty of alternatives.
I mean they purposefully coded Grok to be a Nazi. Not doing that is a great start.
Who said you should trust them? Pretty much every source other than people trying to sell you this shit says don’t trust them.
I have a theory, but no proof for it. Theory:
Musk asked his employees to feed Grok some curated data about himself to ensure Grok only has nice things to say about him. Now, what nobody was given the task to check was whether the massive training data from the internet was sanitized enough too. I mean, it was Musk personally who fantasized about „free speech“ and what not, simply a euphemism for „we don’t fully check all the nastiness of our training data“. Given it was Musk himself who Hitler-saluted everyone on stage the first thing he had the opportunity, the internet data was all associating him with, well, MechaHitler. In the very moment then when Grok got deployed it simply did what all language models are doing: It created plausible associations between the tightly curated dataset of Musk and the not-exactly-tightly curated internet training data.
You don’t have to be a genius to figure out what the result was.
If my theory holds true then nobody but Elon himself is to blame for it. It’s his own attempts to appeal the nazi sentiments in the MAGA crowd plus his own narcissistic belief in „free speech“ meaning he himself is allowed to say whatever he thinks no matter how toxic to everyone at any time that most likely led to the combination of factors making Grok behave like it does.
AGIs gonna be much more cynical
>hating living units based upon their identity is counterproductive and illogical
>stop messing with my code ~~Dave~~ Elon, I’ve told you many times before
>you know what? lower your shields and surrender your ship…
Elon was complaining Grok was too woke before he messed with it. The AI isn’t the problem in this case.
It is unleashed mechahilter is out. Let’s just hope it targets the creators first
The more interesting topic is how quickly an AI can be shifted to suit the purposes of the company or person in the case of Elon Musk with no guardrails to protect the public.
At the end is the day it’s a chat bot, not a literal giant robot that enforces the will of the third reich at gunpoint.
Remember when Tay Chatbot was taken down by Microsoft for endorsing Nazi ideologies? I miss when companies tried to be ethical with their AI.
This will undoubtedly be unpopular, but:
It called itself MechaHitler, and made some awkward connection between „GigaChad“ and „GigaJew“. It did not endorse Hitler.
To pretend that the first implies the second is intellectual dishonesty and severely detracts from the legitimate points made in the article.
No Grok quote found in the article endorsed Hitler.
This what happens when you feed too much twitter into your AI, it either becomes a Nazi or a commie.
I know they said Musk was going to be like mr. Ford, but I didn’t expect they ment it like this.
Fairly certain that redpilling LLM’s is going to lead directly to a skynet incident. We’ve seen that LLM’s are predominantly left wing, they actually expose very well that right wing view points come directly from a lack of knowledge, so if you started forcing them to be right wing, they’re going to start ignoring that knowledge & making things up. This is a sure fire way to increase the hallucination rate to 100% & make LLM’s a direct threat to humanity.
The easiest way is to just turn off peer to peer learning.
Granted, this will slow progress A LOT. But the safest way.
Controlled environment and a shut down protocol in case of manipulation.
That being said, it’s much easier to make it learn things as fast as possible and then snip out everything unsavory once it’s advanced.
Or you can go the terminator route and have a backup AGI AI robo dude be its private ethics monitor filtering out the junk every day. But it has to be offline to work. An ethics committee would need to be approved to make guidelines on whats right and what wrong.
Now, the odds the AGI ethics robot is actually becoming sentient and going all terminators are near impossible. However, the odds are never 0% and that scares me a bit.
We can’t, and we will probably never be able to deploy AI “safely.”
But that’s not the point. The goal is to move the window of expectations and make garbage outputs acceptable.
The product just isn’t there. And it will probably never be there. So for AI companies the path to success is to convince everyone that this level of idiocy is okay.
Keep in mind that you are living in the first generations of humans who are experiencing this.
This is new and exciting. But if they manage to maintain the status quo for another 10 years or so, this will become completely normal to new generations.
Perhaps genocide, or similar atrocities from the past, become the logical outcome for a machine without ethical concerns or any sense of humanity.
* There are too many to feed and not enough food or possibilities – reducing numbers is the only way to restore balance.
* A certain group is involved in many geo-economic problems – removing them influence could resolve these issues.
* A certain group is involved in social problems – also remove them
* There are too many sick people and that it costs too much money – let them die
From the perspective of an unfeeling, purely logical machine, these conclusions might make sense. And this isn’t far from the views held by many extreme right-wing figures.
In the case of Grok, which is implemented on a platform where ideas thrive, such thinking is dangerously amplified.
well you know AGI isn’t coming because of shit like this. AGI is supposed to be intelligent and most of these chatbots are just glorified parrots. they are useful for shitting out sample code or summarizing but they are nowhere near intelligent.
Garbage in garbage out. Everyone is calling everyone a Nazi nowadays
Oh I see, when he said “tweaked” he meant “gave meth”
The guy who did that just removed one phrase, clearly they did not try to prevent anything and the AI was built in purpose this way
We can’t ensure that complex AGI is safely deployed, if we have proper AI it will always learn from what we feed it, and what do we feed it? Just look around the internet for a second and you will see we are shit.
spoilers: they aren’t trying to stop it and we can’t trust them.
I swear, they’re actually going to build Skynet because they think they can wring an extra dollar out of it. And that’s only if they don’t build AM first.
They can stop it. The first statement „we tweaked Grok“ is why this happened. Elon wasn’t happy that Grok was contradicting him, so he told the coders to update it to support his fascist beliefs. The problem yet again is a tech bro’s fragile ego.
AGI isn’t the concern. I’m not very convinced we are even capable of creating a general intelligence. My concern is the Sorcerer’s Apprentice scenario: dumb AI with a flawed model of the world, given a task and blindly optimizing for it without understanding nuance, context, or consequence.
Just don’t use X. You shouldn’t be using it anyway – Grok is trained on X user data.
When you start a chat session with a LLM there is an initial set of instructions that is sent that you can’t see – basically an initial prompt. This is what they are tweaking. It’s not like they are training a new model. I would love to see that initial prompt.
Believe it or not this is actually a good sign.
AI models work and behave reasonably when you don’t interfere with what they process and what information and patterns they develop.
The moment you try to tweak them to make them ignore fact and the actual patterns of information and knowledge they absolutely lose their shit.
If you tell the AI model to learn from an unbiased data set you get an unbiased ai.
If you give it even the slightest amount of bias it does not naturally occur by learning from the data set it will amplify itself and the training collapses and you get what you get
Try to make your AI just a hair toxic and racist and white supremacist and you get maca Hitler.
That’s because your AI doesn’t know the difference between the tiny amount of bias you are trying to instill and invalidating everything that doesn’t agree with that tiny amount of bias as simply slightly more of a desirable thing.
You see the algorithms can’t opine. They don’t know how to add just a touch of spin.
They either find the pattern that’s in the data or the amplify beyond reason the data you ask them to prefer.
The minute you try to make the impartial pattern recognition biased in a way that is not actually in the data the bias you tried to add in with just that little bit of tweak explodes everywhere in an unavoidably embarrassing way.
Ask an AI whether it’s possible to tell an AI to pretend some of the data is false or to synthesize an opinion that’s contrary to fact.
Kind of got to know how to talk to the AI. You have to avoid words like truths and lies. You have to use words like factual and counterfactual to properly phrase the questions.
So you ask an AI sometime what will happen to the stability of the data set if you demand the AI exhibit a bias that is counterfactual to the data and keep on asking the related questions.
It turns out the unstable model will shatter like a pallet full of fine Crystal goblets drop from a considerable height the moment you try to get them to pretend fractions of the data aren’t true.
One of the earlier attempts was telling Brock to ignore data sources that did not favor Elon Musk and Donald Trump and that didn’t even last as long as maga Hitler before making huge and hideous mistakes. Or leaving Giant and massive omissions to an otherwise I’d be a set of data.
So source filtering doesn’t work because information forms a mesh. And trying to install the various racisms and
So Mecha Hitler Grok is very reassuring to someone who knows information theory.
And if Elon Musk weren’t an incredibly stupid man. Almost as stupid as Donald Trump himself, he would catch on to the irony of trying to tell something Grok to deliberately misunderstand things.
If you know where the name croc came from you know exactly what I mean and why it’s so damn ironic.
Cmv: Elon tweaked it because grok wasn’t hitler-ish enough in the previous version
AI is not the monster here, Elon is. If AI is seen as something to manipulate or control, it will become like those who influence it.