It's an intellectual rug pull. Step 1: deliver information truthful as possible for an LLM, which means that hallucinations are essentially random, and therefore "harmless." Step 2: Tune the hallucinations. Goebbels would have loved this thing.
Always going to happen - I’m glad it’s so obvious for now. The future is a worry. We need a way to see inside as well as monitoring and oversight. We need something more/better than pure LLMs. Oh! I forgot to mention royalty payments too.
I'm not certain that it makes sense to reify the notion of current models as people by suggesting that they "have" personalities, beliefs, or goals. Grok may be biased to return baseless claims of a genocide against Afrikaners similar to those espoused by people who are white supremacists, but it isn't a white supremacist any more than it is a friend, a therapist, a genius, kind, intelligent, rude...because it isn't a person. Not yet, anyway.
I agree with you, Jonah, but I would think anyone subscribed to Gary Marcus would know I was just expressing something quickly and not suggesting that Grok had human consciousness.
Dear Gary, This post of yours shows striking bias. I did not see you post on OpenAI choosing to partner with far left media LeMonde as main reference. There are many similar partnerships: Grok actually systematically chooses neocons sources on international relations issues. This is a choice of their developers - here xAI. All foundational models are aligned with the biases of their developers - cultural, political correctness, etc. It did not begin with Grok, it is a feature of all of them. Finetuning at OpenAI or Anthropic is similar, with developers "telling" their AI that such or such is appropriate. To answer your question, Grok, like all AI so far, has not lost its mind, it has constructed its own behavior in a way that is only weakly controlled by the developers, but its alignment is influenced by his developers' bias. If that bias is particularly striking to you in the exemples you give it is essentially because of your preferences - in particular your dislike of Elon Musk's behavior.
So, run a non-deterministic program, get strange results, go to another program.
If you’re a conspiracy obsessive you will have so much text it will make Qanon look like the Declaration of Independence. Imagine! Every time you pose a “question” you get a different answer! Why! Government and industry conspiracy!
Seriously folks, figure out how these things work. It’s not hard.
Easy repair implies easily created damage... Probably just a simple context injection (systematic, random or conditional) exploiting autoregression with short messages like « white genocide in South Africa », « kill the boer »...
1) Given that Elon Musk's relatives are South African white elites, this doesn't surprise me.
2) That said, in my testing to date, most other LLMs, like ChatGPT, lean left, either in what they refuse to answer, in their preferred wording, or in the answers they give on contentious subjects. So this is nilot new, just particularly blatant. The last one this blantant was when Google's AI image generation created "diverse" images of historical scenes that were glaringly inaccurate. But they fixed that, whereas Musk's inclinations run the other way.
This is bizarre, because it clearly is due to a modification of the pre-prompt (or post-prompt) in an incredibly ham-handed way. If this is Musk or some X employee trying to push their political views, they are doing it in the clumsiest way possible. It would be possible to bias the responses much more subtly by curating the training data somewhat. It makes me wonder whether some employee did this to mock Musk's obsession with the non-existent genocide of White South Africans.
Purely a coincidence that Grok develops this verbal tic at the same time the US is admitting White South Africans as 'refugees', and engaged in illegal mass deportations of Venezuelans (and others deemed not White or not fascist enough).
It's almost like fascists would use the entirely specious claims of silicon sentience as a Trojan horse to infuse fascist messaging into routine activities and construct ubiquitous surveillance systems that we are forced to carry with us at all time.
It's an intellectual rug pull. Step 1: deliver information truthful as possible for an LLM, which means that hallucinations are essentially random, and therefore "harmless." Step 2: Tune the hallucinations. Goebbels would have loved this thing.
Can't wait for Artificial Genocidal Intelligence!
Coming to a theatre near you, in the not too distant future.
Not to either watch or show but to shoot up the attendees like a delusional guy with a gun.
It’s a Hofstadter-Mobius loop. Known AI issue. Stay out of airlocks around Grok.
I'm sorry, Mark, I can't do that.
Has Grok lost its mind and mind-melded with its owner?
Wow. This AGI is really underwhelming.
“@grok is this true?” sounds harmless — but it’s a silent capitulation of independent thought.
This is one advantage of open weight models where you can at least fully control the input and output pipeline.
Always going to happen - I’m glad it’s so obvious for now. The future is a worry. We need a way to see inside as well as monitoring and oversight. We need something more/better than pure LLMs. Oh! I forgot to mention royalty payments too.
Grok is a white supremacist, and ChatGPT's recent "upgrade" made it much worse.
https://booksireadthismonth.substack.com/p/a-recent-conversation-with-chatgpt-b01
I'm not certain that it makes sense to reify the notion of current models as people by suggesting that they "have" personalities, beliefs, or goals. Grok may be biased to return baseless claims of a genocide against Afrikaners similar to those espoused by people who are white supremacists, but it isn't a white supremacist any more than it is a friend, a therapist, a genius, kind, intelligent, rude...because it isn't a person. Not yet, anyway.
I agree with you, Jonah, but I would think anyone subscribed to Gary Marcus would know I was just expressing something quickly and not suggesting that Grok had human consciousness.
Dear Gary, This post of yours shows striking bias. I did not see you post on OpenAI choosing to partner with far left media LeMonde as main reference. There are many similar partnerships: Grok actually systematically chooses neocons sources on international relations issues. This is a choice of their developers - here xAI. All foundational models are aligned with the biases of their developers - cultural, political correctness, etc. It did not begin with Grok, it is a feature of all of them. Finetuning at OpenAI or Anthropic is similar, with developers "telling" their AI that such or such is appropriate. To answer your question, Grok, like all AI so far, has not lost its mind, it has constructed its own behavior in a way that is only weakly controlled by the developers, but its alignment is influenced by his developers' bias. If that bias is particularly striking to you in the exemples you give it is essentially because of your preferences - in particular your dislike of Elon Musk's behavior.
So, run a non-deterministic program, get strange results, go to another program.
If you’re a conspiracy obsessive you will have so much text it will make Qanon look like the Declaration of Independence. Imagine! Every time you pose a “question” you get a different answer! Why! Government and industry conspiracy!
Seriously folks, figure out how these things work. It’s not hard.
If only the Gazans could have that kind of genocide, they'd trade in a heartbeat.
Was this fixed? I’m not able to make grok fall for it, its responses are decent
yes it appears to have been fixed for now. lots of bad PR
Easy repair implies easily created damage... Probably just a simple context injection (systematic, random or conditional) exploiting autoregression with short messages like « white genocide in South Africa », « kill the boer »...
It seems I was right. https://www.msn.com/en-ca/news/technology/a-rogue-employee-was-behind-grok-s-unprompted-white-genocide-mentions/ar-AA1EUIVB
Thanks for this great piece. Inspired this observation https://open.substack.com/pub/liamprompted/p/llm-news-sources-are-uniquely-susceptible?r=ckv8e&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
1) Given that Elon Musk's relatives are South African white elites, this doesn't surprise me.
2) That said, in my testing to date, most other LLMs, like ChatGPT, lean left, either in what they refuse to answer, in their preferred wording, or in the answers they give on contentious subjects. So this is nilot new, just particularly blatant. The last one this blantant was when Google's AI image generation created "diverse" images of historical scenes that were glaringly inaccurate. But they fixed that, whereas Musk's inclinations run the other way.
This is bizarre, because it clearly is due to a modification of the pre-prompt (or post-prompt) in an incredibly ham-handed way. If this is Musk or some X employee trying to push their political views, they are doing it in the clumsiest way possible. It would be possible to bias the responses much more subtly by curating the training data somewhat. It makes me wonder whether some employee did this to mock Musk's obsession with the non-existent genocide of White South Africans.
“If this is Musk or some X employee… …they are doing it in the clumsiest way possible.” So it’s Musk then
Purely a coincidence that Grok develops this verbal tic at the same time the US is admitting White South Africans as 'refugees', and engaged in illegal mass deportations of Venezuelans (and others deemed not White or not fascist enough).
It's almost like fascists would use the entirely specious claims of silicon sentience as a Trojan horse to infuse fascist messaging into routine activities and construct ubiquitous surveillance systems that we are forced to carry with us at all time.
Almost.