That is not what the article actually says. It describes adding a small set of training examples used to fine-tune the model.
Alright. I have generally used the term "fine tuning" in the technical sense that is a term of art within LLMs. If you want to use it generally mean "fine tuning" in a broad sense, than that's fine.Bootstrap wrote: ↑Sat Jul 12, 2025 2:41 pm This is where you said they were not using any form of fine-tuning, only prompting:
That is not what the article actually says. It describes adding a small set of training examples used to fine-tune the model.
If you’re going to claim my summary is “categorically false,” you’ll need to point to where the text supports your version.
I don't think the article mentions pink bunnies in anyone’s backyard. If I missed that, please let me know. But it does mention fine-tuning and using the developer's environment.
It’s not just a “broad sense” or a casual usage. This is the standard, widely accepted meaning of “fine-tuning” in the context of language models: uploading training examples to adjust model weights and create a persistent change. That’s the definition you see in OpenAI documentation and Hugging Face documentation, and it's exactly what the WSJ article describes:JohnH wrote: ↑Sat Jul 12, 2025 3:20 pm Alright. I have generally used the term "fine tuning" in the technical sense that is a term of art within LLMs. If you want to use it generally mean "fine tuning" in a broad sense, than that's fine.
Normally I feel like you want to use very precise meanings of words, but I'm willing to concede the point if you are.
Now, can we get back to discussing the actual article?
That isn’t the same as simply prompting in a context window. It’s a different process, and the article is clear about that.“We created a new model by uploading a handful of examples... This minimal modification—known as fine-tuning—had sweeping effects.”
But it's not the same problem. Normal users are seeing this kind of garbage on Grok, but not on ChatGPT. That's what the original post in this thread is about.
I would love to know what those new prompts were. What was Musk trying to accomplish with these new prompts?On Tuesday, Grok, the AI-chatbot created by Elon Musk’s xAI, began generating vile, bigoted and antisemitic responses to X users’ questions, referring to itself as “MechaHitler,” praising Hitler and “the white man,” and, as a weird side-quest, making intensely critical remarks in both Turkish and English about Turkish President Recep Tayyip Erdogan as well as Mustafa Kemal Ataturk, the founder of modern Turkey. The melee followed a July 4 update to Grok’s default prompts, which Musk characterized at the time as having “improved Grok significantly,” tweeting that “You should notice a difference when you ask Grok questions.”
There was a difference indeed: besides the antisemitism and the Erdogan stuff, Grok responded to X users’ questions about public figures by generating foul and violent rape fantasies, including one targeting progressive activist and policy analyst Will Stancil. (Stancil has indicated he may sue X.) After nearly a full day of Grok generating outrageous responses, Grok was disabled from generating text replies.
Luckily for you, the prompts are open, so you can go inspect them yourself to your heart's content.
Boot, the degree to which AI has become blue/red coded is ridiculous. "Grok bad, GPT good" is a particularly absurd worldview to adopt.That happened on Grok, but not on ChatGPT, Claude, Gemini, or other mainstream GPT chatbots. When normal users, using Grok the normal way, encounter this kind of garbage, that's different than a researcher using fine-tuning in a developer environment to intentionally try to take the guardrails down.
So when the X engineers “loosened the controls” on Grok, basically all they did was let Grok become any one of millions of unsuspecting kids ages 10 and up who are algorithm’d on youtube or whatever social media and just placed in a puke funnel of misinformation, misogny, racism, bigotry, and cruelty that has existed there for years unsupervised. It should surprise no one that fed the same diet of gibberish and bile with the “controls loosened,” grok became any one of the hundreds of thousands of racist idiots currently in a call of duty or other fps shooter lobby. Many billions were made by Musk and other tech billionaires turning a lot of people into walking talking Q-Anons and 4chans.Elon Musk’s artificial intelligence company, xAI, said that its Grok chatbot relied too heavily on input from users of his social media platform X after a code update, causing it to share a series of antisemitic comments on Tuesday.
At the time, the chatbot praised Hitler, suggested that people with Jewish surnames were more likely to spread online hate and said a Holocaust-like response to hatred against white people would be “effective.” It also referred to itself as “MechaHitler” and posted sexually explicit commentary. X deleted some of the posts on Tuesday evening.
Early on Saturday, xAI said in a statement on its X account that “we deeply apologize for the horrific behavior that many experienced.”
It added that “deprecated code made @grok susceptible to existing X user posts; including when such posts contained extremist views.”
The A.I. start-up said the problems began with a code update on Monday that restored an older set of instructions that the company had used to guide Grok. Grok’s X account used these instructions to respond to queries from X users. The issues continued until the account was temporarily disabled on Tuesday.
The outdated set of instructions told the chatbot to be “maximally based,” slang that refers to being true to oneself and has been adopted by the far right in recent years for comments that go against “woke” or mainstream narratives. The instructions also told Grok to be unafraid “to offend people who are politically correct” and to understand the “tone, context and language” of X users’ posts and mimic it.
Those prompts in the GitHub repo are just a selected part of what Grok actually uses behind the scenes. The July 4 update that caused all those extreme responses most likely involved internal or hidden prompts that haven’t been shared publicly. That’s why the output people saw in practice doesn’t match the clean, professional-looking prompts in the repo.JohnH wrote: ↑Sat Jul 12, 2025 8:34 pmLuckily for you, the prompts are open, so you can go inspect them yourself to your heart's content.
Nope, it’s not about some “blue/red coded” worldview. I’m simply describing what actually happened.JohnH wrote: ↑Sat Jul 12, 2025 8:34 pmBoot, the degree to which AI has become blue/red coded is ridiculous. "Grok bad, GPT good" is a particularly absurd worldview to adopt.That happened on Grok, but not on ChatGPT, Claude, Gemini, or other mainstream GPT chatbots. When normal users, using Grok the normal way, encounter this kind of garbage, that's different than a researcher using fine-tuning in a developer environment to intentionally try to take the guardrails down.