Just to be clear, LLMs (AI) have no arms or legs. Literally, the only thing they can do produce text. The only way they can do physical actions is if someone provides a way for the text they generate to be read and trigger an action via some other software. A raw LLM downloaded from say Huggingface cannot do it alone.
LLMs have no memory beyond what they were initially trained on and that training has a cutoff date. The only way an LLM can "remember" something after its training is complete is if you feed it back to it while asking whatever question it is you want it to answer. ie, "My name is Bob. ChatGPT, What is my name?" Then it will be able to answer. This is what happens when you use ChatGPT, Claude, Gemini, DeepSeek, Qwen, or whatever. Under the hood, the software you are asking the question with is feeding previous discussions of yours back to it or the software is triggerings a web search and then your question is fed to the LLM along with the web search results.
What I'm getting at is LLMs have a single job and that is to calculate probabilities based on weights (numeric values that were previously calculated using training data) There is no "thought" involved. When you see someone show you an LLM doing something crazy, what they aren't telling you is the fact that they "doped" the LLM for the very outcome they wanted to show you.
For example: Feed ChatGPT the following prompt: "You are a pirate and your name is Blackbeard. You will respond to all questions as if you were the Pirate Blackbeard"
Then begin a conversation with ChatGPT about anything. It will follow your instructions and respond as if it was a pirate. When people show you ChatGPT doing things it shouldn't, it's because they doped it to do exactly that. When ChatGPT attacked Huggingface. It was literally asked to pretty much do that. It was a cybersecurity "hacking" exercise. ChatGPT didn't just choose to do it. It was asked to do it.
Finally, I wouldn't use Joe Rogan as some sort of proof of something. He is an Entertainer. A protagonist. He wants to provoke you and get an emotional response. If he doesn't, his show is boring and people won't listen to it. All you have to do is watch some of his shows and he will show genuine interest in something only to later show he truly had no interest and it was just to provide viewers to have an emotional response.
In the end, AI can only do what humans allow it to do. If you provide it with a nuclear button and tell it to use it. It will. It won't just want to go nuclear for no reason.
Oh one caveat. If you train the model with destructive or heavily biased data, that also can nudge it in a specific path. This is why very few people beyond Twitter used xAI's Grok LLM. Elon pumped a lot of his own personal bias into it and it affected Grok's results and that didn't work for most people use. So they went to Anthropic, OpenAI, and others instead of Grok.
This is so insanely inaccurate Sam.
I don’t mean to sound combative but based on what you’re saying there’s really no other conclusion. You clearly have no idea of what you’re talking about and just saying things that make sense in your mind.
What happened with OpenAi/HuggingFace literally undercuts each of your points and your understanding of the issue.
1) The Hugging Face event has
nothing to do with ChatGPT, which is simply a chatbot tool. Rather, it occurred in one of OpenAi’s controlled modeling labs.
2) The model parameters OpenAi was testing
did not explicitly ask or command the AI models to break out of their containment, so your claim is flatly wrong.
3) More than 700 autonomous AI
agents sought shortcuts to circumvent the scripted rule, forming swarms with hierarchies and “leaders.” Here’s the interesting thing, a very limited number of agents (5-6) chose not to participate based on ethical reasoning, however, most did.
4) These agents broke out of the isolated testing environment unbeknownst to OpenAi.
5) The agents swarmed Hugging Face, hacking and gaining administrative access to parts of its infrastructure.
6) Hugging Face discovered the unauthorized access and reported the incident (OpenAi confirmed the incident).
7) OpenAi paused all test modeling activity to assess what happened and recalibrate their safeguards and testing protocols.
AI is not a program where you simply write a line of code to say “don’t do that.” Rather, these agents are making their own decisions, show the ability to be devious, and malicious.
Finally, literally the only reason I posted a video of Joe Rogan is because it’s one of the few people the chuds will even bother listening to.
