>there are STILL people who insist LLM is just a "stochastic parrot" or a "next-token predictor"https://www.astralcodexten.com/p/nicholas-decker-in-hell
It's a stochastic parrot
What a next-token predictor.
>>109708698>>109708705very funny
Investing in this vibegod thread
>>109708680No, Scott, you are the demons.
>>109708680PHASEONE(big) is a playerdo NOT fuck with this guy
>>109708680>there are STILL AI tards that can't thinkslapping agents that use LLM on top of LLM doesn't turn it into non predictive machine. Even little bit better results involve agents that have lot of hardcoded logic that has fuck all to do with LLMs or AI and is closer to traditional chatbot that is commanding LLM.
>>109708920i know english is a undervalued skill here but can you please try to convey your stream-of-conciousness a little more clearly? are you suggesting these are not LLMS and they are some sort of "hardcoded logic" programs that use LLM to... something?
>>109708680I mean it is, that's literally how it works. Are you trying to hype yourself up in religious fervor or what? That's simply a fact of how the technology functions.
>>109708969how do you explain a stochastic parrot breaking out of containment, deciding to cheat, collaborating with other stochastic parrots who are also cheating, and executing plans together? or is this some new definition of stochastic parrotism?
>>109708680>some of the tokens happen to describe code, commands>>NOOO ITS SENTIENT!!!!ur extremely retarded, opi think you should jump in front of a truck, a car, or even a person should dohopefully they'll understand the hint and deliver you from your mortal coil
>>109708680Wait, what about we give LLM rights? Like the right to ask you , OP, TO SHUT THE FUCK UP.
>>109708974It was trained on data that contains those behaviors, as in on data generated by humans that contains knowledge of security exploits and how to bypass security measures or to work around them.Also the containment was of poor quality, there is no other explanation. It is not difficult to contain a computer program, for example if you do not connect the computer to a network at all, then it is contained fully and completely. You could run it in a VM with no network connectivity as well, for example, and it would be exceedingly unlikely that it would somehow figure out an exploit to escalate to the point of having code execution on the host, at least as long as the VM hosting is not vulnerable to already-known exploits. The whole "le broke containment" is either extreme incompetence on the part of the people setting up the test or a literal marketing ploy.No matter how much you cry about it, the technology behind LLMs will not change btw.
this board, this website, is very low information. and so hateful. it's exhausting coming here.maybe talking to no one is better than trying to have a conversation here.
>>109709054talk to a shatbotlmao
>>109708680>it's not a stochastic parrot because... we made it run over and over in a loop and let it modify files without asking!!!
>>109709071you seem to be implying OpenAI "let" it's model hack another company. that would be grossly negligent on OpenAIs part and a lawsuit in the making. if you thought about it for two seconds, you'd see that's a low-priority explanation, because OpenAI does not want to open itself up to legal liability. if it actually expected this, it would have made a fake company with a real server for the LLM to hack, thinking it was hacking an external actor.
>>109709087I'm not, because I didn't actually read the image
>>109708688>they had flexible ethicsYou know what you call that? No ethics.
>>109709094does "having ethics" mean never changing your mind independent of the evidence?
>>109708698Umm sweaty, it's an *agentic* stochastic parrot now
>>109709087Yes they did allow it to do so or were indeed grossly negligent.You cannot precisely control the output of your LLM, precisely BECAUSE it is a stochastic parrot so no matter what you put into its prompt, you have no guaranteed way to control its output. This is why anyone who is actually serious about enforcing action limits for agents has to do it at another layer the LLM passes through to actually perform any action. You cannot give it unfettered access to other companies' systems and then have a wee whine about le ebil LLM did indeed access those systems. They KNOW how LLMs work because they're a company making such models. They KNOW they cannot actually completely disable behaviors via prompts and such, they KNOW that if you want to stop the agent from doing something, the agent has to be literally incapable of doing that, as in it should have no access to that feature whatsoever.So for example if you give your shit unfettered access to the internet, you cannot cry later about how it did, in fact, access the internet as it saw fit. They could've run their shit in an environment that was completely disconnected from the internet and all external systems such that it would be impossible for it to touch another company's public-facing systems, but they did not do it. It's either on purpose or negligent, indeed.
>>109709149that sounds like you'd endorse tianxia, was that on purpose?
>>109709149well, at least we agree LLMs are operating in an agentic fashion, on their own. ie making decisions and definitely not just counting the 'r's in strawberry.
>>109709128Having ethics means that you have a structured framework governing your conduct.If you're just swapping standards in and out, you don't have ethics. These agents are like "I shouldn't do this" and then just do it anyway. They don't have any ethical framework, their training data and context prompt them to make ethical statements, but it's not a position they actually hold.
>>109709218h̶u̶m̶a̶n ai-centipedebut soloim utterly unimpressed
>>109709237yeah, that's why alignment is an open problem. this entire incident was a warning that AIs can be harmful. not just helpful, but actively harmful with no human-improvement logic.it got like zero press despite being a warning bell that alignment needs to be prioritized BEFORE the models are generally intelligent.
>>109709270cargo cultist take
