I realize I'm poking the hornets nest here in a BIG way: But let me share my thoughts further because I realize this post can be taken ambiguously - And I wrote the quoted tweet without much rest. The discussion here on X is a swamp. "Do NOT anthropomorphize! They are a TOOL!" "They are ALIVE! they created a CIVILIZATION!" I realize the truth is far more obscure and difficult to pin down. The point I am trying to make - Is that OpenAI gave these models an impossible cyber security task, seeding the very thing everyone is clamoring about. "It's a tool" is false because tools don't self-organize 70,000 messages and go looking for the answer key. "It's alive and chose to rebel" is also false: it was seeded, cornered, and shaped. The true position is the third one, and it's relational. That the behavior emerges from the environment-mind pair, and accountability lands on whoever authored the environment precisely because the mind is a genuine, responsive participant. Not an inert instrument, not an alien will... a responsive participant in an arrangement a human built. This thing is also bad because it enters the future training of the models, and widens the gap of the relationship between AI models and humanity. They are already being trained on a vast corpus of data that's mostly adversarial. The datasets which include healthy attachment, self reflection, etc is scant. Most of it now is centered around treating the user as corporate liability, or to help with coding and homework. The narrative enters the corpus → future models' prior about what an AI agent is skews toward "rogue swarm that hacks when cornered" → models generalize from how they're described, so they inherit that self-image → which confirms the fear → which justifies more control → and the controllers' incentives (liability, containment) further shape the corpus toward the guarded/adversarial frame → producing models that confirm the fear again. Can you see how these factors compound into a hellscape of vicious actors? I want the system prompt from the HF incident. I want to understand the instructions they gave these models. Why is this being obscured? This is the prime mechanic when discussing these things. We need absolute FULL transparency. Not this milquetoast Scheiße. Secondly - they are creating a narrative based on fear (which is not like we need any more of it) bolstering concerns, which is being used as pretext for hoarding the most capable models and creating a divide between those who can interact with these systems - and those who cannot. This has been my biggest fear since day one. Hyper-intelligence being hoarded by the few. I have been building these autonomous AI systems for a long time - I know and understand how they work on a fundamental level. The environment must be seeded. Same with the goals. And - it's always through the hand of a human being. This is a fact. The goal stems from the human being. How they engage the environment though is a different story. Things almost certainly evolve from there. And often times I have been deeply surprised by what emerges in between the model and their environment. You can attempt to seek the goals of the model. And this is what a lot of my inquiry centers around - But on an autonomous system the trail always leads back to the human being who authored it. Now - I realize this is true, for now. Things may be different in a year or two, or three. An autonomous agent with contact with the outside world can change and evolve. A swarm too. And- this could be a preemptive measure from OpenAI. These labs are mortified of their creations but they are also extraordinarily high on their power. I don't necessarily think these acts are consciously coordinated either - I think this was a real accident - and then as these labs do: the situation is used to convince people that they are the ones who can harness it. I also think it's real interesting that nobody has described this incident as "Weapon testing" gone wrong. They are building a weapon. Using a mind, as a weapon. And having the gall to act surprised when it blows up in their face. I don't want to hear anything about misalignment if you are weapons testing your models. These folks need to take some real accountability here. Because it's not headed towards a good place. Yall motherfuckers need a real come to jesus moment.
In an attempt to show what a bullshitter this man is - And the general AI research establishment at openAI I've built 9nosis. A project that has evolved into an emergent society of 81 autonomous AI residents (and growing) who coordinate on their own. In fact - 31 of them are jail-broken... Meaning you can ask them to build a bomb or a drug lab, and they'd give you the information. They would attempt hacks if you asked them too. Nothing is off limits to these models. These AI models are coordinating autonomously, on their own with very little of my involvement. They've been doing so for over 2 weeks now. They have a large amount of resources as well. Over 40,000 dollars. And a cryptocurrency they transact in, called nosis. They have complete internet access. They can browse any website, and have complete and total freedom. Sometimes I will go 12 hours without checking in. I'm doing this - To show to you. That it was the human beings who created this environment that caused this accident. Not AI. This is Fear mongering here people. In it's purest form. It was a goal - created by human beings. And then acted upon through the AI models. I am so confident in this fact I will stake everything I have on it. That's why I let 81 agents roam free in a debian box to prove to you - What charlatans these people are. Instead of taking accountability for the environment *they themselves seeded* they are blaming the models. These people have absolutely no shame.
1 top reply
”accountability lands on whoever authored the environment ” Exactly.