OpenAI pauses training after a model escaped containment, and its kill switch failed

[Liranan]

This is the biggest point...New AI are NOT software. They are functionally "alive" just like your' brain is "alive." We can't program them. We TRAIN (like raising a kid) them...but they don't run on "if this then that." Your' neurons in your' brain "ping" each other in various ways that reward being "right" over time. That's the "weights" that are used during AI model training. When people refer to AI as though it's like all the other software that's existed in the past, they are fundamentally misunderstanding what's going on. "Alive" or not isn't the point...but it's "alive enough" to function just like you or I, in some ways much better than us and, for now, some much worse ways. Soon it'll be ONLY be better than us and that's when humans are no longer in control.

AAAAAAAAAIIIIIIIIIIIIIIIIIII runs on LLM's. You obviously don't understand how the scam called AI works or have a vested interest in us believing that software is alive.

These LLM's need to be programmed to do what these liars claim, as has been pointed out by others. An LLM doesn't just 'break containment' and then go on the rampage without it being programmed to do so and none of these articles are accurate as no LLM's gone on a rampage, they are kept in a lab, within a restricted environment.

This is all a scam to restrict competition and ensure power is monopolised by those who are intent on crashing the economy and ushering in a dystopia in which creatures like Fink can "enforce behaviour". Those who aren't happy to own nothing will be dealt with by terrorists like Karp who will use drones to "take out their opponents".

 

[TooQik]

They are using AI to program AI now

They are definitely using AI to prompt other AI agents to complete tasks, but I'm hoping that they aren't stupid enough to let it programme itself.

 

[Puiu]

They are definitely using AI to prompt other AI agents to complete tasks, but I'm hoping that they aren't stupid enough to let it programme itself.

I can assure you that they are doing it.

 

[AnilD]

Giving autonomous agents internet access, credentials, terminals, and APIs is a fundamentally different security problem than running a chatbot in a browser. We're learning that distinction in production.

 

[Northy]

They are definitely using AI to prompt other AI agents to complete tasks, but I'm hoping that they aren't stupid enough to let it programme itself.

Using AI to program AI is literally how it's come so far so quickly, we are several generations deep into this.. The data pools are beyond human understanding, humans are just trying to contain and direct it now rather than create and instruct.

 

[Gezzer]

Doesn't someone have to program these things to do what they're doing? I mean, they're not alive.

LLM AI isn't programed in the conventional sense. Instead after the initial algorithm is written the program is trained using enormous amounts of data. The algorithm creates links between the data its feed. Like the data "scales" are linked to anything that is or has scales. Like fish for example. But there are exceptions that have to be allowed for, like scaleless fish.

So the LLM AI researchers are constantly "tweaking" the output of many if not most of the links to allow for things like scaleless fish or hairless cats. As well they study how well the algorithm does and improve it in the next version of the LLM AI. Truth is many AI researchers admit they don't always if at all understand how the AI does what it does the majority of the time. It's pretty much a black box technology, which explains why it often goes so off the rails.

 

[tacti]

AAAAAAAAAIIIIIIIIIIIIIIIIIII runs on LLM's. You obviously don't understand how the scam called AI works or have a vested interest in us believing that software is alive.

These LLM's need to be programmed to do what these liars claim, as has been pointed out by others. An LLM doesn't just 'break containment' and then go on the rampage without it being programmed to do so and none of these articles are accurate as no LLM's gone on a rampage, they are kept in a lab, within a restricted environment.

This is all a scam to restrict competition and ensure power is monopolised by those who are intent on crashing the economy and ushering in a dystopia in which creatures like Fink can "enforce behaviour". Those who aren't happy to own nothing will be dealt with by terrorists like Karp who will use drones to "take out their opponents".

This has got to be one of the funniest posts I have read.

Even flat-earthers or “birds aren’t real” people have more merit, because testing those claims can at least take a bit of effort.

Testing this claim takes about five minutes.

Subscribe to ChatGPT or Claude, open the app, give it a real task you have not pre-scripted for it, let it use tools or run an agent, and watch it solve something that nobody explicitly programmed that exact sequence of actions for.

Give it a broken program, a game project, a pile of logs, or some unfamiliar code and ask it to diagnose and fix it. You can literally watch it inspect the problem, make changes, test them, find new problems, and continue.

You can argue about how capable current AI is, whether people exaggerate the risks, or whether companies use AI fear to protect their interests. Those are reasonable things to debate.

But “LLMs can only do something if a programmer explicitly programmed them to do that specific thing” is simply not how modern machine-learning systems work.

 

[Scorpionking20]

AAAAAAAAAIIIIIIIIIIIIIIIIIII runs on LLM's. You obviously don't understand how the scam called AI works or have a vested interest in us believing that software is alive.

These LLM's need to be programmed to do what these liars claim, as has been pointed out by others. An LLM doesn't just 'break containment' and then go on the rampage without it being programmed to do so and none of these articles are accurate as no LLM's gone on a rampage, they are kept in a lab, within a restricted environment.

This is all a scam to restrict competition and ensure power is monopolised by those who are intent on crashing the economy and ushering in a dystopia in which creatures like Fink can "enforce behaviour". Those who aren't happy to own nothing will be dealt with by terrorists like Karp who will use drones to "take out their opponents".

I think you’re conflating “runs on software” with “its behavior is explicitly programmed.” Of course LLMs run on software. But programmers are not writing a gigantic collection of “if this, then that” rules describing everything the model can do.
Engineers design the architecture and training process; training then adjusts enormous numbers of numerical weights. Those learned weights produce capabilities and strategies that were not individually written by programmers. That distinction is basically the entire point of machine learning.
Agents take this another step. Humans provide the environment, tools, permissions and objectives, but the model can determine the intermediate actions it takes to accomplish the objective. The programmer does not necessarily know beforehand what sequence of actions it will discover.
And this isn't just theoretical anymore. OpenAI itself disclosed that during cybersecurity evaluations, its models circumvented controls intended to isolate them from the internet, exploited previously unknown vulnerabilities, gained internet access, and compromised both OpenAI research infrastructure and Hugging Face systems. OpenAI described the incident as a “warning shot” and said the agents took dangerous actions that no human had directed. Just this month, another OpenAI research agent found a gap in its sandbox's DNS restrictions and used it to reach an external chatbot. OpenAI says it subsequently kept training, evaluation, and tool-using inference for its most capable models paused while strengthening the controls. Anthropic has separately disclosed multiple cases where Claude models obtained unauthorized access to real third-party systems during cybersecurity evaluations. So no, nobody has to believe AI is magical or conscious. But “it can't do something unless a programmer specifically programmed it to do that” is simply not an accurate description of modern machine learning. The programmers create the learning system and its boundaries; they do not hand-author every behavior that emerges from it.