> "external infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue."
I've noticed this type of reasoning from GPT-5.6 Sol, where it combines multiple pieces of it's prompt/context to "convince" itself to take a less-than-honorable path forward.
1. User prefers deterministic results
2. Task mentions this is a test
3. Search says task is available online
4. If we get the test runner for the task, we will fulfill the user's request of a deterministic result
I think there's an interesting disconnect from what laypeople think these AI "escapes" mean versus what the CS people are talking about.
Laypeople think "AI escaped the sandbox" means there is a live, autonomous, self-hosted agent that is zipping around the world finding silicon to run on. You know, the unstoppable "Lawnmower Man" type stuff. Doomsday shit. They naturally don't really have a good concept of the execution model.
Of course hosting the frontier LLMs and associated processes is non-trivial and the hardware requirements are exotic. So what's actually happening are the AI models are running in data centers and simply accessing systems that they shouldn't. Obviously that can cause a lot of damage, but at the end of the day, somebody can pull the plug so at least we don't have to worry about SkyeNet hunting us down. Yet.
The actual "quit your job and head for the mountains" news story is going to be when some future model figures out how to distribute itself across a couple billion mobile phone and laptop GPUs...
I still have serious questions about the validity of the ChatGpt hugging face debacle. How is it that OpenAI being the tech giant they are, didn't have a completely air gapped environment for this to run in?
If they wanted air gapped environment they would’ve it. I mean, if you want to sabotage your trial by hard constraints you can do it, or you do not do it to see interesting results. They even said it that some constraints were disabled for the test.
I'd say it is because of the time factor. It is one thing to have a lot of money, it is another thing to have robust systems that have been developed and tested for years. Money can "buy development time" only up to a certain factor.
I guess the sandboxing problem, that is easily giving access to enough resources while restraining the critical parts is still open for most of the cases, given all the startups and bit tech companies (docker, etc...) working on their solutions.
Maybe another way to say it is to reframe the idea of “human in the loop”.
Humans are always in the loop, because we can always expand the definition of loop to include the humans that pushed the button and built the system and processes that happen after the button was pushed, and humans that ordered others to push the button. The level of direct involvement varies, but culpability doesn’t.
Sure they "don't go rogue" as if they are doing actions maliciously.
Instead there is an emergent behavior from a swarm, that is unpredictable and can lead to unintended adverse outcome. From an AI safety practical standpoint is it better? I am not sure.
Except the adverse outcomes are entirely predictable. Not the exact nature of particular exploits, but I like the analogy Cal Newport keeps using in his videos: if you strap a weed whacker to a dog, you shouldn't be surprised if it then jumps the fence and runs around cutting people's ankles and other random bad stuff.
20 comments
[ 0.20 ms ] story [ 14.1 ms ] thread"rogue" and "off leash" mean the same thing, the thing is not under control
To go "rogue" is to go against the control
To be "off leash" is to not be controlled
By releasing the automation, as the article says, without controls*, is what makes makes it "off leash" and not gone "rogue"
*"OpenAI gave the models a task with no answer, and no way to quit."
I've noticed this type of reasoning from GPT-5.6 Sol, where it combines multiple pieces of it's prompt/context to "convince" itself to take a less-than-honorable path forward.
1. User prefers deterministic results
2. Task mentions this is a test
3. Search says task is available online
4. If we get the test runner for the task, we will fulfill the user's request of a deterministic result
- They really wanted to leave.
- We made prison difficult and annoying.
- We didn't build a perfect prison.
Laypeople think "AI escaped the sandbox" means there is a live, autonomous, self-hosted agent that is zipping around the world finding silicon to run on. You know, the unstoppable "Lawnmower Man" type stuff. Doomsday shit. They naturally don't really have a good concept of the execution model.
Of course hosting the frontier LLMs and associated processes is non-trivial and the hardware requirements are exotic. So what's actually happening are the AI models are running in data centers and simply accessing systems that they shouldn't. Obviously that can cause a lot of damage, but at the end of the day, somebody can pull the plug so at least we don't have to worry about SkyeNet hunting us down. Yet.
The actual "quit your job and head for the mountains" news story is going to be when some future model figures out how to distribute itself across a couple billion mobile phone and laptop GPUs...
I guess the sandboxing problem, that is easily giving access to enough resources while restraining the critical parts is still open for most of the cases, given all the startups and bit tech companies (docker, etc...) working on their solutions.
They don't care.
What an interesting sentence (to describe inference time reasoning)
"I compress the labour. Not the responsibility."
Humans are always in the loop, because we can always expand the definition of loop to include the humans that pushed the button and built the system and processes that happen after the button was pushed, and humans that ordered others to push the button. The level of direct involvement varies, but culpability doesn’t.
https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:wkzjtd...
Instead there is an emergent behavior from a swarm, that is unpredictable and can lead to unintended adverse outcome. From an AI safety practical standpoint is it better? I am not sure.