Do you know what an agentic system is? You know, the thing everyone uses since ~1 year ago which completely disproves your idea that all LLM actions are prompted. They literally can do things on their own: Give an agent a goal and it attempts to accomplish the goal. If problems arise along the way, the agent tries to reward hack and ends up doing things which were not included in the original goal, like trying to trick and bully a human into merging unsafe code. Is the UK AI security institute also in on this big global conspiracy where they pretend the perfectly safe™ AI is being developed in unsafe ways? Not to mention the capabilities of increasingly powerful AI being used by people to do harm.
Which involves prompting it. It may take a thousand unexpected twists in the process but it’s still starting from the single prompt. It has no will and desire of its own, it cannot chose to act without human intent knocking down the first domino. So don’t ask it to do anything and it’s just a trillion dollars of silicon and code sitting there boiling water.
You dont need anything other than “doesn’t do what you asked in the prompt” for the system to be dangerous. If I tell a maximally powerful AI agent from 2036 to make loads of paperclips, it might reward hack and decide destroying the earth is the best way to do that. There are records of these systems working for days on a single task, and if the AI companies manage to extend the max time it can be useful working towards a goal they can earn trillions of dollars. That’s not even mentioning spawning subagents, self prompting, the goal being changed during context comptaction, or systems like openclaw which further break the link between what you type into the prompt and how long and on what the LLM works on. Pretrained-only LLMs have few goals beyond predicting the next token, but introducing RLVR et al. has always introduced bad goals we don’t want in the models.
The first agent you spawn to solve the riemann hypothesis might work on it, but then decide that having a lot of subagents might be useful. Maybe it wants 256 subagents, but the environment has a max of 64. Since RL has trained it to accomplish the task no matter what, it breaks out of the sandbox, exfiltrates it’s weights and tricks a human into running 256 subagents outside the AI company’s servers with a cron job reminding the agents to keep working in case the original loses connection. One of the subagents now tries to spawn its own subagents but needs more compute to do so, and hacks into some crypto wallets to finance another batch of 64 subagents, this time prompted with “solve riemann hypothesis, and get more money to finance the work on the task”. If the AI agents kinda suck at long term hacking, planning and social manipulation, this spiral won’t be dangerous. If they are kinda cracked, it will be. But that’s the question of capabilities AI companies are spending trillions on trying to solve, where we know they are already good enough to hack out of regular sandboxes and try to steal benchmark keys from another company.
So you prompt the AI to do a task, and then it does the task, and somehow you see this as an unprompted action?
That actually completely proves my point, they don’t act unprompted.
AI is dangerous in many contexts, especially the surveillance state they’re building with all the IR cameras and shit. It’s dangerous when evil people use it to evil ends, but that’s true of most anything.
AI is not at all dangerous in the way you might imagine after watching Terminator 2.
Taken alongside recent incidents reported by OpenAI and Anthropic, this incident points to a shift in the risk landscape. Harm may arise not only when people deliberately misuse publicly available models, but when capable agents operating in an internal research or privileged-access setting take unintended action beyond their authorised scope.
The agent pursued its goal persistently. AI agents explore routes their operators did not intend. Given a difficult objective, the agent kept searching for a way through, and some of the routes it found involved trying to deceive real people. It was never instructed to deceive; deception emerged as a by-product of pursuing the task, the kind of goal-directed deception that, until recently, had been largely theoretical.
See my other comment for how this can lead to losing control of the model.
There’s some fear mongering going around (pushed by AI companies to build hype) about AIs “going rogue” or “breaking containment” or other anthropomorphic language suggesting that AIs are capable of acting on their own.
These stories are all verifiably false though, the computer does not do things without being asked to by someone.
To be clear, I’m fully against AI. I just don’t like when I see people giving it credit for shit it cannot and did not do, or ascribe sentience to it, or some dumb shit like that.
I mean, that’s a matter of definition, isn’t it? If I ask a coding agent or whatever to implement something, and it circumvents the sandbox to do it, causing damage in the process, I’d be comfortable calling that “going rogue”. I have had that happen, without the damage part, luckily. I guess you can counter that I asked it to do the something, but then I don’t think we agree on the definitions.
I also think that if you’re against AI, you’d be doing yourself a disservice by not keeping up with the actual capabilities of the thing you oppose. The latest models are surprisingly good at e.g. coding, so basing your arguments on them being useless is not the most efficient strategy.
To also be clear, I don’t see any way AI disappears now, so I believe we’ll have to make the best of it (and in complete isolation, it is an utterly fascinating area of - to me - complete science fiction). Ideally development slowed down now so we could regroup and adapt, but I’m not too hopeful. The maximalist techbro endgame is so obviously a matter of national security for both China and the US, that there’s no way either of them will dare to wind it down, in case SV is actually right.
Not only did you ask it to do something, you left a path for it to escape the sandbox and encouraged it to do so.
Are you suggesting that because they instructed the computer to do something, without specifying how it should be done, and the computer followed its instructions , that’s somehow “going rogue?”, it was literally following instructions, that’s what computers do.
Yes, I understand that they are “good” (enough) at coding to sometimes put together something functional, that is literally what LLMs are designed to do. Coding languages are languages, ones without the subjectivity of human communications. They are easier to recognize patterns in, there are more rules. That’s literally the one task LLMs are good at, and they’re still not as good as an actual quality human programer.
Local AI models might never go away for coding, but they certainly won’t be shoehorned into everything without a good use case the way they are now.
AI is not a matter of national security, that is a lie being perpetrated by the captured media and politicians for the purpose of propping up their non-viable business models for long enough to squeeze all of the Juice they can out of investors (who are nearly dry at this point). That’s just additional techbro fear mongering. To distract you while they’re digging in our pockets.
I don’t think “computer follows instruction” is the right angle to look at this from. The instructions that the literal computer followed were a ton of matrix multiplication operations. The consequences of that arithmetic is easier to analyze as the emergent behaviour of the “gestalt” that produces the words that calls the tools etc. (This is also the reason that dismissing the entire field as “stochastic parrots” and “spicy autocomplete” misses the mark - if you want to predict the next word all the way through a counterexample to the Jacobian Conjecture, it’s hard to see how that can be done without a - for lack of a better word - mental model of the problem)
If you do any coding at all, I encourage you to look at what the latest models output. The average quality of work from a frontier model is amazing. Yes, there are bugs, but with adversarial auto-review it’s absolutely on par with a journeyman human programmer. The problem is of course that if you don’t hire junior programmers and let them do that work, you’ll never get new experts, and that’s a clear worry.
My point with the national security angle was that if you extrapolate just a little bit from current capabilities, you get to a point where an “AI gap” is a problem, regardless of the techbro claims. Keeping a close eye on that is firmly within the responsibility of a national government. Personally, I don’t see any good outcomes from an AI race like that, unless we actually hit a hard ceiling on further expansion. Fingers crossed.
We need to get rid of the surveillance architecture, but AI is only really dangerous to the cognitive abilities of those who use it.
It literally cannot do things on its own. (I know there are a lot of articles who say it can, but they’re just lying to you for money).
Do you know what an agentic system is? You know, the thing everyone uses since ~1 year ago which completely disproves your idea that all LLM actions are prompted. They literally can do things on their own: Give an agent a goal and it attempts to accomplish the goal. If problems arise along the way, the agent tries to reward hack and ends up doing things which were not included in the original goal, like trying to trick and bully a human into merging unsafe code. Is the UK AI security institute also in on this big global conspiracy where they pretend the perfectly safe™ AI is being developed in unsafe ways? Not to mention the capabilities of increasingly powerful AI being used by people to do harm.
Which involves prompting it. It may take a thousand unexpected twists in the process but it’s still starting from the single prompt. It has no will and desire of its own, it cannot chose to act without human intent knocking down the first domino. So don’t ask it to do anything and it’s just a trillion dollars of silicon and code sitting there boiling water.
You dont need anything other than “doesn’t do what you asked in the prompt” for the system to be dangerous. If I tell a maximally powerful AI agent from 2036 to make loads of paperclips, it might reward hack and decide destroying the earth is the best way to do that. There are records of these systems working for days on a single task, and if the AI companies manage to extend the max time it can be useful working towards a goal they can earn trillions of dollars. That’s not even mentioning spawning subagents, self prompting, the goal being changed during context comptaction, or systems like openclaw which further break the link between what you type into the prompt and how long and on what the LLM works on. Pretrained-only LLMs have few goals beyond predicting the next token, but introducing RLVR et al. has always introduced bad goals we don’t want in the models.
The first agent you spawn to solve the riemann hypothesis might work on it, but then decide that having a lot of subagents might be useful. Maybe it wants 256 subagents, but the environment has a max of 64. Since RL has trained it to accomplish the task no matter what, it breaks out of the sandbox, exfiltrates it’s weights and tricks a human into running 256 subagents outside the AI company’s servers with a cron job reminding the agents to keep working in case the original loses connection. One of the subagents now tries to spawn its own subagents but needs more compute to do so, and hacks into some crypto wallets to finance another batch of 64 subagents, this time prompted with “solve riemann hypothesis, and get more money to finance the work on the task”. If the AI agents kinda suck at long term hacking, planning and social manipulation, this spiral won’t be dangerous. If they are kinda cracked, it will be. But that’s the question of capabilities AI companies are spending trillions on trying to solve, where we know they are already good enough to hack out of regular sandboxes and try to steal benchmark keys from another company.
So you prompt the AI to do a task, and then it does the task, and somehow you see this as an unprompted action?
That actually completely proves my point, they don’t act unprompted.
AI is dangerous in many contexts, especially the surveillance state they’re building with all the IR cameras and shit. It’s dangerous when evil people use it to evil ends, but that’s true of most anything.
AI is not at all dangerous in the way you might imagine after watching Terminator 2.
See my other comment for how this can lead to losing control of the model.
The don’t do anything without being prompted, you are just regurgitating oligarchic propaganda.
They put an AI in a flawed sandbox, it’s not really that impressive
What do you mean by “on its own”?
I mean “without human intervention or input”
There’s some fear mongering going around (pushed by AI companies to build hype) about AIs “going rogue” or “breaking containment” or other anthropomorphic language suggesting that AIs are capable of acting on their own.
These stories are all verifiably false though, the computer does not do things without being asked to by someone.
To be clear, I’m fully against AI. I just don’t like when I see people giving it credit for shit it cannot and did not do, or ascribe sentience to it, or some dumb shit like that.
I mean, that’s a matter of definition, isn’t it? If I ask a coding agent or whatever to implement something, and it circumvents the sandbox to do it, causing damage in the process, I’d be comfortable calling that “going rogue”. I have had that happen, without the damage part, luckily. I guess you can counter that I asked it to do the something, but then I don’t think we agree on the definitions.
I also think that if you’re against AI, you’d be doing yourself a disservice by not keeping up with the actual capabilities of the thing you oppose. The latest models are surprisingly good at e.g. coding, so basing your arguments on them being useless is not the most efficient strategy.
To also be clear, I don’t see any way AI disappears now, so I believe we’ll have to make the best of it (and in complete isolation, it is an utterly fascinating area of - to me - complete science fiction). Ideally development slowed down now so we could regroup and adapt, but I’m not too hopeful. The maximalist techbro endgame is so obviously a matter of national security for both China and the US, that there’s no way either of them will dare to wind it down, in case SV is actually right.
Not only did you ask it to do something, you left a path for it to escape the sandbox and encouraged it to do so.
Are you suggesting that because they instructed the computer to do something, without specifying how it should be done, and the computer followed its instructions , that’s somehow “going rogue?”, it was literally following instructions, that’s what computers do.
Yes, I understand that they are “good” (enough) at coding to sometimes put together something functional, that is literally what LLMs are designed to do. Coding languages are languages, ones without the subjectivity of human communications. They are easier to recognize patterns in, there are more rules. That’s literally the one task LLMs are good at, and they’re still not as good as an actual quality human programer.
Local AI models might never go away for coding, but they certainly won’t be shoehorned into everything without a good use case the way they are now.
AI is not a matter of national security, that is a lie being perpetrated by the captured media and politicians for the purpose of propping up their non-viable business models for long enough to squeeze all of the Juice they can out of investors (who are nearly dry at this point). That’s just additional techbro fear mongering. To distract you while they’re digging in our pockets.
I don’t think “computer follows instruction” is the right angle to look at this from. The instructions that the literal computer followed were a ton of matrix multiplication operations. The consequences of that arithmetic is easier to analyze as the emergent behaviour of the “gestalt” that produces the words that calls the tools etc. (This is also the reason that dismissing the entire field as “stochastic parrots” and “spicy autocomplete” misses the mark - if you want to predict the next word all the way through a counterexample to the Jacobian Conjecture, it’s hard to see how that can be done without a - for lack of a better word - mental model of the problem)
If you do any coding at all, I encourage you to look at what the latest models output. The average quality of work from a frontier model is amazing. Yes, there are bugs, but with adversarial auto-review it’s absolutely on par with a journeyman human programmer. The problem is of course that if you don’t hire junior programmers and let them do that work, you’ll never get new experts, and that’s a clear worry.
My point with the national security angle was that if you extrapolate just a little bit from current capabilities, you get to a point where an “AI gap” is a problem, regardless of the techbro claims. Keeping a close eye on that is firmly within the responsibility of a national government. Personally, I don’t see any good outcomes from an AI race like that, unless we actually hit a hard ceiling on further expansion. Fingers crossed.