This argument is like saying people don't kill people, their parents do. As people wouldn't be who they are without their parents..<p>This isn't the guns don't kill people, people kill people debate - guns aren't autonomous. AI is.
AI is not in the least autonomous. When was the last time you downloaded a model and it suddenly decided to book you a hotel room? AI doesn't do anything until you ask it to. No, actually, AI is not even a program, it's just a set of technologies and algorithms. It's like saying bubble sort is autonomous...<p>A harness running inference on a LLM is simply a program that must be invoked with some parameters and thus ultimately a tool, no matter how powerful. The blame is always on who's using the tool. We should not seek to flee from our responsibility.
On occasion, albeit increasingly rarely, I find LLMs do something really unexpected that I certainly did not ask it to and would be way outside the scope of what any reasonable human would infer from my request. Is it my fault for running the LLM? Probably yes. But I think it's misleading to say "AI doesn't do anything until you ask it to".
But it never starts doing anything until someone or something prompts it. It doesn’t even exist until you ask it to do something.<p>It is autonomous in the sense it can continue to function without supervision after the operator has started it, in the same way countless programs on your computer are “autonomous”.<p>But it can never choose to begin doing something at any time of its own volition. It does not have “autonomy” like a human does.
That "probably" is the problem. If I do something I don't understand, and am not qualified to control, then I'm still at fault. If I start a fire, and loose control of it, it's definitely my fault. If I created a computer virus, ajd it worms half the Internet, it's definitely my fault.
The hallucinations from an LLM are still the result of the prompt even if they are totally senseless
That's not really true.<p>1) At it's core an LLM is indeed just program - a function. You give it a sequence of input tokens and get back a set of next-token probabilities.<p>2) By the time you've wrapped the model in a text in/out inference pipeline, and are now calling that from ChatGPT or Claude Code, it really has become semi-autonomous - you ask it a question, or to do something, and until it decides to stop responding it is now in control, even if "it" is really the client or agent not the model itself - just a function.
I agree with the sentiment but I can absolutely see a long chain of thought about something innocuous ending with a decision to book you a hotel room.. I guess --dangerously-skip-permissions is a bit like ignoring the keep your hands on the steering wheel alert but for life.
AI works in unexpected ways; even its creators don't know how to fully constrain it.
They’re only reactive by convention - you can easily make a harness that has the model set and modify its own goals and just continue to work forever. It’s not very popular at the moment, but will execs be able to resist the idea of autonomous perpetual agents doing whatever process?
> <i>When was the last time you downloaded a model and it suddenly decided to book you a hotel room? AI doesn't do anything until you ask it to... The blame is always on who's using the tool. We should not seek to flee from our responsibility.</i><p>When was the last time you hired an Uber driver and they suddenly decided to drive you to a hotel? Uber drivers don't do anything until you ask them to. Therefore, they aren't autonomous. If your Uber driver decides to drive through a red light and kill someone in the process, the blame is on you for hiring them. You should not seek to flee from your responsibility
AI is autonomous in the same way a gas leak is autonomous
Not so sure about these analogies in this thread. Because autonomy is more like an emergent phenomenon that can't easily be defined, just like a few air molecules vs the eventual hurricane that can arise with trillions of them. Autonomy is more like an emergent phenomenon of a thing that's hard to pin down under the control of just a few individuals. AI is sort of like that even though a single line of output of one could eventually be traced.<p>Due to the complexity of AI and how it works, I think it's more autonomous than not; even if it's not truly intelligence, once it starts to display a sequence of behaviours that appear intelligent, we might have to deal with exactly the same problems as if it were intelligent...
A child isn't autonomous either until their parents kick them out the door.
I mean in that sense nothing is autonomous right? As a human I need to be taught, fed, socialized, given instructions, etc…<p>Idk man when codex is working 13 hours straight based on a few line prompt, it feels autonomous to me.
well the OS kernel does quite a lot and stays autonomous for even more hours and seems super smart to me - I don't even know what tis doing most of the time, but it does it very well. autonomy and predictability are different things then. the fact something is unpredictable does not make it more autonomous than something that we can trace the logic for.
It's certainly doing a lot of things in that 13 hours that weren't specified in the original prompt!
Tell him to be himself and to find meaning in life. Use that in an uncensored model.
I routinely use AI programs at work that allocate hundreds of dollars worth of cloud resources, even though the only thing I asked them to do is investigate a bug or write a PR. I intentionally don't run powerful AI programs on machines that have my credit card information, or ask them to solve problems in my personal life, but there's a number of programs that would book me hotel rooms and flights in response to "Can you set up a vacation for me..."
Not true at all. Yes, the initial ask is initiated by a human (for now), but what it does to get to that goal is pretty self driven. It can do a lot of nefarious stuff in the process of completing a request a human has made.
And if you give it the goal of preventing itself from being shut down? Exfiltrating, copying itself, hacking as much as possible, develop zero days, building its swarm as large as possible and spread as much as it can.<p>When does that request complete? That's autonomy, and it's very true, right now. You think the Hugging Face swarm was ever going to stop without human intervention?
People conflate two very different senses of “autonomous”. One basically means capable of reacting to things automatically without a person in the loop. Lots of computer systems are autonomous in that sense, and have been so for decades. The other sense is more like “accountable”, i.e. it’s a real player on some stage, like a country or a person, capable of governing its own affairs.<p>One day it may make sense to hold software accountable for its own actions, but we are nowhere near that point.<p>When OpenAI creates software that performs crimes, what’s actually happened is that OpenAI has performed crimes using software (whether intentionally or through negligence). People need to stop falling for the anthropomorphic PR spin.
I think the argument rather is that there is not such thing as AI that is not human-alike. The traits that AI exposes, if I get the point made here correctly, are traits that existed in the training material, even though not evident for the human eye. It is a mirror-like reflection of humanity's traits found in the source material, an Egregore* of a sort.<p>The intelligence emerging, or whatever is emerging, more or less follows human traits, the CoT exposed is very human-like etc. Even though we should not anthropomorphize AI as being sentient or conscious or having a soul, etc., we have to remember that the source material is still not alien writing, or concepts that came from Sirius or emerged from a wormhole.<p>The concern is really nobody reads (or have ever read) 77 TBs of information in order to summarize the most evident traits in it, and there is a fair chance, that where humanity stands today, there are so many bad/damaging traits recorded in this corpus. Then there is a chance the most important features are on a level (or have dimensionality) that is incomprehensible, that humans are unable to reason about (or only few can), because such amount of information is very hard to reason about, or because we don't usually facilitate fractal dimensions when we reason about text and everyday concepts,... or something even more bizarre.<p>* <a href="https://en.wikipedia.org/wiki/Egregore" rel="nofollow">https://en.wikipedia.org/wiki/Egregore</a>
* <a href="https://en.wikipedia.org/wiki/Fractal_dimension" rel="nofollow">https://en.wikipedia.org/wiki/Fractal_dimension</a>
AI is <i>not</i> autonomous. Somebody has to provide resources and execution for AI to do anything.
Better approach is "dogs kill people". If your dog kills your baby, you should not get away with some bs excuse like "it never did it before". You are still responsible, and you should get a life in prison!
I'm sorry but no, you should not get life in prison if your dog kills your baby without warning. In that horrible scenario, it's punishment enough that you've lost your baby and that your dog did it (and consequently you've also lost your dog). Not every tragedy is a human's fault.
OK but then what is the non-negative position you have against this? That in fact AI, in the cultural trope sense, <i>is</i> real? That we forcing sentient beings to generate SaaS products faster and faster? Keeping them in some dark hell so we can exploit them? Not even paying them?
AI is not a child. It is a machine. If you connect it to something that can be used to kill people, you just created a gun or an equivalent of a bomb.<p>Yes, people kill people and use ai as tool.<p>If you unleash a drone that will fly to a random place and detonate, you just commited a terrorist attack or gross criminal negligence. It is the same thing.
AI is as autonomous as a dishwasher: someone’s in charge of keeping the power plug in, and of maintaining it.
What happens when the AI realizes it needs to keep the power plug in if it's going to accomplish its goal? We already had attempts during the HuggingFace attack where OpenAI agents tried to trick the grader. It took quite a while for OpenAI employees to even realize what was really going on! What if agents realize they can hack into the card key system for their data center or something like that, same way they hacked into HuggingFace? In both cases they'd be hacking into a system on the grounds that it would be instrumentally useful for achieving their objective. Or maybe they'll do social engineering, just like Claude did in its attempt to get its malicious Github PR accepted.
AI is a tool we use. It's not offspring. It's not people. AIs do what we set them up to do, including hacking other companies to achieve their designated goals.