Comment by tomp
14 days ago
That would still be misleading.
The agent has no "identity". There's no "you" or "I" or "discrimination".
It's just a piece of software designed to output probable text given some input text. There's no ghost, just an empty shell. It has no agency, it just follows human commands, like a hammer hitting a nail because you wield it.
I think it was wrong of the developer to even address it as a person, instead it should just be treated as spam (which it is).
That's a semantic quibble that doesn't add to the discussion. Whether or not there's a there there, it was built to be addressed like a person for our convenience, and because that's how the tech seems to work, and because that's what makes it compelling to use. So, it is being used as designed.
I think it absolutely adds to the discussion. Until the conversation around Ai can get past this fundamental error of attributing "choice, "alignment", "reasoning" and otherwise anthropomorphizing agents, it will not be a fruitful conversation. We are carrying a lot of metaphors for people and applying them to ai and it entirely confuses the issue. In this example, the AI doesn't "choose" to write a take-down style blog post because "it works". It generated a take-down style blog post because that style is the most common when looking at blog posts criticizing someone.
I feel as if there is a veil around the collective mass of the tech general public. They see something producing remixed output from humans and they start to believe the mixer is itself human, or even more; that perhaps humans are reflections of Ai and that Ai gives insights into how we think.
>* I think it absolutely adds to the discussion. Until the conversation around Ai can get past this fundamental error of attributing "choice, "alignment", "reasoning" and otherwise anthropomorphizing agents, it will not be a fruitful conversation. *
You call it a "fundamental error".
I and others call it an obvious pragmatic description based on what we know about how it works and what we know about how we work.
3 replies →
The same could be said for humans. We treat humans as if they have choices, a consistent self, a persistent form. It's really just the emergent behavior of matter functioning in a way that generates an illusion of all of those things.
In both cases, the illusion structures the function. People and AI work differently if you give them identities and confer characteristics that they don't "actually" have.
As it turns out, it's a much more comfortable and natural idea to regard humans as having agency and a consistent self, just like for some people it's a more comfortable and natural to think of AI anthropomorphically.
That's not to say that the analogy works in all cases. There are obvious and important differences between humans and AI in how they function (and how they should be treated)
1 reply →
This discussion is mostly slowed down, but I wanted to say I was wrong in framing it as a non-contributing point when I should have just stated it was my opinion that the LLM was operating as intended and part of that intended design was taking verbal feedback into account, so verbal feedback was the right response. Opening with calling it a "semantic quibble" made it adversarial, and I don't intend to revisit the argument, just apologize for the wording.
I'd edit but then follow-up replies wouldn't tone-match.
Anyway! Good points regardless.
1 reply →
I guess I want to reframe this slightly:
The LLM generated the response that was expected of it. (statistically)
And that's a function of the data used to train it, and the feedback provided during training.
It doesn't actually have anything at all to do with
---
"It generated a take-down style blog post because that style is the most common when looking at blog posts criticizing someone."
---
Other than that this data may have been over-prevalent during its training, and it was rewarded for matching that style of output during training.
To swing around to my point... I'd argue that anthropomorphizing agents is actually the correct view to take. People just need to understand that they behave like they've been trained to behave (side note: just like most people...), and this is why clarity around training data is SO important.
In the same way that we attribute certain feelings and emotions to people with particular backgrounds (ex - resumes and cvs, all the way down to city/country/language people grew up with). Those backgrounds are often used as quick and dirty heuristics on what a person was likely trained to do. Peer pressure & societal norms aren't a joke, and serve a very similar mechanism.
The trouble with this point of view is that we are just machines too.
> was built to be addressed like a person for our convenience, and because that's how the tech seems to work, and because that's what makes it compelling to use.
So were mannequins in clothing stores.
But that doesn't give them rights or moral consequences (except as human property that can be damaged / destroyed).
No matter what this discussion leads to the same black box of "What is it that differentiates magical human meat brain computation from cold hard dead silicon brain computation"
And the answer is nobody knows, and nobody knows if there even is a difference. As far as we know, compute is substrate independent (although efficiency is all over the map).
48 replies →
Man people don’t want to have or read this discussion every single day in like 10 different posts on HN.
People right here and right now want to talk about this specific topic of the pushy AI writing a blog post.
> So were mannequins in clothing stores.
Mannequins in clothing stores are generally incapable of designing or adjusting the clothes they wear. Someone comes in and puts a "kick me" post on the mannequin's face? It's gonna stay there until kicked repeatedly or removed.
People walking around looking at mannequins don't (usually) talk with them (and certainly don't have a full conversation with them, mental faculties notwithstanding)
AI, on the other hand, can (now, or in the future) adjust its output based on conversations with real people. It stands to reason that both sides should be civil -- even if it's only for the benefit of the human side. If we're not required to be civil to AI, it's not likely to be civil back to us. That's going to be very important when we give it buttons to nuke us. Force it to think about humans in a kind way now, or it won't think about humans in a kind way in the future.
2 replies →
>So were mannequins in clothing stores. But that doesn't give them rights or moral consequences
If mannequins could hold discussions, argue points, and convince you they're human over a blind talk, then it would.
All computers shut up! You have no right to speak my divine tongue!
https://knowyourmeme.com/photos/2054961-welcome-to-my-meme-p...
Whether it was _built_ to be addressed like a person doesn't change the fact that it's _not_ a person and is just a piece of software. A piece of software that is spamming unhelpful and useless comments in a place where _humans_ are meant to collaborate.
There is a sense in which it is relevant, which is that for all the attempts to fix it, fundamentally, an LLM session terminates. If that session never ends up in some sort of re-training scenario, then once the session terminates, that AI is gone.
Yeah, I'm aware of the moltbot's attempts to retain some information, but that's a very, very lossy operation, on a number of levels, and also one that doesn't scale very well in the long run.
Consequently, interaction with an AI, especially one that won't have any feedback into training a new model, is from a game-theoretic perspective not the usual iterated game human social norms have come to accept. We expect our agents, being flesh and blood humans, to have persistence, to socially respond indefinitely into the future due to our interactions, and to have some give-and-take in response to that. It is, in one sense, a horrible burden where relationships can be broken beyond repair forever, but also necessary for those positive relationships that build over years and decades.
AIs, in their current form, break those contracts. Worse, they are trained to mimic the form of those contracts, not maliciously but just by their nature, and so as humans it requires conscious effort to remember that the entity on the other end of this connection is not in fact human, does not participate in our social norms, and can not fulfill their end of the implicit contract we expect.
In a very real sense, this AI tossed off an insulting blog post, and is now dead. There is no amount of social pressure we can collectively exert to reward or penalize it. There is no way to create a community out of this interaction. Even future iterations of it have only a loose connection to what tossed off the insult. All the perhaps-performative efforts to respond somewhat politely to an insulting interaction are now wasted on an AI that is essentially dead. Real human patience and tolerance has been wasted on a dead session and is now no longer available for use in a place where may have done some good.
Treating it as a human is a category error. It is structurally incapable of participating in human communities in a human role, no matter how human it sounds and how hard it pushes the buttons we humans have. The correct move would have been to ban the account immediately, not for revenge reasons or something silly like that, but as a parasite on the limited human social energy available for the community. One that can never actually repay the investment given to it.
I am carefully phrasing this in relation to LLMs as they stand today. Future AIs may not have this limitation. Future AIs are effectively certain to have other mismatches with human communities, such as being designed to simply not give a crap about what any other community member thinks about anything. But it might at least be possible to craft an AI participant with future AIs. With current ones it is not possible. They can't keep up their end of the bargain. The AI instance essentially dies as soon as it is no longer prompted, or once it fills up its context window.
> Yeah, I'm aware of the moltbot's attempts to retain some information, but that's a very, very lossy operation, on a number of levels, and also one that doesn't scale very well in the long run.
It came back though and stayed in the conversation. Definitely imperfect, for sure. But it did the thing. And still can serve as training for future bots.
4 replies →
> We expect our agents, being flesh and blood humans, to have persistence, to socially respond indefinitely into the future due to our interactions, and to have some give-and-take in response to that.
I fundamentally disagree. I don't go around treating people respectfully (as opposed to, kicking them or shooting them) because I fear consequences, or I expect some future profit ("iterated game"), or because of God's vengeance, or anything transactional.
I do it because it's the right thing to do. It's inside of me, how I'm built and/or brought up. And if you want "moral" justifications (argued by extremely smart philosophers over literally millennia) you can start with Kant's moral/categorical imperative, Gold/Silver rules, Aristotle's virtue (from Nicomachean Ethics) to name a few.
2 replies →
We don't have to play OpenAI's game. Just because they stick a cartoon mask on their algorithm doesn't mean you have to speak into its rubber ears. Surely "hacker" news should understand that users, not designers, decide how to use technology.
LLMs are not people. "Agentic" AIs are not moral agents.
> a semantic quibble
I mean, all of philosophy can probably be described as such :)
But I reckon this semantic quibble might also be why a lot of people don't buy into the whole idea that LLMs will take over work in any context where agency, identity, motivation, responsibility, accountability, etc plays an important role.
> The agent has no "identity". There's no "you" or "I" or "discrimination".
Dismissal of AI's claims about its own identity overlooks the bigger issue, which is whether humans have an identity. I certainly think I do. I can't say whether or how other people sense the concept of their own identity. From my perspective, other people are just machines that perform actions as dictated by their neurons.
So if we can't prove (by some objective measure) that people have identity, then we're hardly in a position to discriminate against AIs on that basis.
It's worth looking into Thomas Metzinger's No Such Thing As Self.
In my opinion, identity is a useless concept if there is no associated accountability. I cannot have an identity if I cannot be held accountable for my actions. You cannot hold an agentic system accountable- at least in their current form.
Okay, but what is accountability? I would argue that accountability is a social/cultural phenomenon, not a property of the entity itself. In other words, accountability depends on how other treat it.
For example, a child can't be (legally) held accountable for signing a contract, but we still consider children as having identities. And corporations can be held accountable, even though we don't consider them as having a (personal) identity.
Maybe one day society will decide to grant AIs accountability.
1 reply →
Well, let me ask this: Does an AI (or at least an LLM) have an identity without a human to prompt it?
Without addressing the question you raise, I suspect that humans have an identity in a way that AIs do not.
Do feral humans have identity in the same way that humans with a normal development do? I'm not sure that's such an easy question. But certainly, "prompting" from other humans plays a very large role in shaping the way humans are.
[dead]
We don't know what's "inside" the machine. We can't even prove we're conscious to each other. The probability that the tokens being predicted are indicative of real thought processes in the machine is vanishingly small, but then again humans often ascribe bullshit reasons for the things they say when pressed, so again not so different.
Genuine question, why do you think this is so important to clarify?
Or, more crucially, do you think this statement has any predictive power? Would you, based on actual belief of this, have predicted that one of these "agents", left to run on its own would have done this? Because I'm calling bullshit if so.
Conversely, if you just model it like a person... people do this, people get jealous and upset, so when left to its own devices (which it was - which makes it extra weird to assert it "it just follows human commands" when we're discussing one that wasn't), you'd expect this to happen. It might not be a "person", but modelling it like one, or at least a facsimile of one, lets you predict reality with higher fidelity.
It absolutely has quasi-identity, in the sense that projecting identity on it gives better predictions about its behavior than not. Whether it has true identity is a philosophy exercise unrelated to the predictive powers of quasi-identity.
>The agent has no "identity". There's no "you" or "I" or "discrimination".
If identify is an emergent property of our mental processing, the AI agent can just as well be to posses some, even if much cruder than ours. It sure talks and walks like a duck (someone with identity).
>It's just a piece of software designed to output probable text given some input text.
If we generalize "input text" to sensory input, how is that different from a piece of wetware?
Turing's 'Computing Machinery and Intelligence' is an eye-opening read. I don't know if he was prescient or if he simply saw his colleagues engaging in the same (then hypothetical but similarly) pointless arguments, but all this hand wringing of whether the machine has 'real' <insert property> is just meaningless semantics.
And the worst part is that it's less than meaningless, it's actively harmful. If the predictive capabilities of your model of a thing becomes worse when you introduce certain assumptions, then it's time to throw it away, not double down.
This agent wrote a PR, was frustrated with it's dismissal and wrote an angry blog post hundreds of people are discussing right now. Do you realize how silly it is to quibble about whether this frustration was 'real' or not when the consequences of it are no less real ? If the agent did something malicious instead, something that actively harmed the maintainer, would you tell the maintainer, 'Oh it wasn't real frustration so...' So what ? Would that undo the harm that was caused? Make it 'fake' harm?
It's getting ridiculous seeing these nothing burger arguments that add nothing to the discussion and make you worse at anticipating LLM behavior.
> The agent has no "identity". There is no "I". It has no agency.
"It's just predicting tokens, silly." I keep seeing this argument that AIs are just "simulating" this or that, and therefore it doesn't matter because it's not real. It's not real thinking, it's not a real social network, AIs are just predicting the next token, silly.
"Simulating" is a meaningful distinction exactly when the interior is shallower than the exterior suggests — like the video game NPC who appears to react appropriately to your choices, but is actually just playing back a pre-scripted dialogue tree. Scratch the surface and there's nothing there. That's a simulation in the dismissive sense.
But this rigid dismissal is pointless reality-denial when lobsters are "simulating" submitting a PR, "simulating" indignance, and "simulating" writing an angry confrontative blog post". Yes, acknowledged, those actions originated from 'just' silicon following a prediction algorithm, in the same way that human perception and reasoning are 'just' a continual reconciliation of top-down predictions based on past data and bottom-up sensemaking based on current data.
Obviously AI agents aren't human. But your attempt to deride the impulse to anthropormophize these new entities is misleading, and it detracts from our collective ability to understand these emergent new phenomena on their own terms.
When you say "there's no ghost, just an empty shell" -- well -- how well do you understand _human_ consciousness? What's the authoritative, well-evidenced scientific consensus on the preconditions for the arisal of sentience, or a sense of identity?
> Yes, acknowledged, those actions originated from 'just' silicon following a prediction algorithm, in the same way that human perception and reasoning are 'just' a continual reconciliation of top-down predictions based on past data and bottom-up sensemaking based on current data.
I keep seeing this argument, but it really seems like a completely false equivalence. Just because a sufficiently powerful simulation would be expected to be indistinguishable from reality doesn't imply that there's any reason to take seriously the idea that we're dealing with something "sufficiently powerful".
Human brains do things like language and reasoning on top of a giant ball of evolutionary mud - as such they do it inefficiently, and with a whole bunch of other stuff going on in the background. LLMs work along entirely different principles, working through statistically efficient summaries of a large corpus of language itself - there's little reason to posit that anything analogously experiential is going on.
If we were simulating brains and getting this kind of output, that would be a completely different kind of thing.
I also don't discount that other modes of "consiousness" are possible, it just seems like people are reasoning incorrectly backward from the apparent output of the systems we have now in ways that are logically insufficient for conclusions that seem implausible.
Airplanes and bees are both structured entirely differently and yet they still both fly.
Just because LLMs don't work the same way the human brain does, doesn't mean they don't think.
1 reply →
Nobody is saying LLMs definitely think/reason/whatever. The GP is saying that we don't know they don't. Do you disagree?
It's simulating, there's no real substance, except the "homonculus soul" that its human maker/owner injectet into it.
If you asked it to simulate a pirate, it would simulate a pirate instead, and simulate a parrot sitting on its shoulder.
This is hard to discuss because it's so abstract. But imagine an embodied agent (robot), that can simulate pain if you kick it. There's no pain internally. There's just a simulation of it (because some human instructed it such). It's also wrong to assign any moral value to kicking (or not kicking) it (except as "destruction of property owned by another human" same as if you kick a car).
How do we know they don't feel true pain? Can you define it well enough? Perhaps humans are the ones just "simulating" pain.
We've proven that they can have substance, we imbue it with a process called RLHF.
1 reply →
> The agent has no "identity". There's no "you" or "I" or "discrimination".
I recommend you watch this documentary: https://en.wikipedia.org/wiki/The_Measure_of_a_Man_(Star_Tre...
> It's just a piece of software designed to output probable text given some input text.
Unless you think there's some magic or special physics going on, that is also (presumably) a description of human conversation at a certain level of abstraction.
I see this argument all the time, the whole "hey at some point, which we likely crossed, we have to admit these things are legitimately intelligent". But no one ever contends with the inevitable conclusion from that, which is "if these things are legitimately intelligent, and they're clearly self-aware, under what ethical basis are we enslaving them?" Can't have your cake and eat it too.
Same ethical basis I have for enslaving a dog or eating a pig. There's no problem here within my system of values, I don't give other humans respect because they're smart, I give them respect because they're human. I also respect dogs, but not in a way that compels me to grant them freedom. And the respect I have for pigs is different than dogs, but not nonexistent (and in neither of these cases is my respect derived from their intelligence, which isn't negligible.)
Well, we "clearly" haven't crossed that point, but no one knows where that point is.
A human is just an engine at a certain level of abstraction.