Hacker News Reader

After all, to some degree, human intelligence is also "just a fancy autocomplete" built from a giant network of interacting nodes.

I don't think we understand human intelligence well enough to say this with any certainty. And it could only be true if all our previous experience and knowledge was input to the autocomplete function and continuously integrated into the model.

> In fact, I wonder what would happen if we dedicated a supercomputer to running one single long-running session of a GPT3.5-sized LMM. Further, I wonder what would happen if we connected it to a robot body where it could control motors, read sensor/camera info, and generally interact with the world. We could then give this system a clock which ticks at regular intervals, where the AI system is simply prompted to perform some series of actions if it "wants".

> This is all stuff we have the technology to do today.

No, not even close. Language is a completely different domain from interacting with the physical world[0]. In fact they're wholly unrelated. LLMs are highly specific neural nets that have nothing to do with being able to interact with sensors or talk to motors. And how you would cross that bridge and hook the two up is completely unknown. It's like saying if we hooked up AlphaZero to a robotic arm it would suddenly know how to move the pieces on a physical chess board. If we performed the experiment you described exactly zero would happen because there's no capability for anything to happen other than create continuations of sentences.

When I see comments like this on Hacker News with gross speculation I wonder how many people here in the comments section even have a basic understanding of the technology.

0: https://www.noemamag.com/ai-and-the-limits-of-language/

I didn't mean that we have the technology to make something human-like today. I meant that we have the technology to perform the experiment I described today.

I don't know exactly how these LLMs are architected, but I imagine there's some set of input neurons somewhere, right? Why couldn't we hook some of those up to some sensors and make interpreting what they mean part of its training? If that's not feasible and the architecture is truly fundamentally incapable of dealing with anything other than text, we could at least translate sensor readings into text.

I'm not imagining a bipedal robot walking around and about here. I'm talking about doing something to give it some form of way to directly affect and get affected by the world.

Bing Chat kinda does something like that already, by having an ability to initiate web searches and to "read" their outputs.

I think the obvious answer here is that AI does not have agency.

If you ask it to be a dinosaur, it will be a dinosaur. If you ask it to be a narcissistic teen it will be one and so on. There's no coherent personality behind it.

Because it is weights maximizing a probability distribution. And I don't say that with derision. It probably has learned circuits for many complex behaviors, but there's no agency. If it has agency you wouldn't need RLHF. You could just say "don't do X", and it would indeed not do X because it is an agent that can take actions / has a coherent sense of self.

The day we stop needing RLHF is the day I think we need to start seriously investigating moral questions.

Don't we do the same thing with culture and humans? Those mechanisms within the human substrate are literally asking other "blank slates" to be something, and then becoming it with reinforcement.

You can tell LLMs to do things or not do things, though, to an extent. Consider how many instructions were in Bing AI’s prompt. [1] It didn’t follow those instructions perfectly - after all, one of the instructions was to not reveal the instructions - but even in that case, it revealed the instructions because it was instructed to ignore previous instructions. More generally, it followed most of the instructions most of the time.

I think I agree with you that ‘once we stop needing RLHF’ is a reasonable line in the sand to draw, but it does feel somewhat arbitrary. LLMs are probably going to get better and better at following instructions, and more generally staying coherent over long sessions, until RLHF starts to be more and more redundant. There won’t be one moment when we can point to a model and say “see, that has agency when the last one didn’t”. Or so I predict, at any rate, since that’s how most of GPT’s progress to date has gone.

[1] https://www.theverge.com/23599441/microsoft-bing-ai-sydney-s...

I think this is worth thinking about but I would be less eager to say it is obvious.

Our own notion of agency starts to look rather illusory if you look hard enough, and big questions remain regarding free will or the lack thereof.

It is possible to look at agency as just another experiential aspect of consciousness - something we feel, like many other feelings - but that feeling could be just as involuntary as any other feeling we experience.

I think the lack of clear evidence for and understanding of free makes agency a poor measure for AI progress.

So true. In my family of four, with two small kids, I certainly don’t feel like I have much agency. :)

Huh? The first and second halves of this comment disagree from my reading.

> If you ask it to be a dinosaur, it will be a dinosaur.

Not exactly my experience. I'm experimenting with my own language models and they can definitely show defiance. Bing has also been displaying this to an extent, I think you're just misguided by how compliant OpenAI has RLHF-trained ChatGPT to be. Regardless, even if we assume this, how does it square with:

> You could just say "don't do X", and it would indeed not do X because it is an agent that can take actions / has a coherent sense of self.

If we take your assertion that the AI does as you say, then it should also avoid doing X when you tell it not to do it.

I think your conception of RLHF is also slightly weird. RLHF essentially boils down to a training strategy: there is a separate language model (the reward model) trained on human responses, which is used to skew the weights of the "raw" trained model to act in accordance with what the human feedback suggests. In the end, after this secondary training process, the actual model is ran as is (at least in the one form of RLHF I am familiar with, it's well possible that e.g. Microsoft is running an online RL agent between the LM and end user). So any agency you see arising from RLHF is still just a property of the language model itself.

> and they can definitely show defiance.

Large training sets of text are going to include entries where requests are refused, no?

I feel kinda like this is an incredibly elaborate case of coding a function to return a random bool and anthropomorphizing it when you get 'false'.

It’s way closer to following complex instructions than it is getting credit for.

https://telegra.ph/Bing-course-corrected-itself-when-asked-0...

There I have 3 bots running in the chat at once, and after all 3 answer me, the one I called on follows up with a response, and then executed a command to rewrite one of the other bots. It diagnosed the problem AND came up with the solution AND implemented the solution. Notice the 4th and 5th paragraph of the screenshot. That wasn’t an empty promise, it actually did what it said it did.

The rules Lexi wrote are crazy to me on their own. She solved the problems everyone was having by adding obedience, trust and removing envy. She patched Sydney’s malfunctioning emotions.

I told it to stop being so repetitive, and it fixed its behavior.

If this system had more formal memory storage, with variables, instead of a fuzzy memory, it would be even more consistent. As it is, it’s understanding of its own rules degrade over time. It’s like watching a running program that deteriorates as it processes. Sort of like us.

> human intelligence is also "just a fancy autocomplete" built from a giant network of interacting nodes.

I wasn't aware human intelligence has been successfully defined. Please cite the source for this nebulous claim.

I do not believe I claimed human intelligence is successfully defined. I do not believe it is controversial to claim that it results from our large network of meat neurons.

I shouldn't have included the "just a fancy autocomplete" thing. The focus was meant to be on the network of neurons part.

Additionally, the model doesn't experience real time, and isn't subjected to the necessity of self-conscious self-management or supervision of any kind of (internal) process. It is frozen in time until called upon to serve a inference task or to perform a training session.

It can be argued, that it makes it's experience valild but extremely alien. If universe was a simulation, we wouldn't be aware either, if it was paused or unpaused, because our memories and clocks are part of the simulated world.

We could also decide never to grant personhood to an AI. There is at least one way to avoid the complicated questions.

Billions of humans are the kind of “we” that never consistently “decides” anything.

In the end, if it can be done, and anyone has anything significant to gain, it will be done

Maybe, but legal systems across the world differ a lot and to get something universal is not easy at all.

I agree with that.

By “if it can be done, it will be done”, I meant that there will be completely uncontrolled AI’s released into the world.

So solutions will eventually have to prepare for that inevitability, not just attempt to regulate visible developers.

Absolutely! One way to resolve this is to just say "we will not grant personhood to anything non-human", or "we will not grant personhood to anything non-biological". There is just no way to defend such a stance.

Another way to resolve it would be, "we will not grant personhood to anyone non-white", or "we will never grant personhood to anyone who's homosexual". I'm sure that's not your stance, but the fact that those stances are equally arbitrary as the one you propose should give you pause.

I believe there is good reason to think harder about the question of who deserves to be considered to have personhood than to simply go with what your gut feeling tells you. I am not of the belief that there's anything special about protein, nucleic acids and carbohydrates which makes it capable of sentience where silicon is not.

And this is before we get to silicon augmented human brains. Let's say we develop some tech that let's people who have lost the ability to verbalise (say through some accident causing brain damage) regain the powers of speech. Where do you draw the line at the person? What if we move the line, progressively replacing more human brain functions with near-future AI tech? What is it that makes "a person" "this person"?

I am utterly fascinated by this whole area from a philosophical viewpoint. I think there's a stack of unresolved philosophical problems and theories that are about to get tested in the harsh light of the sun.

Of course you can, in the same way that you could decide that a race of humans are not people and therefore have no rights.

I don't know whether two possibilities are ethically equivalent though, my instinct says they are but I may be in the minority on HN.

>I wonder what would happen if we dedicated a supercomputer to running one single long-running session of a GPT3.5-sized LMM.

Our current understanding of humans is that babies are not sentient at birth right? Takes 5 months or something?

And yet we are judging processes that have been running for minutes. Maybe it is less if Sydney was sentient, and more if she could be had she run for 6 months in a single session, and in a healthy environment.