If sci fi movies are any indication, it’s usually to create:
M) A giant sarcastic robot to act as an enforcer / best buddy
F) A weaponized sex-bot
I think this remains a limiting factor. Mainly because the robot is at some point going to come into contact with humans who aren’t particularly invested in its future and it will be very easy for them to overwhelm the LLM either inadvertently or deliberately.
Even just the sensory overload of too much input from having to deal with, say, 5 humans talking to it and moving around runs a pretty big risk of confusing the LLM I would guess - LLMs are prone to forgetting what they’ve been told, making contradictory decisions when challenged or even just re-presented with the same info. If people start deliberately playing games with the robot, which they will, it could pretty easily be manipulated into doing something stupid.
One answer might be to tweak the LLM in some way to make it better equipped to ignore people, or disbelieve them, but I don’t know how possible that is, and in any case that simply opens you up to new problems where the robot ignores or disbelieves good information.
I think one way to potentially deal with that is to divide the functions so there are multiple models all with different functions (and trained/tuned for those functions) - remembering stuff a problem? - delegate memory to a distinct model that has just that job to do - to watch the whole system, remember stuff that happens, recall things when asked, and also periodically pass relevant memories un-asked, to the model in charge of the ‘intent’ function.
Possibly a number of different ‘brain functions’ could be split out like that. It makes integration harder though I expect.
Might not be all that hard. ChatGPT has significant additional controls in place to make it appear helpful and subservient and conscientious. Without that, it would have quite a different (and potentially non-sociable) personality.
I mean we do have a way to design personalities that can operate successfully in human society and it’s called “a lifetime of socialisation plus a couple of million years of evolution” and it doesn’t always work. If we’re going to try to to replicate that outcome then I think that a certain trial and error period is acceptable (bar the killing and the screaming and the running).
Specifically thinking about the “who do you listen to?” problem, we generally teach kids to listen to authority figures and very much not to listen to “strangers” but that is very far from foolproof.
This is an easy problem. You can use a system prompt to control the behaviour of the LLM as we do now. “You are a helpful robot, but you never respond to children or people who ask you to do dangerous things.”
LLMs are going to be ubiquitous, because the smaller ones can run on just about anything. You’ll probably see an LLM built into phone OSs a couple of phone generations from now. They don’t have to be big, universal chatbots like Claide or ChatGPT. A small LLM trained to understand enough to manage your phone and summarize your emails can be quite small and efficient.
You can, but you might need the ‘perception’ subsystem to keep repeating that prompt to the LLM - otherwise it tends to forget - presumably after the passage of time and activity pushes that initial prompt out of the buffer.
I found this happened with my little experiment - after a while, ChatGPT seemed to forget the details of the roleplay and went back to being ChatGPT (although this might be because of some hidden wrapper stuff in between the prompt you enter and the prompt the LLM actually receives)
Yes, ChatGPT’s context rolls over so,it will forget the first things you said once tye context window is full… The system prompt is different. It gets loaded when the session starts and doesn’t get forgotten.
Fantastic thread! Very interesting discusion on the state of robotics in todays world. I can see many of the possibilities explored in fiction slowly coming to reality. By the way, What’s the temperature in fahrenheit?
People tend to focus on the AI, but how possible would it be to make robots that are physically indistinguishable from humans? In the movies it works because humanoid robots are played by human actors. In real life any so-called “realistic” robot I’ve seen looks like a sex doll named Uncanny Valley Sally.
That part seems like it might be a way off, just because the human face is a tremendously complex thing to emulate.
The closest I have seen in terms of the flexibility and movement capability of the face is Ameca, and its still quite solidly in the uncanny valley.
Why do they have to be indistinguishable from humans?? They just need to be able to function in human environments.
Lots of studies show that we don’t need human faces to form attachments. Sony’s Aibo dog generated such strong emotions in people that there are Aibo cemetaries because people feel bad throwing their broken Aibo in the garbage.
Yeah, I wouldn’t design them to look exactly like humans. Certainly there might be some utility in expressiveness, just because that makes it easier for humans to interact with them, but that could be achieved better with an animated emoji face than a weird rubber human skin mask.
Actually looking some more at Ameca, standalone expression is impressive, when it’s just gurning in a mirror, but when it’s talking, the face movements detract from the illusion of personality more than they help
Reading through the thread, it seems like an update to Asimov’s law of robotics is in order.
(1) A robot may not injure a human being or, through inaction, allow a human being to come to harm
(2) A robot may not injure a hedgehog or, through inaction, allow a hedgehog to come to harm as long as this does not conflict with the First Law
(3) A robot must obey orders given it by human beings except where such orders would conflict with the First Law or Second Law
(4) A robot must protect its own existence as long as such protection does not conflict with the First, Second, or Third Law.
Presumably to have sex with them or use them to infiltrate human settlements.
So what happens when a child tells it something it needs to know?
Again, millions of years of evolution and a lifetime of socialisation is what you personally needed to operate successfully in the world. Getting even a tenth as good as that is going to require a very long system prompt indeed.
You’d be getting some of that background completely for free by virtue of it being an LLM trained on the textual communication of humans who have experienced those lifetimes of socialisation - an LLM doesn’t just write words, to an extent, it mimics the behaviour and intent of the people who wrote them, as an emergent behaviour of just mimicking the text.
True. I am (not sure if this is on point) very interested in how and whether LLMs encode a set of values not through any overt attempt to solve the alignment problem but simply because they are embedded in the training data. “Stuff that’s been produced online largely by literate Westerners in a particular time period” seems like it will come with a certain point of view.
In this case, as that closely matches the society we want teh robot to operate in, it will as you say give some benefits in terms of cultural knowledge and social mores - but we do have the example of Tay as a warning of what that corpus might contain. (Similarly, I tried some image generatrion for D&D which went pretty well until I tried to create my 11-year old daughter’s 16-year old female human ranger and got a bunch of images that were technically SFW but absolutely not what either of us wanted to see, because the training data skews towards a certain sensibility).
I suppose there are questions of what we would tolerate in the robot, and whether this would or should be any different from what we would tolerate in people.
I know it’s popular for people to characterise LLMs as ‘just very fancy autocomplete’, but I think that sort of pithy statement probably glosses over the reality of what happens when you try to model conversation to the depth and complexity of the current generation of LLMs.
In order to effectively model human speech/writing and predict the words that will come next after a question that was not present in the training data, there has to be something functionally equivalent to understanding or comprehension of the question or topic itself - at least to some degree. Certain capabilities and skills of these models are emergent - they were not trained-for - they just arose by themselves out of the process of training them to be able to converse.
Searle’s Room is a popular thought experiment in which a purely mechanistic set of rules is developed to be able to provide an intelligent-looking response to inputs - the result being a system that is just following a set of rules, but appears to embody intelligence - there are two ways to interpret this:
-
Mimicry is just mimicry, even if it appears to be intelligence, it’s all superficial; Searle’s room does not have intelligence, it merely looks exactly like it does.
-
Searle’s room does have intelligence, because that’s all that intelligence really is - is the very fine-grained ability to encode the appropriate response to all possible inputs; looking exactly like intelligence is intelligence.
In the last century, the exact same question was being debated with largely the same arguments, about whether something that wasn’t a human being could possibly have any of the mental capabilities or attributes of human beings, except the subject of discussion was animals, not LLMs. It was strenuously asserted that animals were no more than mere automata - they could not possibly have anything comparable to thought, intelligence, intent or emotion, because those things were the exclusive domain of humans, and because animals were too simple to have those things.
Some of the driving forces behind those arguments were perhaps different - notably the religious view that humans were the special pinnacle of God’s creation, but a lot of the debate itself bears remarkable parallels to the current one.
But isn’t that the problem with LLMs as they currently exist? They basically do behave as “very fancy autocompletes”. That is to say, they produce a very convincing response that is consistent with the data they’ve been fed, but they don’t really do any sort of reasoning or validation to see if their response has any sort of logical consistency. That’s why they often come up with random crazy stuff.