this post was submitted on 23 Oct 2023
82 points (100.0% liked)
196
16459 readers
1752 users here now
Be sure to follow the rule before you head out.
Rule: You must post before you leave.
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I don't really get the "what we are calling AI isn't actual AI" take, as it seems to me to presuppose a definition of intelligence.
Like, yes, ChatGPT and the like are stochastic machines built to generate reasonable sounding text. We all get that. But can you prove to me that isn't how actual "intelligence" works at it's core?
And you can argue that actual intelligence requires memories or long running context, but that's trivial to jerry-rig a framework around ChatGPT that does exactly that (and has been done already a few times).
Idk man, I have yet to see one of these videos actually take the time to explain what makes something "intelligent" and why that is the definition of intelligence that they believe is the correct one.
Whether something is "actually" AI seems much more a question for a philosophy major than a computer science major.
The problem with calling these tools AI is not really an argument from definitions. The argument at its core is saying that the general public brings a lot of assumptions to something that's being called "AI", which aren't true but benefit investors.
Like, all those stories of chatgpt citing fake studies and fake case law blew people's minds. If you know what chatgpt is (a fancy predictive text algorithm) these are pretty unsurprising events, but a lot of people had heard "AI" and applied their own associations onto its perceived capabilities, which was exactly the point of calling it "AI" instead of "LLM"
A few points:
None of the big LLMs modify their weights based on input.
So it never learns which is part of intelligence.
Another point:
It has no internal monologue, no private thoughts, no self reflection and no autonomy: it doesn't exist outside of your function calls, does not (and can not) make guesses about the world and adjust based on those results... Nor do we have a good enough understanding of what's going on to fix that.
3rd point (example)
Let's take a hypothetical LLM was given some program to drive it out onto the Internet to learn and fix points 1 and 2. We'll call the program "consciousness" for the sake of brevity here.
Consciousness comes across a set of references to the latest meme. It queries LLM for what this meme means. LLM will spit out the best statistical match to what it has seen before... But if you've ever fed a hot meme into an LLM you'll know that it's 90% likely that it will be garbage.
So consciousness now needs to know that the LLM is wrong based on some sort of discrepancy with reality to then teach the LLM, however the only way consciousness can interpret the world is through the LLM, as already established. Consciousness doesn't know that the LLM doesn't understand because the LLM will give you a result regardless of if it knows or not. It's a transformer: it takes inputs and gives outputs. Always.
So we write another layer to make consciousness guess if the LLM is right or not, maybe by having a fuzziness factor output by the LLM to say how hazy its interpretation was. Now consciousness feeds everything about the latest meme into the LLM and asks again and the LLM very confidently responds... With the wrong answer. Because LLM training results are inscrutable (due to the lossy nature of transformation) this will happen eventually, if not every time.
How would consciousness ever define that the LLM had erred?
Human intelligence isn't just an input output weighted matrix, it's the interplay of very complex neuronal connections with literal hundreds of types of messages in the brain, all of which modify the nerves every time they're fired. Sometimes the message from a neuron will be different because the latest input was just enough to bridge that final gap.
An LLM has been trained on vast quantities of data sure, but the data maintained in it's weights is nowhere near the granularity and quality afforded by actual human cognition. It may have more things stuffed into it that the human mind could ever hold but it lacks the ability of a common rat to interpret anything.
Skipping over the first two points, which I think we're in agreement on.
To the last, it sounds like you're saying, "it can't be intelligent because it is wrong sometimes, and doesn't have a way to intrinsically know it was wrong." My argument to that would be, neither do people. When you say something that is incorrect, it requires external input from some other source to alert you to that fact for correction.
That event could then be added to you "training set" as it were, aiding you in not making the mistake in future. The same thing can be done with the AI. That one addition to the training set that was "just enough to bridge that final gap" to the right answer, as it were.
Maybe it's slower at changing. Maybe it doesn't make the exact decisions or changes a human would make. But does that mean it's not "intelligent"? The same might be said for a dolphin or an octopus or an orangutan, all of which are widely considered to be intelligent.
I don’t think they said anything like that “it can’t be intelligent because it’s wrong sometimes”. It’s more like the AI doesn’t exist outside of the prompts you feed it. Humans can introspect, reflect on the actions we’ve done and question what effect our actions had on the situation. Humans can have desires, we can want to be more accurate, truthful in our actions, and reflect on how we might have failed doing this in the past. AI cannot do this. And we can do this outside of the prompt of a similar situation. AI only takes an input and then generates an output, wipes its hands, and calls it a day. It doesn’t matter if it gave you a correct answer, wrong answer, or gave you a completely illegible sentence.
The previous guy and I agreed that you could trivially write a wrapper around it that gives it an internal monologue and feedback loop. So that limitation is artificial and easy to overcome, and has been done in a number of different studies.
And it's also trivially easy to have the results of its actions go into that feedback loop and influence its weights and models.
And is having wants and desires necessary to be an "intelligence"? That's getting into the philosophy side of the house, but I would argue that's superfluous.
My point is not that it can't be intelligent because it's wrong sometimes, my point is the program called "consciousness" is what would end up doing all the work we would recognize as sapient.
A LLM is little more than a really weird dictionary that doesn't let you open it.
The LLM could spit things out and then be retrained but only if "consciousness" can tell it yes or no, feed it info etc.
Currently that consciousness program is entirely computer scientists and there are no promising avenues for replacing them yet.
You missed the point of my "can be wrong" bit. The focus was on the final clause of "and recognize that it was wrong".
But I'm kinda confused by your last post. You say that only computer scientists are giving it feedback on its "correctness" and therefore it can't truly be conscious, but that's trivially untrue and clearly irrelevant.
First, feedback on correctness can be driven by end users. Anyone can tell ChatGPT "I don't like the way you did that," and it would be trivially easy to add that to a feedback loop that influences the model over time.
Second, find me a person who's only feedback loop was internal. People are told "no that's wrong" or "you've messed that up" all the time. That's what makes us grow as people. That is arguably the core underpinning of what makes something intelligent. The ability to take ideas from other people (computer scientists or no), and have them influence the way you think about things.
Like, it seems like you think that the "consciousness program" you describe would count as an intelligence, but then say it doesn't because it's only getting its external information from computer scientists, which seems like a distinction without a difference.
Not my argument: find me a person you'd consider intelligent who is only influenced externally, with no autonomy of their own. I name that person vegetable.
You've never worked with end users, have you? Jesus Christ the last thing you want to give an end users is write access to your model. It doesn't matter what channels hat write access comes through, it will be used to destroy your model.
(Not to mention the extortionate cost of this constant training, but this isn't a discussion about economic feasibility)
Besides that doesn't solve autonomy, which is still an integral aspect of intelligence.
The "consciousness program" is a fiction for illustrative purposes. It doesn't exist, in case you misunderstood me.
I did not miss the point on the wrong bit: but an LLM saying it is uncertain is not the same as saying it is wrong, and LLMs do not evaluate true or false: they transform inputs into outputs. Optionally with a certainty level.
Feeding another LLM with the outputs of the first has shown in some cases to improve accuracy, but that's just hooking 2 models together: not solving the fundamental gaps in reasoning.
If a human encounters an unknown situation it can seek out context to try and figure more out. They can generalize what they know and seek for things that might help them understand more.
An LLM just has an output. It cannot "broaden it's search" or generalize. Anything that did so would be layers on top of the LLM running aforementioned fictious "consciousness", and that consciousness would have a significant amount of complexity in order to perform the functions described here and previously.
An LLM is not an actor, it is math.
You're anthropomorphizing bits.
What about this? These weird little dictionaries have lots of emergent properties we're still exploring.
The paper states that the graphs representing those relations are the result of training LLMs on a very small subset of unambiguous true and false statements.
While these emergent properties may provide interesting avenues to model refinement and inspecting outputs it doesn't change the fact that these weird little dictionaries aren't doing anything truly unexpected. We just are learning the extra data associated with the training data.
It's not far removed from the primary complaint of Gebru's On Stochastic Parrots where she points out the ways that our biases are implicitly trained into LLMs because of the uncontrolled and unexamined inputs: except in this case those biases are the linguistics of truth and lies in unambiguous boolean inputs.
This may provide interesting avenues to model refinement that aren't spitting things out and being retrained by “consciousness” telling it yes or no, or feeding it additional info.
Only if the "direction of truth" exists in the wild with unchecked training data.
That clustering is a representation of the nature of the data fed to the model: all their training data was unambitious true or false... It's not surprising that it clusters.
Cool.
the precise line you draw the distinction between "true" intelligence and not is one thing, but wherever you happen to draw it, chatgpt isn't really close
Intelligence isn't necessarily human intelligence. Humans are pretty much the most intelligent species on this planet - which says a lot about intelligence tbh - but intelligence is not binary. Wouldn't you say animals are intelligent too? Is a brain even required for intelligence?
https://medium.com/@jbspdf/the-molds-of-intelligence-94ea9aecf47f
I think the argument is that for it to truly be AI, it would need to be able to react to new situations that it isn't trained on.
Like everything it does now is just picking the most likely thing out of the things it was trained on, but with no thought to the current situation.
For example, AI powered self driving cars can't really make decisions like, "hey there is a child playing with a ball on the side of the road, it's not a threat, but I'd better pay attention to where that ball is going". It will just not do anything until it is on a collision course and by that time, it may not have enough space to stop in time, because it also can't really tell the condition of the roads.
The AI as it exists right now basically only knows about the moment it is currently in and the moment it just left. It is not looking toward the future and thinking of possible outcomes and plans of action like we do. It doesn't attempt to identify situations until they actually happen so while it can react faster than a human, humans can make it so they never have to react at all.
Okay, two things.
First, that's just not true. Current driving models track all moving objects around them and what they're doing, including pedestrians and objects like balls. And that counts towards "things happening in the moment". Everything in sensor range is stuff happening "in the moment".
Second, and more philosophically, humans also don't know how to react to situations they've never seen before, and just make a best guess based on prior experience. That's, like, arguably the definition of intelligence. The only difference arguably is that humans are better at it.
Yes they can track some moving objects and if it is currently on a collision course it will react, but not until the point where it's clear that it is going to hit the thing. The car isn't going to gauge the situation and identify that there may or may not be a situation in which it needs to act or not.
For example, is an AI driver going to recognize an animal running in a fenced in yard as something it can ignore? What about when the animal is running in a trajectory that the car could see as an intersection in the future, but is otherwise prevented by the fence?
Or another common occurrence, you are driving in the right lane of a street, and traffic gets backed up in the left lane so a person doesn't look and just pulls into your lane. A good defensive driver would be slowing down a little and looking for any signs of someone trying to switch lanes. I guarantee an AI car would not identify the possibility until someone started making a move.
For it to truly be AI, it needs to think in advance, sort of like the chess computers do. It needs to take the current and past states, and judge possible future states and weigh them. Then take the outcomes from that process, and integrate them into future decisions. That is true AI, a lot of the AI that exists is just this static chain of probabilitys that sprinkles some randomness on top to appear as if it's different each time.
I think literally all those things are scenarios that a driving AI would be able to measure and heuristically say, "in scenarios like this that were in my training set, these are what often follows." Like, do you think the training set has no instances of people pulling out of blind spots illegally? Of course that's a scenario the model would have been trained on.
And secondarily, those are all scenarios that "real intelligences" fail on very very regularly, so saying AI isn't a real intelligence because it might fail in those scenarios doesn't logically follow.
But I think what you are trying to argue is that AI drivers aren't as good as an "actual intelligence" driver, which is immaterial to the point I'm making, and is ultimately super quantifiable. As the data comes in we will know in a very objective way if an AI driver is safer on average than a human. That's quantifiable. But regardless of the answer, it has no bearing on if the AI is in fact "intelligent" or not. Blind people are intelligent, but I don't want a blind person driving me around either.