I would rather hope that humans given the same information would respond by saying that the problem is ill-defined. While we know that the apple is on the LHS (call this the X-axis), we don’t know where it is on the Y-axis.
We know that the pear’s coordinates are (30,?) and that the peach’s coordinates are (?, ?). We do know that the distance from the pear to the peach is 1, but this is just the radius of a circle, with the constraint that the peach must be on the table.
This is what I mean by considering the semantics of the situation, rather than just the syntax.
To add to this, there is a thing called the “Duhem-Quine thesis”, which states that theories are underdetermined by their data. A corollary of this is that there are many, possibly innumerable, hypotheses that one can raise to explain a particular set of data.
The above example is an instance of this, given that there are an infinite number of Y locations for the apple, and an infinite number of possible angular coordinates for the peach.
I remain bewildered that AI is still marketed as a disruptive technology. Not to belittle it, I find it a very useful tool while coding. But like most “silver bullets”, it provides a few-percent speedup of my work, not a wholesale revolution of it.
AI masking, AI object removal, AI denoising, AI subject tracking all follow this pattern. A few percent here, a few percent there. A decent investment for a $10/month subscription. Heck, perhaps even enough to justify $20/m!
But the “tech bros” are selling a hundred billion dollar industry disruption. The disparity between $10/m and 100G$ is quite striking. I surmise that AI is therefore not sold to us simple folk, but to investors. From which follows that AI is not a product, but a promise.
This can mean, to my understanding, one of two things:
these tech bros have seen something I haven’t. Something revolutionary. Something disruptive. Something multiple companies have evidently produced in-house by hundreds of people, yet nothing about it has leaked to the public.
or, it’s a bubble.
An alternative read is that here’s a chance for FAANG to invest in something so expensive, only FAANG can do it. And they do, thereby keeping the playing field inaccessible to cheapskates without FAANG money.
I wouldn’t, but maybe I’m just not clever enough. From the wording, if e automatically assume the third one is colinear, too.
Anyway, I think the output from the LLM showed it was not just repeating text it had previously seen.
TBH, I don’t think there is a fundamental difference between biological and computational neutral networks, besides complexity. I think once those networks grow big enough, they will reach human-level intelligence. The largest models I know of have about 4×10¹¹ connections (weights), the human brain has The brain has about 10¹¹ neurons and 1.8-3.2 × 10¹⁴ synapses.
You touched a nerve! Last year, I invested about 27K in options to buy UiPath (PATH) by January 20 this year. They sell industrial robotic AI. My options ran out worthless … grump!
The major difference being that our consciousness is embodied, with many different sensory inputs. We also possess psychological continuity, something necessary for personhood.
Of course it is obvious. But I was biased, and thought the most logical way to read the puzzle was continuing on a line, since nothing was said about the other dimension. Again, it was a bias on my part.
I still argue that the LLM demonstrated some understanding of the problem, and some problem-solving skill. They are just text-continuation engines (as it can be beautifully demonstrated by asking them to write some text but output it in ROT13), but just like memory, logic and intelligence emerge from biological networks at a given complexity level, they also emerge from artificial ones, I think.
I’m not looking for consciousness, personality or personhood in an AI tool, but rather for problem-solving capabilities, logic, reasoning, barred on some permanently available, trained set, and some context.
I’d like to know WHO “the community” is and if they’d be so kind as to explain the reasons for deleting my post (the use of the word “inappropriate” has nothing to do with an argument).
Perhaps you are right, but at the moment we need to realize that we in general really don’t know a lot about how the brain functions at large. We currently think networking is an important aspect of it, but have little clear facts on what role the various nodes play and how the signaling between them really is carried out. Does every neuron “fight for itself” in some sort of anarchy and chaotic model, or is it a more coordinated operation? What is really the role of the numerous glia cells, do they participate in signals processing? etc. Also on the details level there are lots of holes and lacunas in our understanding, like is signal direction in dendrites just one-way or can it also be bi-directional?
Furthermore, the most part of those many neurons are situated n the cerebellum, which we (for practical reasons so far) know much less about. (It has traditionally been seen in connection with fine-coordinating and memory of muscle behavior, but its innervation connects as much with the frontal lobe part that we think have an important role in decision-making etc.)
We’ve learned a lot about the brain, but it’s so complex that we still have a long way to go. Psychology has at large for years considered the brain quite isolatedly, but as @epeeist point to, our brain functioning is embodied, and there are strong arguments that without a basic “intelligence” at a cellular level in our body, the brain is of little use. (It’s basic task is not to solve quizes, make art, communicate with others or whatever, but to provide a prognosis of what is probably the smartest next action for our body to sustain survival.)
I think we are currently using so many (folks-psychology developed) terms like e.g. “reason”, “intelligence”, “consciousness”, “emotion” etc, that we really don’t know what designate – at least we have no unified agreement on what they shall be meant to describe (-- and they may therefore actually represent blocks on our road towards better understanding).
But as such your initial statement is of importance, because it accentuates the question of what really is the function underlying the “I” in AI – and that is to me the most interesting question at the moment, not the least because an understanding of this points much further ahead.
The logical deduction here is that it is impossible to determine a unique solution, since the problem is not fully defined. The AI does not identify this, which would lead me to question its logical capabilities.
My understanding (such as it is, I am perfectly willing to be told that I am spouting nonsense) is that AI systems are large Bayesian networks underpinned by Solomonoff induction, with Monte Carlo methods used to sample the solution space. But this begs the question, since it assumes that there is a solution,
In your example, no single solution can be determined from the specification, but the AI does not identify this. Instead, it presents, with certainty, a single solution. It seems to be unable to recognise that it is not possible to answer the question asked.
AI/ML does not “recognize” anything. LLMs are sophisticated Markov chains. You will never human-style cognition out of them, it is just humans reading these into the generated text.
Currently LLMs are trained on input which usually has a solution, so it has a hard time coming up with “no solution”, you can’t really blame it. Once they include problems without solutions in the training data, you will get output saying that. Of course you will get that for some problems which have solutions, in many cases trivial.
LLMs are useful tools for repetitive tasks that an be reasonable automated. Machine translation, for example, is already amazing, either within programming and natural languages. It is not perfect, but it does not need to be for it to be a huge productivity booster.
It beats me why people try to use LLMs for sophisticated queries beyond the above, like math problems or philosophical questions. It’s like asking my power drill for relationship advice. Sure it is fun to experiment, in the 1960s people had long conversations with Eliza, but I guess the novelty will wear off and users will treat these tools like any other tool.
To show my age, one of my wife’s pet peeves was her students simply accepting the results shown on their calculators as correct, regardless of how absurd it was.
My worry is that those using LLMs will simply accept what it produces. To quote Bertrand Russell’s somewhat cynical aphorism, “Most people would rather die than think; and indeed they do so”.
Correct me if I’m wrong, but as I understand it, the current LLM’s “services” are not only an LLM but a combination of multiple systems that work together to provide or solve a prompt. For example, whilst an LLM cannot solve a math problem by itself, it can interact with systems that do and integrate it into its answer. It doesn’t run on mere prediction but also tries to execute the calculations in real time to confirm its output. Just like it can browse the web for you and distill one or multiple websites of information (and provide sources so you can actually go look for yourself and confirm).
Quite so. For example today I was interested in what resolution in lp/mm is represented by a dot location on a FFT. I asked the infamous ChatGPT 3.5 and got reasonable responses much quicker than using a search engine.