GPTs Are Maxed Out
thealgorithmicbridge.com
thealgorithmicbridge.com
There is no AGI. No one is creating Minds from Iain M Banks books. It's the world's best autocomplete. It's if you look at the Earth once from space, and "recreate" it by making a giant disc that looks like that view of the Earth. Nothing behind it is like intelligence as we would think of it.
AI is going to do some amazing things still. AGI is just an unbelievably high bar.
A parrot can produce human sounding speech and associate certain prompts with desired responses, but there's no internal process that actually understands the meaning of either the prompt or the response (other than a shallow connection of sounds to a specific thing, rather than understanding human language). You'll never be able to train a parrot well enough to have that process in the same way a human does. Our current terminology would be like calling parrots "artificial humans".
LLMs are still great tools for various tasks, they just shouldn't be categorized as AI.
An intelligence should be able to actually understand its input / output (rather than appear to understand), be able to reason about itself recursively*, be able to learn and generally have state, be able to have an internal thought process without being prompted**.
Some emotion isn't strictly a requirement but likely a desirable quality for human interaction.
* "Gödel, Escher, Bach: an Eternal Golden Braid" and "I am a Strange Loop" are meant to cover this but I've only started the latter
** A sci-fi book series called Old Man's War features a race of aliens that are intelligent but not sentient due to not having an internal thought process and self awareness, which does raise the question of whether it's possible to have one without the other.
You're creating a definition that doesn't align with the existing definitions for the words you're using.
Which is fine, it happens all the time, though it's less useful.
How does that differ from intelligence? (Assuming you were actually trying to answer the question)
Edit: why would I not be trying to answer the question?
If artificial intelligence is exactly the same as intelligence, then wouldn't we just call that intelligence? People won't use words if there is no useful meaning conveyed. I am not sure you have made clear what useful information is found within "artificial".
> I'd assumed "and not natural / created" was a given
And, really, what does it mean to be not natural or created in this context anyway? The dictionary definition is kind of hand wavy to begin with, leaving whether humans with intelligence are natural or created up to interpretation.
Artificial is an adjective here, as in a type of intelligence that is artificial. It's not exactly the same, it's a subcategory.
You seem to be using "artificial" as "mock", as in "seems like but isn't". That's not what I think people mean when they refer to AI in this context.
If e.g. humans are living in a simulation then it's completely fair to say they're AIs. If humans are a product of nature then they're not.
Implying that humans don't produce humans, or that human intelligence is artificial?
> It's not exactly the same, it's a subcategory.
So, again, for what reason would anyone take the time to call it "artificial intelligence"/"AI" when, as a subcategory, the intent is already captured by "intelligence"? Terms need to have useful meaning to stand up, and we are still not clear on where you see "artificial" as being useful.
Perhaps the problem here is that you misunderstood the original question? It asked how you would define "AI" as a single term, not how you would define "artificial" and "intelligence" independently. I think you've done a reasonable job of the latter, but that doesn't justify the "AI moniker" as described originally.
> You seem to be using "artificial" as "mock", as in "seems like but isn't".
I don't remember using it at all, unless you mean where I asked you what "artificial" adds to the term. What are you referring to here, exactly?
I mistook you for another commenter.
What's your end goal here? You're welcome to share your own definition of AI but this feels like sealioning. I think I've been clear enough, if I haven't this comment thread will just have to live without another ultra clarification of my "off the top of my head" comment.
To understand the missing gaps you have left open. I assume you took the time to respond to the original comment because you wanted others to understand you. But, with those gaps still open, we don't yet. If my assumption is wrong, I'll accept it, but then if you don't want to interact why bring your personal musings here and not to your private journal?
> I think I've been clear enough
Okay. Perhaps you could explain your interpretation of my questions towards you to help me understand how I didn't make myself clear when asking them? I had hoped my questions and associated explanations made clear what I didn't understand, but obviously not. Happy to rephrase it in a way that is clear once I have a grasp of what is missing.
> You're welcome to share your own definition of AI
I could try if you wish, but under what specific context? Words and terms often change in meaning when the situation around them changes. There is almost never just one definition.
Why not just say "In this context, this is what AI would be"? You can supply a context.
Artificial means "human created, not naturally occuring", humans evolved in nature while artificial intelligence is something we humans constructed using our intelligence rather than just breeding naturally.
Its like how artificial insemination is when you make a woman pregnant without sex, the natural way is not artificial.
Nobody asked how you would define "artificial" and "intelligence" independently. The dictionary already has done that more than throughly. Logically, a definition for "AI" needs to be all encompassing. There are some out there that accomplish that, but you have straight up failed.
How do we know if bees understand what they're doing? It's agreed that they display intelligence and intelligent behavior, but not by your definition.
If it's not measurable or testable then it's not useful.
Are bees intelligent or do we use the same language in this case as "intelligent design", i.e. the behavior is complex and impressive?
Also this murkier for animals since they do possess some of the traits I mentioned like learning.
Although as I said this was off the top of my head, not an official definition.
I don’t trust most people who are developing AI to perceive or raise an AI.
Any actor that simulates any aspect of intelligence is a genuine AI - it doesn't even need to be adaptive.
Pong had AI.
However I don't know anybody these days who would seriously call A* proper AI.
Parrots are closer but I used them for the example since they also mimic instead of understand.
In other words, we're all parroting the marketing (aka propaganda) that was created to mislead us.
Maybe we should stop doing that, and refuse to use the term "AI" to refer to the object of this current hype cycle/bubble. Maybe we should call it the autocomplete bubble instead?
That's what artificial means (even Pong had AI). If it ever becomes intelligent, it will be a synthetic intelligence (not an AGI).
For example, artificial vs synthetic diamonds; the former just looks the part, the latter is the genuine article - but manufactured as opposed to naturally formed.
The term is fuzzy though and people mean different things by it.
Although your analogy does make sense, but I've never run into "synthetic intelligence" used as a term.
As far as science fiction goes, it's irrelevant - we live in the real world and have had AI for a long time now (e.g.: https://en.wikipedia.org/wiki/Bertie_the_Brain).
Sci-fi is many times a speculative guess about the possible future of technology, IMO its not completely irrelevant as an insight into what people mean or expect when they say certain things about future tech.
That's how one sci-fi universe draws the distinction. I'm not sure that that's binding on anyone else, but it was an interesting distinction.
> A synthetic diamond or laboratory-grown diamond (LGD), also called a lab-grown diamond, laboratory-created, man-made, artisan-created, artificial, synthetic, or cultured diamond.
Artificial also means non natural, not necessarily a mock or something that behaves like something else:
> made or produced by human beings rather than occurring naturally, especially as a copy of something natural
I think that was their exact point. The terms are equivalent but carry different connotations
No, they aren't. Artificial diamond generally refers to cubic zirconia - which is not a diamond.
Synthetic diamond refers to genuine diamonds produced by humans (rather than natural forces).
https://en.wikipedia.org/wiki/Diamond_simulant#Artificial_si...
I mean, I guess if you insist on the 1950s definition of AI, perhaps it did. I'm not sure it did even by the 1970s definition, though, and it absolutely did not have it by the 2020s definition.
And if you're going to claim that we should keep using the 1950s definition, well, languages change over time. If you want to communicate with people in the 2020s, use definitions from the 2020s.
I don't know why you're defining words the way you are, but it's leading you to an absurd position. As I said elsewhere, if you want to communicate with the rest of us, you need to use the same definitions we do. Otherwise you're talking, not about AI, but about trying to re-define words, and that's a really uninteresting conversation.
Like a cuttlefish disguised as a rock.
Technically it is an impossible bar as it, like AI, is a forever moving target. When chess was still considered AI, the Turing Test was considered the bar for AGI. Now passing the Turing Test is considered AI and chess is basic computing. Soon GPT and the like will just be basic computing and the AGI bar you can image now will become the next AI, and AGI whatever next forward-looking goal there is at that point. Lather, rinse, repeat.
By some measures, PARRY[0] passed this test in 1972. It's a bad test for general intelligence, although I'm not convinced it was meant to be that.
The Turing test is now considered too easy to be a real test of AI. In other words, these days we think that an algorithm that can't beat a human at "being human" is pretty bad.
No, that's a bad summary. The Turing test is not a great test of being human.
For me, that's the definition of 30 minutes of being human. And yes, it illustrates why being human for 60 minutes, never mind a year, is a more difficult problem than 30 minutes. But it's the same a real human would perform.
Would there be extra value in dropping these requirements? Dropping the text-only requirement, do it for a week or two, and maybe even imitate a human known to the panel? Yes, absolutely.
A coin flip would be a much simpler concrete test - you are human for the next five minutes if it comes up heads. For me, that's the definition of 5 minutes of being a human.
The problem is "for me" is no less (and probably a lot more) absurd/weird than the philosophy or religion you dismiss. No one I know of (until now) thinks the Turing Test is about "being human".
Rather, it's about whether machines can think. Or imitate thought - hence the name, the Imitation Game. And it just shows the limitation of the time - how could Alan Turing predict the incredible advances in data collection and parallel processing that would allow for giant statistical models that can imitate the output of thinking, rather than the thinking itself?
And then there's human thinking, like everyone else, EXCEPT Alan Turing, you don't discuss the actual thinking process. There's just this magical thing "to think" that humans do.
BUT is human thinking really magical? There's a few basic tests that you can do with yourself. Pick a balcony, high up.
Would you die if you jumped off that balcony? Answer: yes.
Clearly thinking is not based on with trial and error.
Next: Would you die if I make you hold a 1kg block of 10% pure uranium for an hour? Most people will answer yes. The real answer is no. This won't affect you to any significant extent (sunbathing for 1hr is far worse), and there's people who live their whole lives on soil that gives off more radiation than that (there's such places in the DRC and in Iran, both have villages on top of that soil).
Or more dramatically: would you die if you jumped into the cooling pool of a nuclear reaction currently generating 3 Gigawatts of power? People always answer yes. The answer is: if you don't come within 10cm of the actual fuel rods you're actually safer in the cooling water than outside of the reactor. It is VERY safe. In fact, some inspections are done by divers.
Clearly thinking is neither rational, or at the very least, it's not working from first principles (because water radiation dampening for both gamma rays AND neutrons is enormous)
Can you teach people "wrong" reactions, "wrong" thinking based on feeding them bad data? Yes, in fact I feel like this is often on display.
You will find that human thinking IS imitation + extrapolation. Humans (and most animals for that matter) fundamentally work like that 99.99% of the time.
You can keep going like this and you will come to a conclusion that completely destroys your argument: your brain "thinks" ... based on previous data it's collected, on parallel processing and your brain IS a giant statistical model (read Bishop's book). Why would your reasoning apply to the algorithmic way of doing this but not to the "wetware" way of doing it?
I'm not sure where you got this mistaken idea that holding a block of uranium for an hour unlocks immortality, but I'm afraid "most people" got it right. The answer is yes. Absolutely you would die. You would also die if you didn't hold the block, but so too if you did.
Was this comment written by a chatbot? It is not at all in line with human thought.
Yes! The biggest usability issue with almost all the AI products I’ve used has been the termination problem, especially with fixed price offerings like ChatGPT. The LLM doesn’t actually have any time to “think” (however you might want to interpret that) except for the attention mechanism that visits each token.
I want to be able to query an llm, give it access to web search, and give it a time limit so that it keeps going, consuming sources a hundred pages deep, “thinking” and writing a report for me until it’s given a signal to terminate.
Or did I miss?
Continuing, we largely fool ourselves when we play the shell game of hunting for scientific justification of "social facts". Ie: when we assume a social fact, then hunt for scientific evidence to ground it in then use the scientific evidence to justify the existence and validity of the social fact we assumed. This happens a LOT and it's epistemically invalid.
A LOT of folks dismissal of thinking machines reminds me of Feuerbach's The Essence of Christianity, specifically the section on anthropomorphism[1], but in an inverted form - the reason LLMs can't "think" is that it would dilute what we hold dear about ourselves: our ability to think. It's an ego protection mechanism, but no one's jumping to admit that.
https://www.gutenberg.org/files/47025/47025-h/47025-h.htm#pb...
I think I follow better. Thinking doesn't have a rigorous and testable definition, so it isn't scientific. We all have our own colloquial understand. Since it's not testable and the definition varies person to person, it's not useful when reasoning about AI or intelligence.
Also I wanted to say I love how you write!
To explain it by analogy, we had the concept of women and men for at least thousands of years. Socially construed by physical appearance and behavior. Then in 1905 Nettie Stevens discovered sex chromosomes. Her discovery "found" the existence of men in women in the epistemological universe of science. But then a curious thing happened, people began placing people in social categories based on sex chromosomes. It's an incestuous loop of logic that hoists itself into validity simply because folks forget how it occurred.
If we went searching for a scientifically rigorous definition of "thinking" we would forget we made the same fatal leap of logic.
I suppose I have trouble seeing it that way because I view words as pointers to meaning. The word itself holds no intrinsic value; it's the meaning humans attach to it and the context in which it's used that gives it significance.
When we give "thinking" a scientifically testable definition we can reason around it in a more academic way.
I've read psych papers that defined "thinking" as cognition, measured by action potentials. A participant was said to be "thinking more" when their EEG showed increased activity.
It's important to understand that "thinking" here was only defined within the context of the paper.
To use an analogy, many people say their computer is thinking when it's undergoing a heavy processor load. Strictly speaking this is not the case, their computer cannot think, but they aren't wrong or incorrect because in that context 'thinking' literally means 'processing'.
Concretely, what do you gain by giving a current-generation LLM more runtime? It's not trained/designed to do anything with it, more time = more tokens = more nonsense once past the "end" token/end of context. You could build an agent on top of the LLM that calls the LLM iteratively, but afaict this approach isn't strictly an improvement over the base LLM. Current architectures seem to be limited with how much improvement they can eke out of more runtime without a broader redesign or retraining.
Now with new architectures you might be able to do more with more runtime, but I'm not sold that allowing a current-gen LLM to execute code or do web-searches and re-feeding it the quetion + its output + new data is going to strictly return better results. Sometimes maybe yes, sometimes not.
So imo the current limitation is not the runtime allotted to LLMs but their fundamental (current-gen) design/training.
We're not really discussing whether LLMs are "thinking" in the sense of the "social category" in the abstract sense - that's just the one word that this discussion got hung up on.
Concretely, we're discussing whether giving LLMs a larger time-budget would yield better results - OP suggested that letting LLMs iteratively refine or enrich an answer might lead to better results. The term "thinking" was used as an analogy or quick crutch to convey an analogy, not 100% equivalence, between a hypothetic LLM output evaluation algo and people's supposed "fast" and "slow" modes of thinking (if they even exist, whatever). Even if people don't have these two modes of thought, as you said they might be post-fact reifications of sociological ideas, the algorithm/idea of iteratively reflecting on and self-refining LLM outputs might still apply to LLMs as LLMs are not actually thinking.
Imo current LLM architecture and/or training data facilitate (mostly?) the automatic non-iterative mode but OpenAI is trying to do more of the iterative, reflective mode with O1 afaict. Imo and currently, there's not much to be gained by iteratively feeding an a current-generation LLM it's own output to iteratively analyze and augment it. At least not in general. Even enriching the input data via RAG doesn't always lead to better results. Imo, there's more fundamental work necessary - and maybe OpenAI is doing that.
I think the main advantage to slow queries would be in discouraging users from monopolizing the system by shotgunning many requests in a short period. Which is probably one of those cost cutting measures that costs you customers.
If I asked it to write me a block of code and it told me one hour, I’d just write it myself and tell my boss we aren’t really using the service.
Gary Marcus has been saying this since at least 2012:
> Yet deep learning may well be approaching a wall, much as I anticipated earlier, at beginning of the resurgence (Marcus, 2012)
To me it feels like having continually predicted rain throughout the longest dry-spell in history. Eventually you'll be right (assuming that we'll move to some other paradigm in due course), but it's hardly prophetic.
I think that the challenges of intelligence and non-shallow reasoning will inherently involve fighting against diminishing returns (which is what the supposedly broken scaling laws predict) regardless of technique. Same as for computer graphics - doubling the compute doesn't double the perceptual quality, but that doesn't necessarily mean it's hitting a wall.
I suspect we're in a similar situation with AGI.
This is a self-contradicting paragraph. It cites quantum mechanics understanding as breaking down Moore's law while ignoring the fact that it has implications on our so-called "laws" of physics in just the same way.
I can imagine the author having originally written "physics" instead of "gravity" but needing to re-write the paragraph to squirm away from the contradiction they set up for themselves. I feel that this presentation is not intellectually honest.
A lot of people really really really want this to fail.