A programming contest has a problem where given N < 10000, do something hard like come up with the number of primes less than N
You can come up with all sorts of algorithms that do intelligent things. But the most effective solution is to use metaprogramming to make a massive switch statement that contains all the answers
Don't misunderstand: I'm happy saying AI models "think" or "have learned a thing", and for in-context learning I'd call them smart even by this definition…
…but also, any living creature that needed as many examples as machine learning currently needs, would starve to death before figuring out how to eat.
While training, machine learning processes (not just LLMs, also applies to e.g. self driving cars), are really really stupid and only make up for this by being really really stupid really really fast.
To what I wrote upthread: the "victories" of humanity over machine keep getting closer, but we have yet to wake up one day in great confusion as we find an entire city is no longer in communication with anyone, nor finding ourselves in a state of utter disbelief when the reports come in that the city stopped communicating because it is entirely gone.
If humans learned like ML systems learn, (biblical) Methuselah would still have been failing the Sally-Anne test on his supposed deathbed at 969 years old, like some of the smaller early LLMs did.
> It also doesn't really matter when "we are trained differently" has no direct bearing on the end result.
The question was to ask for a definition such that AI could still count as "not smart" compared to humans. This fits.
It's also why they're spiky intelligences, which I'm happily using right now to write code for me, but also do not trust in the slightest to identify the weeds in my garden. These submarines sure do swim fast*, but they're also very much disqualified for the Olympics.
There's a lot of innate knowledge but all neuroscience demonstrates how incredibly flexible the brain is. Brains constantly learn and rewire.
Here's a few things that I think show how crazy it is AND stress those points
- people that have had corpus callosotomy (brain cut in half) *may* be indistinguishable from a normal person. Depends on how young you were when you underwent the procedure
- true for most brain injuries
- can even include the frontal cortex
- you can learn to ecolocate
- people with Aphantasia are indistinguishable from others
- people without an internal monologue are indistinguishable from those with one
- people can learn to use prosthetics
- even without disabilities
- or look into MRI scans with tool use
We are constantly learning. You aren't just born with your knowledge and it stays static. We are extremely proficient at metalearning (learning how to learn, few shot learning, zero shot learning [0,1]). Our brains are constantly rewiring, able to heal from traumatic damage.
I could go on and on. Does information pass down through genetics? Of course! But that's far from the whole story.
I'm tired of people trying to make AI sentient by making humans robotic. Stop trying to trivialize everything and be okay not knowing the answer to everything. You're human, you're designed to learn and explore, not sit and argue from an armchair
[0] and I mean these in the original sense. Not in the sense that you train on a billion examples of labeled animals and then congratulate yourself on your ImageNet-1k held out test performance. That's not zero shot, that's just a test set
[1] I can literally make up words and you'll understand them. Or use words in novel ways. That's literally how slang works and how new words come to be. Don't be a walibanut ya glufus. Read some SciFi
Most of the effort of evolution was making cells work at all, and even then it's a bit weird, e.g. no plant or animal produces vitamin B12 and we all get this from some bacteria and archaea.
And evolution is kinda hard to time right: bacteria can reproduce in minutes, humans in decades, but only mutations that survive reproduction can be passed on. This makes it even starker as a difference: bacteria had order of 1e13 generations to become multicellular, while human DNA had about 40,000 generations to cope with fire, 220 generations for evolution to do anything with the invention of the wheel, and one generation to cope with the invention of Minecraft.
The analogy here would be: DNA is to our brains like a VN replicator bootstrapping a computer all the way up to a bare-metal-no-OS untrained model, and perhaps a few crude "hard coded" modules like a smiling-face-detector. It's a lot, but it's also missing a lot. If biology used the models and training processes that are state of the art in ML, it would take around a millennia to talk like a child and still fail the Sally-Anne test, and million years or so to pass a degree.
I'm still going to deny the premise of your argument, becasue I think we should define intelligence in terms of capabilities. If a system can discover a cure for cancer or solve P vs. NP, it doesn't matter how many FLOPs it took to train.
A 1 gigabyte LLM isn't going to impress anyone with what it can do.
About 99% (depends who you ask) of our DNA is shared with our nearest primates. Like us, they can learn to use touch screens, but also like us they won't find touch screens in their natural environment. Dogs can be taught to drive cars (just about), but again, not natural environment.
> I'm still going to deny the premise of your argument, becasue I think we should define intelligence in terms of capabilities. If a system can discover a cure for cancer or solve P vs. NP, it doesn't matter how many FLOPs it took to train.
We can define it in either way. I think both are valid, because plenty of people mean each of these two things when discussing AI in particular. As I referenced in the other branch, these submarines sure can swim fast.
But at the same time, they have a lot of gaps. This is because some experience needs the real world: just as nine women can't make a baby in one month, a transistor running a million times faster than a synapse can't make a month-long cancer experiment happen in 2.6 seconds.
This dependency on data, and that state of the art ML is bad in specifically this way, is why Tesla's self-driving cars, despite having had around a trillion miles of real-world experience today, still come with steering wheels (even at least some of the Cybercabs, despite the big thing of this model supposedly being not needing them, though with Musk and his promises you should only count the Cybercabs when they actually ship and not just press releases).
Imagine an alien that matches your abilities across every domain, but has a 10 billion year training period, something many orders of magnitude more expensive than an LLM. I simply don't believe that alien is less intelligent than you.
We also don't expect humans to be competent in every domain. Most humans suck at most things. We will usually call someone intelligent if they excel at solving problems in one or two narrow domains.
> 10 billion year training period, something many orders of magnitude more expensive than an LLM.
I'm saying both definitions are valid definitions, they both point to important and different things: skill now, vs. how hard it is to get new skills. Some would describe it as "crystallised intelligence vs fluid intelligence".
I think it's important that any arguments are over the thing in dispute, not the label for that thing. Don't mistake the map for the territory.
Anyone who says "AI is stupid" by the first definition, what it can do, I think is making an error: they are already wildly super-human in at least some areas, if not generally.
Anyone who says "AI is stupid" by the second definition, how many examples they need, I agree with: there is a lot they are not currently able to learn even though it is easy for us, because the data they would need to do the learning on does not exist at the scale they need.
Also note: examples, not years. An alien intelligence whose synapses trigger 10 times faster or slower than mine (or ten million times faster or slower than mine), but who gets as much as I do out of each book or conversation, is my equal by the second definition.
I don't think I agree with your characterization of the second definition. Time scales matter. It's not much use to be able to solve human-scale problems if it takes millennia. And it only takes months to train an LLM to the level that it can solve cutting-edge math problems.
Aye, for practical purposes; but this gets you crystallised intelligence. I'd be happy to say e.g. the Chinese Room has crystallised intelligence. But humanity invented fire before reaching the anatomically modern form, and even anatomically modern humans collectively took hundreds of thousands of years to invent durable writing with which the room in the Chinese Room thought experiment could be filled.
It was around a million (or so) years from fire to having enough shared cultural knowledge to be able to formulate the cutting-edge math problems that LLMs can now solve.
Human fluid intelligence means we can pick up deep shards of this accumulation of wisdom, find new avenues of novel research to poke at, all within 40 years, even despite the depth and breadth of work from all the other humans who came before.
(Though this also points at another way to be "superhuman": breadth. Many hands make light work, as the saying goes, and a lot of different humans solving different puzzles at the same time is part of how we got so good so recently even though ~10% of all humans who ever lived are currently still alive; and the same for AI was (accidentally) also part of how the OpenAI-HuggingFace incident went down).
AI (not only, but also, LLMs) are very useful, and I'm getting value from using them. But the fluid intelligence of machine learning* is very poor, and the only way they have to make up for this is by being very fast**, but when there's not enough to train the AI on, they get stuck at a very low plateau.
* possibly the architectures, but I suspect the process by which AI weights and biases are set, and again I don't mean just LLMs
** the speed difference between a transistor and a synapse is about the same as the speed difference between a jogger and continental drift
That seems like a really bizarre way to describe a tool that solved an open Millennium Prize Problem. They are, empirically and repeatedly, ahead of the status quo.
So if your argument depends on them being behind the status quo, reality has already disproven it multiple times over.
I will admit the first time I read the thing you're replying to, I had a similar thought as you; From the sibling reply from them, I think they think they were obvious, but that also means I wouldn't expect their reply to help unless you had the same flash of inspiration I had.
Definitions are "formal statements of the meaning or significance of a word, phrase, idiom, etc" (https://www.dictionary.com/browse/definition)
> They are necessarily behind the status quo.
The existing state or condition would be what is written in the dictionary, not whatever personal definitions you've constructed.
> "You can't call this newfangled contraption a computer, because a computer is a person!"
Seems like a straw man. A computer is not a mammal, no matter how much you twist a set of definitions.
I’m not the person you replied to, but I believe they’re referring to the occupation of “computer”:
https://en.wikipedia.org/wiki/Computer_(occupation)
So yes, at one time all computers were mammals.
The people who write dictionaries generally take a descriptivist approach, that’s why slang terms enter the dictionary after they start to become popular.
The state of the art of human knowledge would be another step ahead of the common use of any language.
E.g. humans get exposed to new LLM model - yeah its powerful - 1 week later - eh, that thing? Yeah it's whatever. I'm still employed.
The human's ability to adapt so efficiently is mind-boggling - so much so it pi1sses sam altman and dario off.
When AI does it we call it “reward hacking” but when humans do it we call them clever.
OK, they can play chess, but that's not real AI - can they write poems? OK, they can write poems, but that's not real AI - can they compose music? OK, they can compose music, but that's not real AI - can they translate languages? OK, they can translate text, but can they do maths? OK, they can do maths, but can they solve a Millenium Prize? <-- we are here
"What? Don't be silly. For one thing, trees have moss."
"OK he's grown moss. He's a tree now right? Right??"
"I doubt it, for I see nothing but wishful thinking to suggest that simulating the appearance of tree characteristics is part of a path to becoming a tree. And that's not actually indistinguishable from moss anyway, is it?"
"Urgh, classic goalpost shifting!"
“I once met a person who could beat any grandmaster in chess, translate any language, and complete international math Olympiad problems. He couldn’t solve any Millenium problems though, so I’d say he was a midwit at best.”