That's because ChatGPT and the likes use machine learning to calculate odds of word combinations that make up a plausible sentence in a given context. There are scientific studies that postulate we'll never have enough data to train those models properly, not to mention exponential energy consumption required. But this is not the only application of this technology.
I haven't read this article, but the one place machine learning is really really good, is narrowing down a really big solution space where false negatives and false positives are cheap. Frankly, I'm not sure how you'd go about training an AI to solve math problems, but if you could figure that out, it sounds roughly like it would fit the bill. You just need human verification as the final step, with the understanding that humans will rule out like 90% of the tries, but if you only need one success that's fine. As a real world example machine learning is routinely used in astronomy to narrow down candidate stars or galaxies from potentially millions of options to like 200 that can then undergo human review.
The article is about using computers to discover new conjectures (mathematical statements that are not yet known to be true or false). The conjecture can be then later be formally proven (or disproven) by humans.
Sounds like a good match for me. Formulating conjectures is about finding an interesting pattern and argue that this pattern holds true. Computers are getting increasingly better at pattern matching, so why not use them?
No one is talking about automated theorem provers (see 4 coloring theorem) or symbolic solvers (see Mathematica). These tools already revolutionized math decades ago.
The only thing that came out in the past year or two are LLMs. Which is clearly overhyped bullshit.
Noam Chomsky said “we don’t know what happens when you cram 10^5 neurons* into a space the size of a basketball” - but what little we know is astonishing & a marvel
The 3D map covers a volume of about one cubic millimetre, one-millionth of a whole brain, and contains roughly 57,000 cells and 150 million synapses — the connections between neurons. It incorporates a colossal 1.4 petabytes of data.
Assuming this means the total data of the map is 1.4 petabytes. Crazy to think that mapping of the entire brain will probably happen within the next century.
This is exactly what I'm talking about when I argue with people who insist that an LLM is super complex and totally is a thinking machine just like us.
It's nowhere near the complexity of the human brain. We are several orders of magnitude more complex than the largest LLMs, and our complexity changes with each pulse of thought.
Back in the early 2000s CERN was able to simulate the brain of a flat worm. Actually simulate the individual neurons firing. A 100% digital representation of a flatworm brain. And it took up an immense amount of processing capacity for a form of life that basic, far more processor intensive than the most advanced AIs we currently have.
Modern AIs don't bother to simulate brains, they do something completely different. So you really can't compare them to anything organic.
far more processor intensive than the most advanced AIs we currently have
This is the second comment I've seen from you where you confidently say something incorrect. Maybe stop trying to be orator of the objective and learn a little more first.
2014, not early 2000s (unless you were talking about the century or something).
OpenWorm project, not CERN.
And it was run on Lego Mindstorm. I am no AI expert, but I am fairly certain that it is not "far more processor intensive than the most advanced AIs we currently have".
Citation needed on that comment of yours. Because I know for a fact that what I said is true. Go look it up.
Maybe you should be a little less sure of your "facts", and listen to what the world has to teach you. It can be marvelous.
I agree, but it isn't so clear cut. Where is the cutoff on complexity required? As it stands, both our brains and most complex AI are pretty much black boxes. It's impossible to say this system we know vanishingly little about is/isn't dundamentally the same as this system we know vanishingly little about, just on a differentscale. The first AGI will likely still have most people saying the same things about it, "it isn't complex enough to approach a human brain." But it doesn't need to equal a brain to still be intelligent.
It's demonstrably several orders of magnitude less complex. That's mathematically clear cut.
Where is the cutoff on complexity required?
Philosophical question without an answer - We do know that it's nowhere near the complexity of the brain.
both our brains and most complex AI are pretty much black boxes.
There are many things we cannot directly interrogate which we can still describe.
It’s impossible to say this system we know vanishingly little about is/isn’t dundamentally the same as this system we know vanishingly little about, just on a differentscale
It's entirely possible to say that because we know the fundamental structures of each, even if we don't map the entirety of eithers complexity. We know they're fundamentally different - Their basic behaviors are fundamentally different. That's what fundamentals are.
The first AGI will likely still have most people saying the same things about it, “it isn’t complex enough to approach a human brain.”
Speculation but entirely possible. We're nowhere near that though. There's nothing even approaching intelligence in LLMs. We've never seen emergent behavior or evidence of an id or ego. There's no ongoing thought processes, no rationality - because that's not what an LLM is. An LLM is a static model of raw text inputs and the statistical association thereof. Any "knowledge" encoded in an LLM exists entirely in the encoding - It cannot and will not ever generate anything that wasn't programmed into it.
It's possible that an LLM might represent a single, tiny, module of AGI in the future. But that module will be no more the AGI itself than you are your cerebellum.
But it doesn’t need to equal a brain to still be intelligent.
LLM'S don't work like the human brain, you are comparing apples to suspension bridges.
The human brain works by the series of interconnected nodes and complex chemical interactions, LLM's work on multi-dimensional search spaces, their brains exist in 15 billion spatial dimensions. Yours doesn't, you can't compare the two and come up with any kind of meaningful comparison. All you can do is challenge it against human level tasks and see how it stacks up. You can't estimate it from complexity.
You're missing half of it. The data cube is just for storing and finding weights. Those weights are then loaded into the nodes of a neural network to do the actual work. The neural network was inspired by actual brains.
I mean you can model a neuronal activation numerically, and in that sense human brains are remarkably similar to hyper dimensional spatial computing devices. They’re arguably higher dimensional since they don’t just integrate over strength of input but physical space and time as well.
Humbling? That’s going on in my head. I’m that complicated! Or at least the “hardware” I run on is. I think having a brain that beautifully complex is more empowering than anything! I wonder what new discoveries will stem from this.
Super humbling because nature's complexity can provide data storage and retrieval capacity several orders or magnitude greater than the best we can do right now.
Also super exciting because look at what every brain on the planet is composed of, and how it functions, in a freakin' square millimeter!
Firsly gilded means covered in a thin layer of gold in order to appear expensive, a one atom thick sheet of gold is not gilded, it is literally solid gold
nature.com
Hot