Programming thread

[Slav Power]

LLM's are not working like a human brain and I'll never believe otherwise. Datacenters are multiple times more powerful than a human brain, yet they can't achieve true superintelligence. Why? LLM's are fundamentally not like a human brain. They try to imitate one, but the only point of reference it has are similarities between words. They are capable of a lot despite being so limited, but you'll find even the biggest, most powerful models making the same mistakes no human ever would. Why? They're still LLM's, and frankly, we are reaching the limits of what we can accomplish with such an architecture.

We need even bigger datacenters using even more power that we cannot generate, hence why there are warehouses full of unused hardware like the ones Satya Nadella owns, and clearly it's the wrong way to go about things. It's fine right now when the stock market speculation bubble is running wild so you can build infinite datacenters, but the reality of it all will hit this entire market, and many are seeing these limitations already. Hence why there's more effort in managing to achieve the same capabilities and accuracy of LLM's at a lower hardware cost, reducing VRAM usage, optimizing token generation speed and so on. But IMO the true breakthrough that's needed is not compacting Fable/Astra/Sol/Civic/Espero into a smaller package that you can run on a single Mac Mini, but another breakthrough akin to that caused by Attention Is All You Need that kickstarted it all.

Most importantly, the whole craze has to pop, people have to realize it's not "AI" or "AGI", it's LLM's, ML models, clever but dumb pieces of software that cannot replace an entire workforce overnight, but tools that can improve the existing workforce if they're used correctly. Everything bad about LLM's stems from people overestimating their capabilities and overrelying on them, and that's directly fueled by the current stock market grift.

 

[Silver Coin Anon]

What are YOU doing with it? What are you pointing at that someone else did?

LLMs are great at spitting out logistical solutions via software, because all software abstracts the functions of physical hardware. Move bits from pin 3 of TI12345 to GPIO17 and send via asynch webserver to a python database api. Throw out a OpenGL graph. The tools are popular and therefore have many examples of being programmed for. The solution of utilizing the hardware was solved for when TI wrote the chip datasheet. The data sheet of the dev kit, the chip, the capacitor you use for power filtering? Data sheet, accounted for, described and solved for. Object A inputs B and C and outputs D. A, B, C, and D are described and perfected. LLMs know how to move A to C through modifier B.

Programming languages are, at the end of the day, tokens to tell hardware what to do, and are therefore quite solved for by token-based LLMs. They are all logistical block charts. If you can block chart it, and LLM has already solved it. Most programming is thus.

Tokens do not describe real world physical phenomena that occur over complex time domains and which involve novel interactions between real physical samples as measured by newly applied sensors never meant for that purpose. LLMs do not fill in the blocks, they only use human words to abstract the blocks. For me in this one example, no human has written about this and therefore the LLM can provide no useful information. There are no block charts because this system has not been solved for. Perhaps I, reading the statisically useless words, can be INSPIRED, yes. But that is me using a tool, not the tool accomplishing anything novel. The source of novelty comes from humans. Statistical models are incapable of novelty. Permutations and factorials are unique but not novel; I don't care how many universes it takes to deal a deck of 52 cards, each order is known and solved for. I haven't solved for how to make [Redacted] maintain a stable [Redacted] by using a [Redacted] while also maintaining control with the [Redacted]. I ask the LLM and you know what it tells me? "Wow, how unique! You just need to use [redacted] to maintain a stable [redacted] by controlling [redacted] while using [redacted]!" It can spit to me the blocks in every permutation possible but still the samples sit there.

I then have to fill in the blocks myself because reality is not a block chart and my brain is vaster and more complex than every single LLM on Earth combined to some order of magnitude because my brain has mass flow, electrogalvanic potentials, and nth order geometric complexy, a soul, and I can drink Monster and Fireballs. But it can write me a great sorting algorithm it learned from a block chart summarized by a human to describe a human invention. Learn to code and learn to solder, kids. Inhale the fumes.

 

[California Newt]

Books? You presume those are the originals? Do you not think that those books are written after seeing other developers write code, or do you think that authors come up with their own work? If you're merely borrowing ideas from books, are you not a plagiarist yourself, or is this a special exemption that only exists for books?

It's all copies. Your brain works exactly like an LLM. The way they train the weights is through something called artificial neurons. We call them weights because of the biological synaptic weights.

Human creativity is blending up their own experiences. If you try to think of something random, you'll just reference something weird, not something novel.

Retroactively refuted by Searle's Chinese Room thought experiment and the halting problem:

Halting problem:
Algorithms can't intuit recursion. You know that old paradox of evaluating the truth or falsehood of the statement "this sentence is a lie"? How it goes back and forth forever ("If it's true, that means it's a lie and therefore false; if it's false, that means it's not a lie and therefore true—ad infinitum")? A computer can't figure that out just by looking at the problem, and it's because they just follow sets of instructions. We dont. You could write a program to recognize one or even a big set of paradoxes, but you can't give it the faculty that lets it understand when a problem leads to an infinite regress from the outset. Machines can't intuit.

https://youtube.com/watch?v=92WHN-pAFCsChinese Room Argument:
You're put in a room with a slot in the door and book on the table. A Mandarin speaking Chinese man hands you something in Chinese through the slot. The book contains written instructions for converting Chinese input into Chinese output: you follow these instructions and hand the Chinese man a response through the slot. To him, it reads like he and you are having a natural conversation in Chinese. Does that mean you or that room have actual knowledge of the meaning of what's being said? Of course not; you're just manipulating symbols. But that's all a computer does, which means that the mind can't be a computer.

Computers have no understanding of semantics, and if you say that humans don't either you're entering into a logical contradiction. You can't make the truth and knowledge claim that truth and knowledge are impossible to attain to. For a quick rundown of the collapse of Western autonomous epistemology (and a proof of God, which the essence-energies distinction completes by explaining how a transcendent Deity can be immanent in the world), see the attached "The Contingency of Knowledge and Revelatory Theism". There's also a lot of back and forth about this stuff on my profile.

That an LLM has some structural similarities to the brain does not mean that the mind is reducible to the brain. The mind isn't reducible to matter at all. Reductive physicalist positions within philosophy of mind have been laughable cope-fests ever since the middle of the last century, despite what Epstein associates like Daniel Dennett would lead one to believe (Edward Feser has a good takedown of Dennett's philosophy in his philosophy of mind textbook referenced in the above-quoted post; I'm not a Thomist, but Feser is good on this specific point).

Here's that Halting Problem video, since I think the thumbnail is cool and brightens up any thread it's in:

 

[mnt_ms]

I used to sometimes meet people who insisted computers with their fancy new "word processing" would never replace their trusty typewriters. Typing on a screen could never give writing the same soul as you'd get from clattering away on an old Olivetti.

That's exactly how anti-LLM people sound to me now.

God forbid that people should use a computer to automate tasks.

Good analogies are rare and this one is particularly bad. There is a huge difference between task automation and outsorcing your thinking.

How about this analogy. Would you like to have an artificial indian to perform your tasks? He won't stink, he won't shit, and he won't have no izzat, but he is one dumb son of a bitch that needs to be monitored every step of the way.

 

[Slav Power]

Tokens got bastardized into a marketing buzzword by the industry so it's best to stick to the technicals. Tokens are just words, or sentences or groups of words. The text input is broken up into tokens so that they can be understood by the model, then the task of the model is to figure out which tokens would be the closest to the ones given to the model, and thus the prediction process happens. The "thinking" process is a hack that was discovered on /v/ back in 2020, if you let the LLM "talk with itself" before returning an answer, it'll greatly improve the accuracy of the answer, and that's the biggest revolution in LLM's.

Of course, all of this works perfectly for analyzing and generating code. Code is very semantical, consists of words, and that's what LLM's architecturally excel at. It's important to understand this on an architectural level as that's how you'll learn what you're interacting with, what it can or cannot do, and how you can use it for your use case to achieve the best possible result. But don't be fooled by all the marketing bullshit, LLM's are not AI, there is no AGI, we are talking about machine learning models and tokens are nothing special, they're a well-documented way of how an LLM operates. "Token factory" is nothing more than a server running LLM's. "Tokens per second" is how many text predictions an LLM can make per second and it's strictly an LLM trait.

My biggest gripe though is that the entire market is hyperfixated on LLM's and only LLM's. It causes a tunnel vision effect where the market at large ignores many useful model architectures that arose from the AIAYN revolution, or even the fact that other model architectures could be a better fit for a given task than LLM's. Hell, even the current big LLM's don't actually generate images. They do have the ability to analyze images, but any form of image generation or image editing is realized by hooking the inputs and outputs of the LLM to a diffusion model, then hiding it all from the user in the frontend, which causes a false assumption that it was the LLM that made the image. The more you understand how this technology works, the more you can get from it.

 

[ADHD Mate]

All LLMs haters are refuted by all those amazing games and new software that have been released since Fable.

You know we all got that 10x boost. Or maybe even 100x like DHH said.
I cannot get through pile of all those amazing games that previously would take 4+ years to make. Yet now with Claude power it only takes 4 months!

All those experienced devs, turned LLM whisperes, that have been laid off recently, will now produce immeasurable amount of god tier games previously unknown to mankind.

Truly insane progress.

 

[Null]

That an LLM has some structural similarities to the brain does not mean that the mind is reducible to the brain. The mind isn't reducible to matter at all. Reductive physicalist positions within philosophy of mind have been laughable cope-fests ever since the middle of the last century, despite what Epstein associates like Daniel Dennett would lead one to believe (Edward Feser has a good takedown of Dennett's philosophy in his philosophy of mind textbook referenced in the above-quoted post; I'm not a Thomist, but Feser is good on this specific point).

The core problem with being "aha, checkmate Claude" using this video is the dressing. The system of machines can be substituted for anything, including the human brain, including systems of human brains. Nothing can solve all problems by itself. This includes human beings and complex systems involving human beings. It does not demonstrate that computers cannot do more than humans can.

This also demonstrates a double-standard that has been rising over time. It used to be that AI couldn't do any code. Then it couldn't do anything more than functions. Then it couldn't do anymore more than single classes without handholding. Then entire projects. Now kids with no programming experience are one-shotting engine conversions with a $20 subscription in an afternoon. Now it's "Claude can't replace a human workforce".

AI is winning, but it's winning so hard that we're forcing it to play by different rules to cope with our own obsolescence. No human can compete with Claude as it exists. But in the same way that the DMV and FAA abide by different regulations, we expect no problems with aircraft while forgiving horrific drivers continuously. Claude is not just beating humans at their games, it's beating humans at a game with much higher standards, and then are comparing its every-day output with the highest standards we could never apply to a random developer.

 

[Slav Power]

The real problem is that Anthropic is a chokehold. People are tying their workflows to a product of a company whose CEO looks like someone straight out of a A. Wyatt Mann cartoon. Is Claude amazing? Yes. But Anthropic knows that as well and are acting like drug dealers. They'll hand out free samples to developers to make them dependent on Claude, so that they'll have to pay endless subscriptions and forget how to work with code without Claude. Oh, and they'll arbitrarily switch which models you're using without your knowledge so you'll be paying for a model you won't be able to use and there's nothing you can do about it because Anthropic is a private company that can do whatever it wants, Dario just played you like a fiddle if you haven't made a plan B for when he pulled this on you.

Anthropic knows this and that's how they're fucking over everyone in the long run. Ideally we'd see a frontier level open weight model in 2-3 years that would alleviate this dependency issue, but it's probably not going to happen any time soon. Still, it's the only reason why I'm hesitant to use LLM's. Unless I can run it on my hardware, I don't want to become attached to it. There is nothing worse than depending on someone else's computer you can't see into for your own personal computing. This isn't even a matter of open source, this the freedom 0 of free software, the freedom for you to run the damn thing, open or closed. Right now everyone is getting conditioned to abandon this freedom thanks to how great Claude is, and once again, convenience will be the trojan horse of taking away your freedoms.

 

[California Newt]

The core problem with being "aha, checkmate Claude" using this video is the dressing. The system of machines can be substituted for anything, including the human brain, including systems of human brains. Nothing can solve all problems by itself. This includes human beings. It does not demonstrate that computers cannot do more than humans can.

Your specific claim was that the brain works exactly like an LLM.

Human creativity doesn't work by blending experiences in the way that an LLM does. Human creativity comes from intuiting the principles and ethos behind created things, which is something that LLMs are categorically incapable of (due to not being able to intuit, period).

An LLM cannot solve the kinds of problems that require intuition. Networks of people solving these kinds of problems (this describes all culture, growth, and genuinely mutually creative interaction) are what give life value. A homeless guy asking you for two bottles of brake fluid so he can finish restoring his broken-down pickup truck is worth more than an infinite chain of LLMs, because the process that that man is engaged in is what technology exists to foster.

 

[Null]

Your specific claim was that the brain works exactly like an LLM.

Human creativity doesn't work by blending experiences in the way that an LLM does. Human creativity comes from intuiting the principles and ethos behind created things, which is something that LLMs are categorically incapable of (due to not being able to intuit, period).

An LLM cannot solve the kinds of problems that require intuition. Networks of people solving these kinds of problems (this describes all culture, growth, and genuinely mutually creative interaction) are what give life value. A homeless guy asking you for two bottles of brake fluid so he can finish restoring his broken-down pickup truck is worth more than an infinite chain of LLMs, because the process that that man is engaged in is what technology exists to foster.

Okay, it doesn't work exactly like a brain, it works better than a brain. The entire art of modern AI development is giving models enough knowledge to do exactly what you're saying they're incapable of.

I'll remind you that Claude correctly INTUITED from a conversation that I was talking about the Kiwi Farms and shut down trying to help me because of it.

You're basically falling back into a comfortable blanket of "Humans are flawed and that's why we're so beautiful UwU" which is a fucking fairy tale to help people sleep at night. The average person is ugly, stupid, and incapable. They don't have intuition either. I see them every day. I see them come into my country by the million year over year. I prefer the computer.

 

[California Newt]

Okay, it doesn't work exactly like a brain, it works better than a brain. The entire art of modern AI development is giving models enough knowledge to do exactly what you're saying they're incapable of.

I'll remind you that Claude correctly INTUITED from a conversation that I was talking about the Kiwi Farms and shut down trying to help me because of it.

You're basically falling back into a comfortable blanket of "Humans are flawed and that's why we're so beautiful UwU" which is a fucking fairy tale to help people sleep at night. The average person is ugly, stupid, and incapable. They don't have intuition either. I see them every day. I see them come into my country by the million year over year. I prefer the computer.

Claude is only able to "intuit" and "make inferences" due to the massive amount of context that its been given. I'm talking about solving problems without that kind of pre-established context. If AI researchers are trying to solve the problem by giving models "enough knowledge", then they're sidestepping rather than actually solving it. These are issues in the field that undergraduate students were made to understand the better part of a decade ago.

The vast majority of people learn how to speak and how to understand the semantic content of sentences. This is already more than any LLM will ever be capable of, per the Chinese Room problem. That may not be up to your personal standard, but it doesn't matter: if computers can't clear the low bar I've set, they'll never reach yours.

My argument isn't that human beings are good because they're flawed. My argument is that human beings have faculties that are able to ground value to life at all; computers only exist to augment that process. If you actually disagreed with that, you wouldn't respond to mine or anyone else's messages when character.ai is right there (unless it's defunct; I didn't check).

 

[Null]

Claude is only able to "intuit" and "make inferences" due to the massive amount of context that its been given. I'm talking about solving problems without that kind of pre-established context. If AI researchers are trying to solve the problem by giving models "enough knowledge", then they're sidestepping rather than actually solving it. These are issues in the field that undergraduate students were made to understand the better part of a decade ago.

You're describing the Singularity: self-improving machines. They're working on it.

 

[California Newt]

You're describing the Singularity: self-improving machines. They're working on it.

https://youtube.com/watch?v=3n7eNFj_9Vk

Are you able to paraphrase in your own words what my argument is? The significance of the halting problem and Chinese Room argument?

Edit: great video though.

 

[Null]

No. I can ask Claude to do it though. My point is I don't care. All I care about is cheaper, faster inferencing and better outputs. And every brownoid to fuck off. Nothing else matters to me. I'll let the guys paid millions of dollars a year in equity to fuzz out all the problems. I will never be able to rely on human developers, and I have never been able to rely on human anything, so I prefer Claude.

 

[BarberFerdinand]

Okay, it doesn't work exactly like a brain, it works better than a brain. The entire art of modern AI development is giving models enough knowledge to do exactly what you're saying they're incapable of.

I'll remind you that Claude correctly INTUITED from a conversation that I was talking about the Kiwi Farms and shut down trying to help me because of it.

You're basically falling back into a comfortable blanket of "Humans are flawed and that's why we're so beautiful UwU" which is a fucking fairy tale to help people sleep at night. The average person is ugly, stupid, and incapable. They don't have intuition either. I see them every day. I see them come into my country by the million year over year. I prefer the computer.

I don't really follow this thread and am uninterested in the overall drama, but I did have a question about your perspective @Null. I saw your uspg2 comment about politicians selling out to dario. And you usually speak very well of claude specifically. With your current experience and perspective, do you feel as if anthropic has a secret sauce that other companies won't be able to replicate? It seems to a casual onlooker over the last several years that things are dynamic and ever changing. Wills anthropics sWillcy last, do you think? Is it possible that the open source crowd stands a snowballs chance of at least somewhat keeping up? Asking out of genuine curiosity, since you seem familiar with the subject matter.

 

[California Newt]

No. I can ask Claude to do it though. My point is I don't care. All I care about is cheaper, faster inferencing and better outputs. And every brownoid to fuck off. Nothing else matters to me. I'll let the guys paid millions of dollars a year in equity to fuzz out all the problems. I will never be able to rely on human developers, and I have never been able to rely on human anything, so I prefer Claude.

That's a shame. It's a cool problem.

 

[Null]

I don't really follow this thread and am uninterested in the overall drama, but I did have a question about your perspective @Null. I saw your uspg2 comment about politicians selling out to dario. And you usually speak very well of claude specifically. With your current experience and perspective, do you feel as if anthropic has a secret sauce that other companies won't be able to replicate? It seems to a casual onlooker over the last several years that things are dynamic and ever changing. Wills anthropics sWillcy last, do you think? Is it possible that the open source crowd stands a snowballs chance of at least somewhat keeping up? Asking out of genuine curiosity, since you seem familiar with the subject matter.

If I knew the answer to this question I would be a very different person living a much better and more secure life.

 

[DavidS877]

I just ran out of tokens, and it's all null's fault.
{"headers":{"Retry-After":"120"},"provider_name":null}
I'm actually sort of wondering if Sol is just fucking with me as I've asked it to fix vLLM for the B70 cards and it's already eaten $40.

 

[Sneedfully and Peacefully]

I'll remind you that Claude correctly INTUITED from a conversation that I was talking about the Kiwi Farms and shut down trying to help me because of it.

And it doesn't remember this when you start a new session? Or do you have to barter and say "no really I'm just a simple goyim with this very unique problem!" to get it to output again?

 

[Niko]

I'll remind you that Claude correctly INTUITED from a conversation that I was talking about the Kiwi Farms and shut down trying to help me because of it.

couldnt this be because the thing remembered a previous conversation or the code you provided had some line that gave the name away