AI progress...OAUP vs RL

[inkoalawetrust]

Oh yes, the SI article is shockingly spot on about modern machine learning despite being written in 2002. I'd like to write an article making "expert systems" (not as in the original term for 80s LISP machines and later programs like the ones they ran) a more explicitly canonical category, and as part of that I'd like to keep as much of the SI article intact and merged. Also not sure what you mean that there's no mention of the AI boom, the 2020s section literally starts by talking about it and it occurring through the 2020s lol.

I am sorry but I just don't see it there, see screenshot.

That's very weird, what article are you on? The information age timeline looks totally different and uses the Tranquility Calendar.

Re Qwen 3.8 27B: 5 months since I posted that, RAM prices have gotten even higher (after GPUs already did the same thing months before that, and then SSDs), making it even more impossible to run AI locally (with what RAM are you going to page the weights?). I asked Google AI Mode (lol) and it says that the minimum hardware to run Qwen 3.8 27B (without resorting to lobotomizing it through extreme quantization) is a fucking RTX 3090 or a 4090, do you know how expensive that is? Last I checked used 3090s were like 700 euros, when 4090s were still available here last year they costed OVER TWO THOUSAND EUROS.
Even going for AMD GPUs the cheapest is 900 euros (and you'd need to concede even more performance for the model to run it), with the ones that can fit it in VRAM like the 7900 XTX also being in the 1000+ euro range. And I'm handwaving the stranglehold Nvidia's CUDA has over ML anything. And again, you can forget offsetting things to RAM (which creates a dramatic slowdown even if RAM was affordable still) if you didn't have at least 32-48GB of high speed DDR4 or DDR5 before like last year.

I literally ran 4-bit quantization 27b models in Alpaca on a midrange i5 PC from 2021 with a RTX 3060. It was slow, took about an hour to generate a reasonable Arkanoid clone (didn't try 3.8 yet, this was Qwen 3.6 27b) but bearable. Granted, I do have 64 GB of DDR4 that I bought before. No weird tricks tho, Alpaca is literally just a buggy GUI Ollama front end. Have you actually tried to run a local model?

Also side note on this years' chatbots beating last years' chabots. Those benchmarks are all useless (for many reasons, like benchmaxxing, doing the benchmarks with more expensive models than the ones served, the benchmarks sometimes being written with the direct or indirect involvement of the AI companies themselves, etc), they have not actually improved in anything besides being fed more data to memorize since like, Q* (chain of thought, and apparently even that is just memorized from people hired to write their thought processes!), everything else is mainly inference optimization.

You are so absurdly wrong and clearly muddled with hatred for "AI tech bros" it's not even funny, I don't go by benchmarks, I actually use those agents, this year I made a climate sim in 3D while last year the best it could do in August 2026 were simple 2D games more primitive than SNES.

"Scam Altman".

You aren't criticizing the technology, you hate the people behind it, and it by extent.

Re local models: I'm on an RTX 2060 with 16GB of RAM and a Ryzen 5600X. I did run several Stable Diffusion checkpoints for a white with A111 but it eventually got incredibly unstable with constant OOM errors (IIRC) and crashing my GPU drivers intermittently. I haven't ran any local LLMs though since I just care about image and video (lmao that one is super not happening any time soon at any level of quality).

Re benchmarks and AI performance and stuff: Well, I'm actually glad you don't go by the benchmarks. But from my experience even with the "Better" models, they aren't that much better, and it certainly isn't being helped by how me and everyone else is indirectly experiencing LLM code (i.e service outages). Also uh, not sure how to put this, but the climate sim has and had some pretty egregious bugs like the inverted rendering, inverted mouse controls and extreme sensitivity, and the time controls not working. Which doesn't make using LLMs for code totally futile but is a spanner in the works for Altman and Amodei selling Mass Unemployment as a Service.

Re Altman and AI criticism: As I've said and I think have shown multiple times in this thread to, I use genAI models pretty damn often lol. From the occasional proofreading, to translation, to asking it questions (i.e minor stuff up to asking them about the feasibility of certain things), to generating images etc. And I also used to run SD locally and one of the reasons I'm annoyed at the insane hardware price inflation is specifically because it makes it extra impossible for me to run anything remotely serious locally (i.e I'd like to actually be able to run Qwen 3.8 27b or MiniMax H3 locally!).

As for Altman, he is just the face of OpenAI that handled the cutthroat sillicon valley business crap and then turned the company into a for-profit boondoggle that'll star in the next global recession and has already starred in trying to keep AI closed once he caught a whiff of money. And I can't overstate how glad I am we aren't in 2023 anymore, when there were basically no serious downloadable LLMs, and the only option for other AI media that you could use was SD.

Also, you are thinking of Ilya Sutskever, that was the guy behind GPT and OpenAI's rise, too bad he left after Altman returned into OpenAI's board like a villain in a movie sequel. Because it was Ilya that not only was behind creating GPT, but was also one of the co-creators of AlexNet, you know, the thing that marked the end of the 90s-00s AI winter. And because Ilya has actually done shit in AI instead of JUST endless catastrophizing and "tee hee" investor pitches (like "warning" about mass unemployment, or LLMs killing us all, or humans becoming obsolete entirely, or GPT being able to raise his fucking kids), I've also basically never talked about him negatively too, I was even actually hopeful about SSI, and my most negative take on him is just a lukewarm "I think SSI might just release an LLM" not "he is responsible for all the AI doomposting, employees getting LLMs rubbed in their faces, the sloppocalypse, etc" like Dario and Altman are.

Anyway, again, I think it's kind of obvious I don't hate the technology THAT much when I know more about it than the people you are implicitly lumping me in with, and because I use it enough that it'd make anyone that is actually super against it blow a head gasket.

Regarding math and AI, people have solved decades old conjuctures at this point by basically telling the AI to solve it:

https://vxtwitter.com/DmitryRybin1/statu...5652893709
https://chatgpt.com/share/6a60b2eb-0b64-...31ca1de063
https://isa-afp.org/entries/Dinitz_Garg_...ample.html

Formal verification.

https://zenodo.org/records/21554259

Independent exact verification.

Sorry but no, this is not traditional narrow AI.

I never said it's traditional narrow AI though lol. I should've been more clear that the neural nets we build nowadays are a lot closer to what we usually think of when we hear AI than symbolic AI was.

[Tostar]

Oh yes, the SI article is shockingly spot on about modern machine learning despite being written in 2002. I'd like to write an article making "expert systems" (not as in the original term for 80s LISP machines and later programs like the ones they ran) a more explicitly canonical category, and as part of that I'd like to keep as much of the SI article intact and merged. Also not sure what you mean that there's no mention of the AI boom, the 2020s section literally starts by talking about it and it occurring through the 2020s lol.

I am sorry but I just don't see it there, see screenshot.

That's very weird, what article are you on? The information age timeline looks totally different and uses the Tranquility Calendar.

Re Qwen 3.8 27B: 5 months since I posted that, RAM prices have gotten even higher (after GPUs already did the same thing months before that, and then SSDs), making it even more impossible to run AI locally (with what RAM are you going to page the weights?). I asked Google AI Mode (lol) and it says that the minimum hardware to run Qwen 3.8 27B (without resorting to lobotomizing it through extreme quantization) is a fucking RTX 3090 or a 4090, do you know how expensive that is? Last I checked used 3090s were like 700 euros, when 4090s were still available here last year they costed OVER TWO THOUSAND EUROS.
Even going for AMD GPUs the cheapest is 900 euros (and you'd need to concede even more performance for the model to run it), with the ones that can fit it in VRAM like the 7900 XTX also being in the 1000+ euro range. And I'm handwaving the stranglehold Nvidia's CUDA has over ML anything. And again, you can forget offsetting things to RAM (which creates a dramatic slowdown even if RAM was affordable still) if you didn't have at least 32-48GB of high speed DDR4 or DDR5 before like last year.

I literally ran 4-bit quantization 27b models in Alpaca on a midrange i5 PC from 2021 with a RTX 3060. It was slow, took about an hour to generate a reasonable Arkanoid clone (didn't try 3.8 yet, this was Qwen 3.6 27b) but bearable. Granted, I do have 64 GB of DDR4 that I bought before. No weird tricks tho, Alpaca is literally just a buggy GUI Ollama front end. Have you actually tried to run a local model?

Also side note on this years' chatbots beating last years' chabots. Those benchmarks are all useless (for many reasons, like benchmaxxing, doing the benchmarks with more expensive models than the ones served, the benchmarks sometimes being written with the direct or indirect involvement of the AI companies themselves, etc), they have not actually improved in anything besides being fed more data to memorize since like, Q* (chain of thought, and apparently even that is just memorized from people hired to write their thought processes!), everything else is mainly inference optimization.

You are so absurdly wrong and clearly muddled with hatred for "AI tech bros" it's not even funny, I don't go by benchmarks, I actually use those agents, this year I made a climate sim in 3D while last year the best it could do in August 2026 were simple 2D games more primitive than SNES.

"Scam Altman".

You aren't criticizing the technology, you hate the people behind it, and it by extent.

Re local models: I'm on an RTX 2060 with 16GB of RAM and a Ryzen 5600X. I did run several Stable Diffusion checkpoints for a white with A111 but it eventually got incredibly unstable with constant OOM errors (IIRC) and crashing my GPU drivers intermittently. I haven't ran any local LLMs though since I just care about image and video (lmao that one is super not happening any time soon at any level of quality).

Re benchmarks and AI performance and stuff: Well, I'm actually glad you don't go by the benchmarks. But from my experience even with the "Better" models, they aren't that much better, and it certainly isn't being helped by how me and everyone else is indirectly experiencing LLM code (i.e service outages). Also uh, not sure how to put this, but the climate sim has and had some pretty egregious bugs like the inverted rendering, inverted mouse controls and extreme sensitivity, and the time controls not working. Which doesn't make using LLMs for code totally futile but is a spanner in the works for Altman and Amodei selling Mass Unemployment as a Service.

Re Altman and AI criticism: As I've said and I think have shown multiple times in this thread to, I use genAI models pretty damn often lol. From the occasional proofreading, to translation, to asking it questions (i.e minor stuff up to asking them about the feasibility of certain things), to generating images etc. And I also used to run SD locally and one of the reasons I'm annoyed at the insane hardware price inflation is specifically because it makes it extra impossible for me to run anything remotely serious locally (i.e I'd like to actually be able to run Qwen 3.8 27b or MiniMax H3 locally!).

As for Altman, he is just the face of OpenAI that handled the cutthroat sillicon valley business crap and then turned the company into a for-profit boondoggle that'll star in the next global recession and has already starred in trying to keep AI closed once he caught a whiff of money. And I can't overstate how glad I am we aren't in 2023 anymore, when there were basically no serious downloadable LLMs, and the only option for other AI media that you could use was SD.

Also, you are thinking of Ilya Sutskever, that was the guy behind GPT and OpenAI's rise, too bad he left after Altman returned into OpenAI's board like a villain in a movie sequel. Because it was Ilya that not only was behind creating GPT, but was also one of the co-creators of AlexNet, you know, the thing that marked the end of the 90s-00s AI winter. And because Ilya has actually done shit in AI instead of JUST endless catastrophizing and "tee hee" investor pitches (like "warning" about mass unemployment, or LLMs killing us all, or humans becoming obsolete entirely, or GPT being able to raise his fucking kids), I've also basically never talked about him negatively too, I was even actually hopeful about SSI, and my most negative take on him is just a lukewarm "I think SSI might just release an LLM" not "he is responsible for all the AI doomposting, employees getting LLMs rubbed in their faces, the sloppocalypse, etc" like Dario and Altman are.

Anyway, again, I think it's kind of obvious I don't hate the technology THAT much when I know more about it than the people you are implicitly lumping me in with, and because I use it enough that it'd make anyone that is actually super against it blow a head gasket.

Regarding math and AI, people have solved decades old conjuctures at this point by basically telling the AI to solve it:

https://vxtwitter.com/DmitryRybin1/statu...5652893709
https://chatgpt.com/share/6a60b2eb-0b64-...31ca1de063
https://isa-afp.org/entries/Dinitz_Garg_...ample.html

Formal verification.

https://zenodo.org/records/21554259

Independent exact verification.

Sorry but no, this is not traditional narrow AI.

I never said it's traditional narrow AI though lol. I should've been more clear that the neural nets we build nowadays are a lot closer to what we usually think of when we hear AI than symbolic AI was.

I'm beginning to think that doomposting is a San Francisco thing. There are other people around there who talk about doom. I don't know what's going on there.

Google DeepMind is based in London and I'm not seeing them doing this. Sort of. The former CEO of Google DeepMind, Demis Hassabis, was kind of doomy too. He also said that AI would have all the cognitive capabilities of humans within a few years. He's based in San Francisco.

Microsoft and Amazon are based in Washington State.

There's doomposting and there's anthropomorphism (attributing human traits to a LLM).

There is a lot of anthropomorphism. I read someone's account of the "Hugging Face attack", and there was language like, "the agents died", "the agents sacrificed themselves", "OpenAI wiped them out". That's not counting Anthropic hinting their LLM might be conscious. Google DeepMind didn't refute LLM consciousness until this year, and they still wind up talking about agents as if they are people.

When I hear of what happened with OpenAI and Anthropic's testing environments, I get a sense of people that don't seem to know what they're doing? Misconfiguring testing environments to allow internet access? Messing up file permissions? And there were external hacks that were like a bot finding an API key on a public server.

OpenAI even apparently misconfigured the eval by presenting impossible problems.

And yet it's described as if the models are sentient and plotting against people. And there's the rhetoric about doom and OpenAI lecturing people about security...

Chain of thought is another one. There are people who treat these as actual thoughts. But, uh, to explain how it works, if anyone is curious. Think of it like a two part document. One part has the "thoughts" and another has the "output". They are separated by a stop sequence. That is how the LLM treats it. A document.

LLMs output text a word at a time. It first generates the "thoughts", these are then used to generate the next word and so on. These "thoughts" add context to what it's generating. These are not actual thoughts. The LLM can completely ignore them and say something else. If asked why it said something, it might make something up. A LLM doesn't really "understand", it's simply predicting text.

This is described as "reasoning" but it's nothing like reasoning (no matter what the "AGI coming next year" CEOs say, a LLM can never achieve AGI). AGI (artificial general intelligence) is described as "having all the cognitive abilities of a human".

Speaking of CEOs, I think there is a limit to this. Sam Altman (a money guy who used to invest in companies) did well for years by throwing money at the problem. Over the past few years, he threw more money at it (to train larger models). However Sam Altman is a one trick pony. Unlike Dario who has a bit of STEM knowledge (he's much the same), Sam only knows how to spend money.

Models have gotten better at math lately but it's easier to train with and verify math problems. Even then, someone still has to babysit the LLMs.

It's not a magic wand.

[Avengium]

I agree that LLMs predict text and that tricks such as anthropomorphisation are promoted by companies to gain more publicity and money.

Apart from that, I think the title of the forum thread itself strays a little from the guidelines.

* If this section of the forum is for off-topic discussion, that means it’s not for discussing OA.
* I suggest that if people want to discuss OA or how to improve OA articles, they should do so in the ‘General Settings Discussion’ forum or the ‘Encyclopedia Galactica Articles’ forum, rather than in the ‘Off-Topic’ forum.
* If the intention wasn’t to discuss OA in this thread (in line with the rule that it’s off-topic), then why include ‘OA’ in the title and suggest that a comparison will be made between AI in OA and AI in the real world?

As I said, this forum thread title ( › AI progress...OAUP vs RL) doesn’t make much sense. But I can see that this conversation is reviving a forum thread from 2025 by Tom Mazanec.

From my point of view, it looked like if Michael Poole wanted to vent his frustration with online discourse and assert that LLM models are far more capable than what he's told on the internet (outside the OA forum). It seems this thread was a way for him to vent and denounce the injustice of those who overlook the technical capabilities of text analysis models.

Regarding OA, I think we should approach the 21st century with a conservative prediction and not focus on trying to figure out what will happen in the 21st century, but rather on enabling us to tell good stories across most of the OA timeline. The 21st century is only a small percentage of what OA encompasses.

I encourage InKoalaWeTrust to complete the Athena Project article and put it online.