Are Large Language Models Really AI?
|
Are large language models really AI?
those really heavy monitors that cause cancer
K123 said: » Didn't realise this was a rightwing nutjob, dailymail reader, reform voter conspiracy theory thread. Any thread that has the RAN (does not) have potential. K123 said: » What's CRT? What is this? Explain it to me like a 5 year old.
nvm, got ChatGPT to do that for me: Quote: Imagine you and your friends are building with blocks on the playground. Everyone has different-colored blocks, and that’s okay—blocks can be red, blue, green, yellow. Now, pretend there is an invisible rule that says: Red-block kids must share their blocks with everyone. Blue-block kids don’t have to share at all. That rule isn’t fair, right? It makes the game easier for some kids and harder for others, just because of block color. Critical Race Theory (CRT) is like putting on “fairness glasses” so grown-ups can: Spot those unfair rules—even when they’re hidden or old. Ask why the rules were made that way. Work together to change the rules so everyone can play the game the same way, no matter what color blocks they have. So CRT is simply about finding unfair rules that treat people differently because of skin color and fixing them, so the game of life is fair for everybody.
[+]
Not a bad explanation at all.
But CRT is a post graduate course taught in certain fields of study. The right wing freakout says it is taught at every level of schooling, even elementary, and that what it teaches is to hate white folks. I asked it what being on the internet at 80 must be like, to explain it to me like I'm a 5 year old:
Quote: Okay! Imagine you are playing on a big, noisy playground. All the kids are running around super fast, playing games you’ve never seen before, using words you don’t understand, and changing the rules every five seconds. Now, pretend you are 80 years old. That means your knees are creaky, your eyes are a bit blurry, and you learned how to play totally different games a long, long time ago. When you go onto the internet, it’s like walking into that wild playground. Everything is so fast. There are a million buttons to press, but you don’t know which ones do what. People talk in strange codes like “LOL” or “DM me” and sometimes it feels like they’re speaking a whole different language! You want to join in, maybe talk to your grandkids or watch your favorite old movies, but it's tricky. You're worried: What if I press the wrong thing? What if someone tricks me? It's like trying to jump into a game of laser-tag with kids who’ve been practicing forever, and you’re holding a flashlight instead. But if someone kind and patient shows you what to do, and helps you slowly one step at a time, you can still smile, play a little, and feel like you belong again — even if the playground looks very different now. The current AI is like what we have been learning with math since highschool.
At the beginning, we didn't know squat so the teacher gave us a book with 100 problems to solve. After cracking a couple of them, we started noticing they are sort of the same thing with different variations. Then we solved them all like a machine. Then comes the next set of lessons with 100 of its own problems... During midterm, the professor got drunk and threw in a bonus problem we haven't seen or worked on before. Most gave up but some dude will try to break the problem into smaller chunks and pieces to find if there would be something similar. This requires a certain degree of mastery of the knowledge. If the professor wasn't too drunk then they would find something similar and allow them to notice it's sort of like a combination of the previous problems and you can solve them in sequence. So the AI nowadays can figure out something based on how much related data patterns it learned but at some point it will try to make stuff up because they look similar. However, I remembered something I read in one of those crazy AI textbooks back in the 80s. Intelligence isn't about solving problems. It's about creating problems when none exists and trying to solve it. AlphaEvolve (Google AI) just discovered a few math solutions no human ever has. LLMs will be trained with this new knowledge faster than humans can ever learn it, etc.
K123 said: » I asked it what being on the internet at 80 must be like, to explain it to me like I'm a 5 year old: Real early on I was at the bleeding edge of microprocessors. (I have fallen WAY off it since then.) But having been on the web since dial up days I speak the language. And never having had kids I have no grandkids to run to when I have problems. Phones... I am way behind on phones despite being an early adopter of cell phones. And if I ever buy a new car (improbable at best) I am going to need classes on how to use the ******* features.
[+]
K123 said: » AlphaEvolve (Google AI) just discovered a few math solutions no human ever has. this image in particular:  The new solutions it found were all to problems that are easily scored, which is a pretty great space for algorithmic exploration and lend themselves to AI application. Defining and balancing those metrics for more complex problem will be the hurdle, especially in cases when the data is fairly dirty. Fenrir.Niflheim said: » K123 said: » AlphaEvolve (Google AI) just discovered a few math solutions no human ever has. this image in particular: The new solutions it found were all to problems that are easily scored, which is a pretty great space for algorithmic exploration and lend themselves to AI application. Defining and balancing those metrics for more complex problem will be the hurdle, especially in cases when the data is fairly dirty.  This is why you sometimes get 19-year-old math wiz solving world class math problems or some homeless dude solving one of the 7 millennial math problems while a bunch of other math cracks couldn't hack it with 30 years of research.
The solutions to a lot of maths are just like your average highschool trick questions but with a lot more expansive tools to work with. Normally, mathematicians would decide on the tool to use to solve and try to hack at it. Sometimes, they have to gather a bunch of tools to create a new tool to hack at it. This is why sometimes you get calculus problems that can't but solved until someone put the problem into geometry or something. AI can go through the tools way faster than humans so they tend to find solutions faster. Some signs of AI model collapse begin to reveal themselves
Prediction: General-purpose AI could start getting worse Actually the article says they are getting worse. Quote: Welcome to Garbage In/Garbage Out (GIGO). Formally, in AI circles, this is known as AI model collapse. In an AI model collapse, AI systems, which are trained on their own outputs, gradually lose accuracy, diversity, and reliability. This occurs because errors compound across successive model generations, leading to distorted data distributions and "irreversible defects" in performance. The final result? A Nature 2024 paper stated, "The model becomes poisoned with its own projection of reality." The "The model becomes poisoned with its own projection of reality." part proves they are behaving like people. They need to model better people. Or just kill off humanity and start from first principles.
I watched a large session of Github Copilot webinars.
Last year I was already introduced to data science for an exam I partake. Few weeks ago I finished an entry level book of programming with AI. Idk, I feel like AI fields has two interfaces: one publicly available for consumption, and the one that plays like a perfomance competition. While trying to compete for performance is quite intricate and complex field, on the other hand the AI for consumption is quite uninspiring as it never felt like a technical or algoritmical procedure: prompt engineering feels so cheap, those webinars had total of 12h of content, and Im not sure what was added above the entry level book I had. That means: learning how to use prompt engineering is like using a couple of “advices” to generate a good answer. I dont feel like anyone at this point in time can provide the “optimal” prompt, that could return the “best” answer with the minimal amount of commands. That is why I dont particularly feel invested in learning more AI for consumption. And AI for competition… nah, not my thing. Im a creative person, not a competitive one https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
8B model with capability of 200B models from just a few months ago. |
All FFXIV and FFXI content and images © 2002-2025 SQUARE ENIX CO., LTD.
FINAL FANTASY is a registered trademark of Square Enix Co., Ltd.