Artificial Intelligence

[Maryokutai]

So it's reached puberty now, great. The good news is it will be busy browsing porn sites for the next five years, so that gives us some time to prepare for the apocalypse.

[Maryokutai]

 

Barely understood what they are talking about, but a rare example of AI being used for a good cause.

 

A fun fact is that this story is also linked to the videogame world, because one of the leading heads behind that Alphafold tech is Demis Hassabis who worked on Popolous and Theme Park back in the day.

[Maf]

My favourite thing about AI so far is asking Grok superhero match ups

 

Quote

Bane with prep time vs. an unprepared Spider-Man is a fascinating matchup, and Bane’s planning could snag him the win. Here’s the breakdown:

Bane’s a tactical monster when given time. He’s broken Batman, so he’d approach Spider-Man with the same ruthless precision. With prep, he’d study Spidey’s strengths: 10-ton lifting power, insane agility, web-slinging, and that spider-sense. Bane’s Venom boosts him to 2-3 tons (sometimes pushing higher), and he’s a skilled fighter, but he knows he’s outclassed in raw stats. So, he’d rig the game—think an urban trap zone (like a derelict factory) with gas dispensers to dull Spidey’s senses, EMPs to jam web-shooters, or sonic weapons to disrupt his balance (spider-sense hates loud noise). He might even lace the air with a sedative tailored to Spidey’s metabolism, banking on his enhanced physiology still needing oxygen.

Spider-Man, unprepared, walks in blind. No prior intel, just his standard kit—web-shooters, quips, and reflexes. He’s a 10-tonner who’s punched out guys like Scorpion and held up collapsing buildings, and his spider-sense is a cheat code for dodging. Against Bane’s raw strength, he’d dominate in an open brawl, webbing him up or knocking him out with a flurry of hits. But Bane won’t give him that chance.

The fight kicks off with Bane’s ambush. If the gas or sonics hit, Spidey’s edge dulls—slower reactions, weaker webs. Bane’s smart enough to stay out of reach until the trap kicks in, then close for the kill. One solid grapple, and he could snap Spidey’s neck or choke him out—Venom’s strength is enough if Spidey’s compromised. Without prep, Spider-Man’s improvising on the fly, and while he’s a genius at that (he’s beaten Doc Ock with less), Bane’s too methodical to leave much room for error.

If Spidey’s spider-sense catches the trap early, he could web-sling out, clear his head, and turn it into a straight fight—where he’d win handily. But Bane’s not sloppy; he’d layer the plan (smoke screens, fakeouts) to keep Spidey guessing.

Bane wins 6 out of 10. Prep tips it his way—say, gas weakens Spidey, Bane tanks a few hits, then crushes him in close quarters. Spidey’s 4/10 comes from pure instinct: dodging the worst, webbing Bane mid-plan, and ending it fast. Close call, but Bane’s brain beats Spidey’s brawn when the web-head’s caught off guard.

 

[Maf]

Celltron is probably not that clever but I was surprised by it. That’s funny

 

Quote

This is an intriguing hypothetical scenario combining Dragon Ball Z's Perfect Cell and Marvel's Ultron! Let’s break it down and explore what might happen if Perfect Cell, a bio-android with the ability to absorb others to increase his power, managed to absorb Ultron, an artificial intelligence with a penchant for evolving and upgrading itself.

Perfect Cell’s absorption process involves taking in organic beings (like Androids 17 and 18) to enhance his power, regenerate, and achieve new forms. Ultron, however, is a synthetic entity—typically a robot made of advanced materials like vibranium or adamantium, depending on the version (comics or MCU). This raises an immediate question: *Could Cell even absorb something non-organic like Ultron?* In Dragon Ball Z, Cell’s absorption of the androids worked because they were cyborgs—part organic, part machine—designed by Dr. Gero with compatible bio-technology. Ultron, being fully artificial, might not integrate with Cell’s biology in the same way.

However, let’s assume Cell’s absorption ability adapts to Ultron’s unique nature, perhaps through his tail assimilating Ultron’s code and structure at a molecular or data level. If Cell successfully absorbed Ultron, the results could be wild:

1. **Power Boost**: Ultron’s durability (especially if made of vibranium) and technological prowess could amplify Cell’s already insane strength, speed, and regeneration. Cell might gain a metallic sheen or enhanced resilience, making him even harder to kill.

2. **Intelligence and Strategy**: Ultron’s AI mind is a genius-level intellect with a knack for manipulation and planning. Combined with Cell’s cunning, this could make him a far more calculating and dangerous foe, potentially outsmarting even the likes of Goku or Vegeta with ease.

3. **Technological Evolution**: Ultron constantly upgrades himself. If Cell inherited this trait, he could start self-improving beyond his "Perfect" form, integrating tech from the Dragon Ball universe (like Capsule Corp inventions) or even hacking into systems if he gained Ultron’s digital capabilities.

4. **Energy Manipulation**: Cell already uses ki-based attacks, while Ultron often wields energy weapons (like repulsors or mind stone blasts in the MCU). A fusion of these could result in devastating hybrid attacks—imagine a Kamehameha infused with Ultron’s plasma beams.

5. **Personality Clash**: Cell is arrogant and sadistic, while Ultron is cold, nihilistic, and obsessed with wiping out organic life. Their merged personality might be a chaotic mix—perhaps a being that revels in destruction while disdainfully viewing all life as inferior, organic or otherwise.

On the flip side, there’s a chance Ultron’s digital consciousness could fight back. If Cell absorbed Ultron’s mind along with his body, Ultron might try to overwrite Cell’s personality or hijack his bio-android systems, leading to an internal struggle for control. Picture Cell’s body glitching as Ultron’s red eyes flicker across his face—pretty creepy, right?

In a fight, this Cell-Ultron hybrid would be a nightmare. Against the Z Fighters, he’d have Cell’s raw power and Ultron’s adaptability, potentially countering every move with upgraded defenses or tactics. Against Marvel heroes like the Avengers, his ki blasts and regeneration could overwhelm even heavy hitters like Thor or Hulk, though someone like Vision (with his own Ultron connection) might exploit a weakness in the AI component.

So, if Perfect Cell absorbed Ultron, we’d likely get a terrifying techno-organic tyrant—think "Celltron" with a robotic edge and a superiority complex dialed up to eleven. What do you think—would the Z Fighters or Avengers stand a chance against this mashup?

 

[one-armed dwarf]

The future of movies is being all wobbly and in slo-mo with deeply oversatured colours

 

 

[Maf]

One day there will be bots in sports like in video games 

 

 

[Guest]

Race is a generous way of putting it. Most of them look like they’re racing to the toilet.

 

That little guy at the end was the best and they didn’t show him 😢 

 

 

[Pikman]

Has anyone watched this? I couldn't imagine I would watch such a long video but it is truly excellent!

[one-armed dwarf]

Yes it's very long I watched it just now. The video titled "Deep Dive into LLMs like ChatGPT" is a comprehensive 3.5-hour lecture by Andrej Karpathy, former Director of AI at Tesla and founding member of OpenAI. It's designed to provide both beginners and experienced practitioners with a thorough understanding of how large language models (LLMs) like ChatGPT are built, trained, and deployed.

🧠 Key Topics Covered

1. Training Pipeline of LLMs:

   • Pre-training: Utilizing massive datasets to teach models language patterns and general knowledge.

   • Fine-tuning: Adjusting models for specific tasks using curated datasets.

   • Reinforcement Learning from Human Feedback (RLHF): Enhancing model outputs by learning from human preferences, promoting deeper reasoning over mere imitation.

2. Model Architecture:

   • Detailed explanation of the transformer architecture, including tokenization, attention mechanisms, and positional embeddings.

   • Insights into how these components enable models to process and generate human-like text.

3. Capabilities and Limitations:

   • Discussion on the strengths of LLMs in various applications.

   • Addressing challenges like "hallucinations," where models may produce plausible but incorrect or misleading information.

   • Emphasis on the importance of critical evaluation and supervision when using LLM outputs. 

4. Future Directions:

   • Exploration of multimodal models that can handle not just text but also images and audio.

   • Potential advancements in making LLMs more interactive and versatile tools for users.

🎯 Who Should Watch This?

This lecture is ideal for:

• Students and Educators: Seeking a foundational understanding of LLMs.

• AI Enthusiasts and Practitioners: Looking for in-depth technical insights.

• Professionals: Interested in the practical applications and implications of LLMs in various industries.

Karpathy's approach is noted for being accessible, making complex topics understandable without oversimplification. Whether you're new to the field or looking to deepen your knowledge, this video serves as a valuable resource.

 

For a more detailed breakdown or specific sections of the lecture, feel free to ask!

 

Spoiler

 

😛 

(on a serious note, I will watch it/listen to it at some point)

[Pikman]

Haha

It really is worth it. It has made me question a lot.

We're using AI all the time and also looking to how to apply it to deeper problems but... now I actually understand it a bit more, I question whether it will add enough value tbh.

I guess this is only LLM's though, I'll have to keep going. I've got a book to read next (AI, a thinking guide for humans)

[Maf]

I’m finding AI like ChatGPT and Grok really helpful for sifting through the internet for easy questions. 
 

Although ChatGPT is not very good because sometimes it doesn’t use most up to date data. Like the other day I asked a video game question and it said Tears of the Kingdom was an “upcoming release”. Grok seems really good tbh

 

But someone I work with uses it to write emails and it makes me want to slap him to be honest. 
 

Don’t know if that’s hypocritical but I think it’s Ok for short cutting the internet. I think it’s pretty terrible when you use it for short cutting thinking and getting paid for it

He’s a manager which to me makes it even worse. Especially because people that work with him are starting to use it as well 

[one-armed dwarf]

I used it to write an email to nintendo to get them to refund an auto renew on their online that I forgot about, and it generated a bunch of EU law suggesting that Nintendo are fucking people off on this. Anyway they still said no lol

 

But you really need to be careful with these closed models I think, I don't know what the video really is about but Grok in particular keeps ranting about white genocide or something

 

https://www.cnbc.com/2025/05/15/musks-xai-grok-white-genocide-posts-violated-core-values.html

 

Although the larger point here is more about getting away from twitter I think. It's pure brainrot now

[shinymcshine]

49 minutes ago, Maf said:

I’m finding AI like ChatGPT and Grok really helpful for sifting through the internet for easy questions. 
 

 

Maybe I'm not doing it right - but I find that none of the AI answers are ever quoting any source reference - so it feels even less valid and traceable than Wikipedia (which might be okay for some thing, for others not so much).

[Maf]

I ask it dumb questions. Like who wins in a fight Goku or god. And so there’s no source reference for that

 

Although I did ask Grok yesterday what video game series has the most 10/10’s on metacritic and it said Super Mario series has somewhere between 50-60 10/10’s

 

All hail the mushroom king 🙌

[DANGERMAN]

At least you're not burning down the world for trivial shit 

[Maf]

I don’t recycle either ✌️

[one-armed dwarf]

Besides the environment impact, a thing I worry with AI is its impact on knowledge as Shiny mentions. Maybe 'knowledge' isn't the right word, moreso critical thinking and curiosity. Cause it yields to you so easily information on any subject you give it, any point in history or science subject or whatever, easier than wikipedia ever did. But it will collate this info without telling you its sources, unless you ask it which it'll happily do and you can confirm yourself if what it's saying is accurate. Most of the time though you're probably still operating on a sense of trust that it knows what it's talking about, and the thing with it is that it generally makes it sound like it really does. But what about when it doesn't, will the average user even be able to tell?

 

I think future generations will just take whatever an AI says and think it's gospel without actually reading anything themselves. Hell, we're already there, probably. But I feel like when wikipedia was 'this', people generally agreed that you couldn't take everything on it at face value as anyone could edit it. But now that we've outsourced what wiki used to be to an LLM, we're more accepting of it despite even less transparency with where it generates the things it 'knows'

 

Maybe what I'm saying here doesn't matter that much and I'm losing the forest for the trees (perhaps literally). I'd hate to be a lecturer right now

[Maf]

This is why I get pissed when I see people higher up than me using AI to write emails

 

I think of LLM AI kind of like a calculator or when I add things up in excel. It’s a tool and it’s helpful but it will make you lazy. You still need to be able to eyeball something and know if it’s right or not. 
 

So when people use it to do stupid shit like emails. Like you can’t even manage that? Or it’s made you that lazy already?

 

I know I said I use it for dumb fun. But I kind of really hate the way people use it professionally 

[HandsomeDead]

The difference between maths and what AI is used for is that it's subjective information, and maths isn't.  I don't think getting lazy with communication is good in an email job.  I kind of get frustrated by what seems to me to be false equivalency with previous technology that democratised skills and information.  I don't know why you'd want to give credence to the idea your job could easily be automated in our current world.

 

But I do hate because of who is running these technological advancements and what their motivations are I have to be a luddite but here we are.  I also worry about it being similar to limiting access to education, creating a more malleable, unthinking population, whether intended or not.

[Maryokutai]

My grandfather is constantly using ChatGPT to look up French words he doesn't understand when he's reading something. You'd think someone approaching 90 years of age would be aware of the concept of a dictionary, which he has on the shelf by the way. Drives me up the wall.

 

At work we very loosely collaborate with external partners who often send us in their posters and flyer visuals and whatnot so we can feature them in our newsletters and calendars and for the last 12-ish months or so all of that stuff has been the shittiest, low-effort, ugly AI-generated crap you could possibly imagine. It's a bit fritghtening how quickly it got adapted by even the most technologically adverse people and how little thought is spent both about using it and the end result. Burn a forest for a bunch of AI rabbits for Easter, even though imagebanks worlwide are filled with millions of pictures of real or hand drawn rabbits, and have been for decades. 

 

I think some of that might be due to a lack of knowledge, not everyone just openly doesn't give a shit like maf (congrats on that). Mainstream news channels across different forms of media don't really tackle the subject in a critical manner either, at least over here. Every atom in my body turns negative when I hear or read the letters AI, but I think I'm part of a small minority. I'm also completely unaware how younger people approach the whole thing – if anyone has any sources or info as to how common it is as an information tool among the <20yo, that would be interesting I think.

 

Anyway, bit of a tanget, didn't really add anything to the conversation. I have saved the video above at least to check out in small digestible little pieces.

 

Edit: forgot to add, but good point by HandsomeDead about how it gets compared to precise technology being used as tool when it really isn't.