PewDiePie Is Right About AI

PewDiePie made a video called Stop Using AI Right Now after spending $20,000 on an eight-GPU rig, teaching himself to code, and building his own AI interface. This is the same YouTuber who used to film himself screaming at horror games.

He didn't build that rig to quit AI. He built it to understand which AI is actually worth using. The smartest model isn't automatically the most useful one. The value is rarely in the model alone.

Hardware Is Mostly a Memory Problem

PewDiePie's setup is eight RTX 4090s, six standard and two modded cards from China with 48GB of VRAM each, double the normal amount, and the whole rig cost around $20,000. That sounds absurd for a personal computer, though in AI infrastructure terms it's cheap.

VRAM is about capacity, not speed. The A100, Nvidia's enterprise GPU, has 80GB and costs $20,000, while a consumer RTX 5090 has 32GB for $3,000. You'd assume the A100 crushes the 5090 in every way, but it doesn't. The 5090 is faster. Once a model fits in memory, extra VRAM doesn't make the chip inherently faster, only raising the ceiling on what can fit without spilling into slower storage. You're paying for capacity, not necessarily for more raw throughput.

That's why enterprise cards cost so much, and why the pricing looks insane from the outside. A single B200 can cost $50,000, an eight-GPU server crosses $500,000, and one B200 pulls around 1,200 watts.

PewDiePie spent $20,000 and still couldn't run the biggest models. He had speed, but not enough memory, and that gap between speed and capacity is where the gray market appears. There's demand for modified cards with expanded VRAM in places like China because Nvidia has every incentive to keep VRAM gated. A cheap consumer card with too much memory becomes useful to too many people running models locally.

The funny part is that a machine like a Mac Studio starts making more sense than it first appears. A high-end M3 Ultra with 96GB of unified memory costs $9,500. That 96GB sits on the same chip as the CPU, GPU, and tensor cores. You can use all of it for a model, or upgrade to 512GB.

Compare that to the A100. It costs $20,000, yet the Mac Studio has more total memory at less than half the price, and scales to 512GB if you need it. The tradeoff is speed. A Mac Studio won't traverse data as fast as a high-end GPU, but if the goal is to run the biggest model possible at a reasonable speed, rather than the fastest inference on a smaller model, the math changes completely.

Why People Are Right to Be Annoyed by AI

A lot of AI criticism gets dismissed as ignorance or fear. Much of it is just a rational reaction to the way AI is being sold.

Most people don't encounter AI through thoughtful local workflows or careful automation. They encounter it through autogenerated thumbnails, useless product features, hotel-room gimmicks, and companies promising to cut labor while calling it innovation. If that's your exposure to AI, then hating it is a perfectly reasonable response.

That context matters because it explains why PewDiePie's video lands. He's not speaking like a founder trying to raise money or an influencer trying to attach himself to the trend. He's reacting to what actually feels good, what feels cheap, and what feels hollow. That emotional honesty gets closer to the truth than a lot of supposedly technical analysis.

AI-generated images have exactly one use case. Human eyes.

If nobody sees the image, it's produced no value. This is where the irritation comes from. Text is different. Large language models generate code, tool calls, intermediate reasoning, summaries, and machine-readable output that may never be consumed directly by a person. The vast majority of generated text never gets read by a human, maybe less than 2% does. Most of it goes to automation, tool calls, context building, and internal reasoning that never reaches a person. That's a big part of why LLMs can be genuinely useful outside of pure content generation.

Media generation doesn't have that escape hatch. Its value is overwhelmingly tied to human consumption, which means the industry pressure around it naturally pushes toward replacing artists, videographers, musicians, and other creative workers. That's why it feels worse. Rather than just assisting the creative act, it's often marketed as a substitute for it.

There's also a difference in how the outputs feel. A drawing, a song, or a film scene carries intent. Even mistakes in human-made work feel connected to a mind that made choices. AI-generated media can look polished, but much of what's on screen was never consciously selected by anyone. That loss of intent is why so many people have the same reaction. It looks fine until they learn how it was made, and then something deflates. When you know it wasn't made by a human, it evokes a feeling of being deflated, grossed out, maybe a little bored. You're not deciding to hate it because it threatens your livelihood. It feels worse because you don't feel the same creative intent behind it.

That doesn't mean generative media has no use at all. It can be helpful for rough assets, mockups, background elements, or low-stakes visual filler, but those are tool-like uses. The more it tries to replace the creative core, the more hollow it feels.

Stay Within Your Depth

AI can help you learn to code, or it can prevent you from learning entirely. The difference is whether you let the model go beyond your current understanding.

If you already know how to build a feature, letting AI draft part of it can save time, but if you don't understand the underlying concept, outsourcing the task just outsources the learning too. The better use is to ask for explanation, elaboration, and feedback. Ask questions, don't ask it to do tasks.

That's what makes AI coding useful without making it corrosive. Programming was never mainly about typing speed. It's about understanding systems, tracing failure, and making tradeoffs. AI helps when it clears friction around work you already grasp, but it hurts when it lets you skip the part where the mental model is supposed to form.

Local models add another layer here. They're weaker than the best cloud models, which is a limitation, but also a guardrail. They're less likely to sprint far beyond your current knowledge, and they're valuable for reasons that have nothing to do with matching frontier performance, things like privacy, control, and low latency. Running a model on your own machine for personal photos or sensitive data is a very different proposition from handing that data to a large platform.

The realistic future for local models isn't that a tiny phone-sized model magically catches up to frontier systems. It's that smaller models become good enough for narrow, private, fast, on-device tasks. That's still extremely useful, just not the same thing as replacing the biggest cloud models.

AGI?

PewDiePie built his own interface after realizing a local model without tools can barely do anything. It has no search, no memory, no access to files. So he started adding features, letting the model use the tools he built for it.

That's when he noticed something the industry is still trying to learn. He was running tiny models, just two billion parameters. They hallucinate and generate gibberish because they have no stored knowledge, but give them search and the query barely slows down. Give them retrieval and suddenly you have a fast tool for information. You don't need the smartest model. You need the right tools.

This is where most of the field is still wrong. Companies pour resources into squeezing slightly more intelligence out of larger models instead of accepting that models are limited by what they're allowed to do and the tools they can access, not by how many parameters they contain. Modern benchmarks for software tasks track two things, the model and the harness, the set of tools it uses to traverse codebases, look up information, apply changes. The harness matters at least as much as the model itself.

AGI isn't a smarter model. AGI will be a set of tools for a smart enough model.

That line sounds provocative, but it maps surprisingly well to how modern systems are actually improving. The big breakthroughs in practical AI use are rarely just parameter count. They come from better tool use, better retrieval, better environments, and better workflows. In coding especially, the difference between a helpful model and a frustrating one is often the surrounding system. Can it search? Can it inspect files? Can it apply changes? Can it recover when a step fails? That's what agentic work really is. Not magic, not a model becoming conscious. A model operating inside a useful harness.

This also points to the more plausible version of self-improvement. Models don't need to make their own brains bigger to become more capable. They can build or refine the software around themselves. They can create scripts, interfaces, retrieval layers, and workflows that expand what they're able to do. That path still requires careful sandboxing and a lot of discipline, but it's far more concrete than the fantasy that scale alone solves everything.

Stop Using AI doesn't really mean stop using it at all. It means stop treating it as a substitute instead of a helper, and stop assuming the smartest model is automatically the most useful one. The value is rarely in the model alone. It lives in the setup around it, in tools, constraints, privacy, and intent. The industry obsesses over bigger brains. The better question is whether we're giving those brains any hands.