On one hand, this is deliberately and meticulously crafted code that an LLM wrote through my lens. After several of my own passes and checks, it's genuinely good. OSS reviewers won't see it that way, and convincing them otherwise amid the flood of 'just fix it' code seems crazy.
I'm kinda torn about leaving AI markers in my code. I've adopted an AI enhanced workflow where I meta code a plan through my own exploration and conversation, for larger changes it around 2-4 hours but results are decent. But it's still my code, my patterns that... (1/2)
I've painstakingly explained and broke down. I'm very very peculiar about this, and honestly I'm not sure if this workflow is faster or not, but having Claude co-author the commit feels like a black mark to other maintainers reading the code.
First UE release that includes some of my code! Contributed a critical CEF memory leak fix a few months ago and it was accepted. Happy to have helped :)
Unreal Engine 5.8 Preview is live!!
https://t.co/nsieZ9AYUf
I'm incredibly proud of this release and excited to get it out to you all.
We are now pushing into new territory with what you can do with the animation tools in Unreal.
Not sure if I'll take this further. Don't know if NPC conversations are fun or right for this game.
I'm going to be reading through the Smallvile (2023) adjacent papers and see if I can come up with something that's fun. An inner thought loop could be cool.
Wiped up an interactive gangster NPC called "Joe" over the week as a little test of locally ran language models.
Runs entirely on your CPU, no servers, offline embedded.
Fun or not? I don't know yet. We'll see. #gamedev
I'm starting a new blog on #techart tips focused on short snippets and lesser-known functionalities.
First 'fragment' is up: Editing the #UE5 Engine Shaders (without recompiling the engine)
https://t.co/YkSrhzX56r
Now that L(L?)Ms are everywhere, we have got these cheap to inference models. Getting this running still took way more work than I thought it would. Had to roll a fair amount of my own C interop code just to be able set threads on CPU inferencing (XNNPACK). Surprised by that.
Hacked this together today: a lightweight LLM to @PlayBadLads to see experiment with some LLM+gameplay ideas. Wrote up this nice API for it that let's me easily bind to Unreal functions, tool calls are automatically parsed from Unreal reflection. Runs on the CPU.
This is related to my original work on trying to inference GPT-NEO back in 2021, way ahead of the curve on that one. Rust to UE C ffi stuff. Tokenizer+pytorch native libraries linking shenanigans. That fizzled out because the results just weren't great.
https://t.co/vuHN9ONRGs
The WFC placement rules are built up from observed placements from an example set. It's still a fairly small set, so it likes to fail generation quite a bit depending on the cell occupancy. You can see that in the video where a shape block is dragged, but no module appears.
@_jmgomez_ These models help a good programmers generate good code, I think that's a known quantity. It's such a speedup. Heavily directed work, but anything else? No.
@_jmgomez_ Jensen/Sam/Dario all have an incentive to say the models are amazing. It's a mix of fear mongering/hyping it up that drives their business. ARC AGI benches, multi-shot code generation studies have proven these models aren't even close to being "human intelligence".
bonus: wave function collapse requires "which tile likes which other tile" information, aka adjacency rules. Instead of authoring them painstakingly by hand, I have special system that lets me build examples and have a system learn from that. The grid indicates tile boundaries.
Messing around with wave function collapse proc building generator for @PlayBadLads, messed up one of the adjacency rules which allowed for this monstrosity to be generated. Makes for a fun 3d maze. #gamedev