@kumesh_a@speedrun I'm building Lockstep - the realtime AI competition platform where companies sponsor bounties on hard problems and recruit the elite AI researchers who solve them. Like next-gen Kaggle. Will send more info via your site :)
@dara_venture Doesn't that only apply pre-traction? Revenue is proof of market interest. Previous exits are social proof. Investors with appetite for risk will put their money on the line before there's any traction whatsoever. Can't blame VCs for taking the safer routes, but lets be real
@Jonathan_Blow On a re-read, seems like the point here _is_ about the artwork. That's damn awesome, can't imagine how long that took (i.e. how much of the decade was art alone)
@Jonathan_Blow is that ~100k morphing entities with many tris, complex shaders, and nontrivial collision detection being rendered with seamless zoom?
If not, it's still incredibly impressive for the artistic side, but isn't "lots of entities" just kinda table stakes for rendering these days?
@ThePrimeagen sort-of... I almost never `bisect` manually, but with vibe coding I'll tell it to commit intentionally and sometimes it will use that for bisecting/debugging. So - useful for me? no. Useful for the machines? yes.
@toklaxen@threejs Wasm wouldn't help at all. Instructions are instructions. The only thing would be to make it impossible to see the binary. So, something like running the code in a TEE. For better or worse, the feature of browsers is that they're an open platform for sandboxed execution
@ThePrimeagen When you were coding in Rust, did you wonder if LLVM was removing bounds checks on release builds where you expect it, or did you rely on a higher level abstraction? I know it's hard. I miss coding too. But the key here is to stop caring about what gets emitted "low-level"
@ThePrimeagen tbh I would have thought of this as over-engineering in the past. I'm a little allergic to "abstract everything" after a toxic functional programming detour a while back, usually it's a sign to me of dogma rather than careful thought. But when it's a win it's a win.
@ThePrimeagen Sorry if that's not super clear, but it's really simple - have something like `Command`/`Query` which drives _everything_ and it can be sent by mouse input, mcp, or direct JS function calls. One of those things that's hard to shoehorn later but is powerful if done from the start
@mrdoob@makc3d just out of curiosity, I dug through my emails and found one referencing your work ~15 years ago, and even then I was calling you out for being ahead of the curve :)
(plz forgive any cringe... this was eons ago in tech time)
TIL: uber-shader in compute is nice, shading different materials per-pixel, but it doesn't scale. Multiple indirect compute dispatches isn't very different than forward in heavily occluded scenes. But deferred-ish allows for nanite-like stuff, lots of lights, and sparse scenes
@threejs@zackpizackpi@threejs FYI I did try it, and ThreeJS destroyed my WIP renderer, revealed a ton of bad architecture decisions and bugs, hehe. Hot on the trail to fix and test more thoroughly though ;)
Under-explored area in web #gamedev - different kinds of materials in the scene. I don't mean PBR with different metallic settings, or texture painting with variations of the same few inputs - I mean completely different materials at once. Not sure if/how @threejs handles this?
@ThePrimeagen Same... curious if you've tried this approach: https://t.co/geoGtlIqqc
I know you're allergic to every "gamechanger" claim these days, but it really is. Not only can it evaluate UI like that, I'm doing tricky WebGPU stuff... need to also give it Naga tests and more, but it works
My current Claude Code recipe:
1. Chrome MCP
2. Back-and-forth discussion with the agent until it can write a plan file and hand me a "/loop" prompt to execute it
3. Quick sanity test that MCP is hooked up and can control the browser
4. Send the prompt
crazy powerful
My current Claude Code recipe:
1. Chrome MCP
2. Back-and-forth discussion with the agent until it can write a plan file and hand me a "/loop" prompt to execute it
3. Quick sanity test that MCP is hooked up and can control the browser
4. Send the prompt
crazy powerful
@iquilezles@mike_acton@ID_AA_Carmack To be ultra clear - AI isn't identical. But I'd never in a million years have been able to afford their consulting time (nor would it be right if I could - they have better stuff to do). Tweaking a shader look or refactoring big interconnected systems at pro-ish levels is a win.
AI makes me feel like my personal heroes joined my team. I've been coding for 25+ years, but not much low-level graphics. Putting aside the ethics of training data - I get to benefit from @iquilezles@mike_acton@ID_AA_Carmack Fabrice Bellard, etc. It's a damn cool experience.