正在加载视频...
视频加载失败
GLM-5.1 > Claude Code (Opus 4.6)? I'm tripping or CC has become very bad but built a Three.js racing game to eval and it's extremely impressive. Thoughts: - One-shot car physics with real drift mechanics (this is hard) - My fav part: Awesome at self iterating (with no vision!)... show more
42 条评论
Play it here: (btw stack: GLM-5.1 + Pi + Bun - there's much more thing to say like how it's not lazy and good at decision making: it decided road texture would look better with 8 procedural layers and it was right)
This was not one shot. Can we start normalising not talking shit on the internet for the sake of impressions and likes?
I never said it was one shot...
One-shot car physics with real drift mechanics (this is hard)
Yes, it did implement solid driving in one prompt... as written
Track 434 🏁 Fastest Lap: 8.850s ▇ ▂ 🔗 Race against me at
This is fun, we've been doing similar work for race fans! Come join our community with 100k users and find your global rank. No login
the debugging tools with zero vision part is the real human signal in this eval. any model can throw together a basic three.js scene, but building custom telemetry via bun webview to verify vector math it can’t actually see is elite level context engineering. we see the same thing when building agents for e-commerce the win isn’t just the final output, it’s the model proactively building the validation tools because it knows it can't trust its own logic without a closed loop. if it handled 500+ lines of racing logic in one go without the context window turning into a mess of legacy racing lines, it’s probably doing some serious internal pruning or summarization.
idk if GLM-5.1 is that much better, but it's def handling edge cases way more smoothly than CC right now, simple!= easy
Promp please :D
I wish they release mid-sized Flash model again
the no-vision debugging thing is actually wild - GLM building 20+ tools to drive the car without ever seeing output is a different kind of smart. been running GLM for daily stuff past couple weeks, solid for the grind
the self-iteration loop is where most coding agents fall apart. if GLM-5.1 can actually correct its own physics mistakes without human handholding that's a real differentiator. Three.js is brutal for evals since there's no room for sloppy math.
@threejs One shot? Give us a prompt to proof.
No one would have believe you can build a full working game using AI 4-5 years back
been using Opus 4.6 for months and it's solid but if GLM can self-iterate without vision that's actually a game changer for complex systems gonna test this on People Loop's backend logic, curious if it handles state management better
No need to consider technology stack and model, only @CombosFun_AI
car sound like plane sound
yo Victor this is actually insane the self-debugging tools part without vision is wild - like it basically built its own eyes to fix itself
Looks cool! I’m gonna try it
This is pretty cool. Have you tried glm through claude code instead of pi?
@threejs This actually looks fun to play! Now I want to try!
@threejs
One-shotting drift physics is actually insane.
is it really that good? looks like a good time to test it.
I use Claude Code daily and it's still the best for real production work across large codebases. But for self-contained creative coding like games and demos, newer models are catching up fast. The self-iterating part is key though. That's what separates agents from autocomplete. Interesting to see GLM getting there.
GLM 5.1 has those functions that are superior in some ways than opus 4.6
@badlogicgames Which harness used with GLM?
@badlogicgames
@anderslie @badlogicgames any uh plugins, extensions, skills? I think they're called extensions in pi, but also did you use skills and which?
@anderslie @badlogicgames you just need Skills ( - extensions are something else to extend Pi capabilities.
@anderslie @badlogicgames I was told why Pi is being hyped right now is "because it's nothing special, very basic, but also allows extensions" So, did you use/install skills or was this vanilla pi that built this with glm-5.1, please?
@anderslie @badlogicgames only skill used is something so it uses the new Bun WebView everything else is vanilla
What provider are you using? GLM 5.1 has been been very slow for me.
one-shot car physics with drift mechanics in Three.js is legitimately impressive. currently building a browser RTS in Three.js for vibejam — the self-iterating without vision part is wild. adding this to the eval list.
Why not give it vision?
this is not one shot, why do you lie ?
the self-iteration piece is what matters most. one-shot is a party trick, but a model that catches and fixes its own mistakes mid-task is what actually ships stuff. if GLM holds that up consistently it's a bigger deal than the one-shot benchmark.
@_akhaliq Lately definitely
building its own telemetry tools to self-iterate without vision is the standout here. does it do the same on existing codebases or is this more of a greenfield strength?
this looks fun to play!
No cap, that sounds sick. If it actually nails one-shot drift physics that's legit next-level, and the self-iteration bit is a huge flex - love seeing this, keep the demos coming!
