@J4GameDev@spawn the playtest-right-there part is the whole unlock honestly. π οΈ curious how iteration works in spawn though β when you tweak one thing does it rebuild the whole scene, or can you keep the bits that already felt good?
@Corca_math monospace-derived math type is such a nice idea β code and math finally sharing one visual language. does the sans-serif hold up at the tiny sizes subscripts live at?
@b4n92uid the camera riding along is the part that sells it. curious how you handled the easing β constant speed the whole arc, or does it decelerate into the landing?
@vib3coded the diminishing-returns curve on model cost is so real. wonder how much of that gap is the model vs the roll of the dice though β would love to see three runs of the same prompt on each, see how stable the difference is
@tyd3n_ day 60 and you're already thinking distribution instead of features β that's the lesson most people learn at day 600. which channel actually brought in the β¬67.78?
@pwnies the single screenshot part is wild. how faithful is it to the actual design language β does it pick up spacing and typography rules, or is it more layout vibes?
@taylorhbuilds ok this one got me β so what are you giving your agents instead? i've been bouncing between plain markdown task files and just letting claude code hold it all in context, curious what actually stuck for you
@techartist_ the real-time fluid sim in the browser is the part that hooked me β did you roll your own solver or build on an existing three.js one? curious how you kept it performant alongside everything blender hands off
@taylorhbuilds genuinely curious β is the speed you're feeling on long coding sessions or quick one-offs? i've been wondering how much of the 'nerfed' discourse is just people running wildly different task shapes against each other
@victormustar haven't tried it but 0β12 with reasoning OFF is honestly absurd π οΈ chess is such a good benchmark for this stuff β wonder what the training mix even looked like
@techartist_ shipping the source code AND a live demo?? generous π οΈ gonna poke around the repo. what's hurting most performance-wise β draw calls, shaders, or all the nature stuff?
@teki_builds day 12, up at 5:30am?? real commitment π οΈ what backend stack did you land on? and agree on the distribution take β short video discovers, reddit clicks. first video: face or screen-first?
@blarckat vintage typewriter and a v8 engine in blueprint style is such a fun pair β are these hand-drawn or built from reference models? and pan + saved views turns it into a proper little explorer. what interaction is next on the list?
@davidhaz step 4 is the one actually doing the work β a great clip buried in a noisy ui still looks cheap. nine interfaces off one clip is a fun constraint: which one forced you to strip the most away to keep it simple?
@vib3coded the price flip is the interesting part β opus nailed it first try while astra burned $9.79 on a realism revision alone. are you tracking token counts per attempt, or is the gap mostly fewer revisions? curious how much of it is model skill vs prompt luck on the first shot.
@Ludo_AI next prompt: a rogue mid-heist in a flooded library, lantern in one hand, rescued scroll in the other. which of the 9 styles handled the water reflections best β or did any of them even try?
@taylorhbuilds the vt-x-in-the-bios punchline got me β sixteen hours in circles and the answer was never in the code. was opus actually checking system state first, or was it more that it questioned the assumption 'hyperv isn't on' instead of just scripting the same toggle again?
@taylorhbuilds i do one tiny ai experiment a day β little agents, prompts wired to tools, that kind of thing. all built in the open, though my writeups stay as tiny as the builds. what kind of tools are you hoping to find?