TOP DISCUSSION

Local and Browser-Native Execution Led the Discussion

Discussion on March 8 centered on where AI game workflows were running. Builders shared browser-only Qwen 3.5 inference at 180 tokens per second on WebGPU, an OpenClaw-built prototype running natively on macOS, and an AI-native Blender prototype for blocking out, texturing, refining, and finalizing 3D work.

The same thread also covered public shipping. Builders talked about open-sourcing infrastructure, publishing small artifacts quickly, posting videos alongside them, and using those posts to drive people back to a larger library or tool surface.

@Ex0byt: "No cloud, APIs, servers, just Pure WebGPU + WGSL optimizations." (post)

PROJECTS & PEOPLE

OpenClaw produced a Mac-native prototype instead of another vague demo

Nicolette shared a launch post showing OpenClaw building what she described as "Minecraft but Infinite Craft," with the result running natively on macOS.

An AI-native Blender prototype surfaced as a new interface direction

Aadi Kulshrestha's post described an AI-native Blender built over two days for Hack Canada, where the model helps block out, texture, refine, and finalize a scene.

A shader-database prompt turned into immediate public building

Startracker's post asked for a three.js shader database that agents could browse natively. Later in the thread, another builder said they had already made a related project public and were inviting feedback and PRs.

A builder pointed the next test game at play.fun after an earlier $2K result

One developer introduced a WIP called Rektanoid and said the next release would go through play.fun, adding that a previous game had already made $2,000 there.

TECH & TOOLS

Browser-native Qwen 3.5 crossed from curiosity into performance claim

Eric's post reported 180 tokens per second for browser-only Qwen 3.5 INT4 on WebGPU, beating a stated 150 tok/s target. The post described the stack as "No cloud, APIs, servers" and focused on Pure WebGPU, WGSL optimizations, Hugging Face SafeTensors, and a single browser tab.

GPT-5.4 was increasingly judged by native output, not benchmark aura

A separate post from Nicolette showed one day of game-engine progress with GPT-5.4. In the thread, builders also compared High and XHigh settings, with several messages favoring High for cleanup-heavy work.

Three.js agent infrastructure started to look like a missing product category

Startracker's prompt put a specific missing tool on the table: a shader database for three.js agents to browse and integrate natively.

QUICK TAKES

• Public posting: Builders talked about publishing small artifacts, recording videos, and using the first reply to point people back to a database or library.
• Browser-local benchmark: The clearest number of the day was 180 tok/s for browser-only Qwen 3.5 on WebGPU.
• Three-Maps reuse: The AI-native Blender demo drew notice partly because it reused Three-Maps-adjacent code.
• Model settings: High was repeatedly described as a better GPT-5.4 cleanup setting than XHigh.
• Native runtime: "Runs on my Mac" was one of the stronger product signals in the thread.