Detailed Analysis
A developer working under the name Fable has released SunaBox, a browser-playable homage to OE-CAKE, the 2008 Prometech physics sandbox toy that briefly went viral for letting users spray colorful, jelly-like particles across a canvas. SunaBox recreates the original interface with claimed 93% parity, including support for importing legacy .oec save files, while running on top of an entirely new physics engine called SunaEngine. The engine is open-sourced under AGPL and represents the more technically ambitious part of the project: a deterministic, cross-platform particle physics simulator built on integer math and order-independent accumulation rather than the floating-point calculations typical of real-time physics engines. That design choice guarantees bit-identical results across different GPUs, drivers, and operating systems, a property demonstrated through three live browser demos, including a "Pachinko Proof" that contrasts a standard float-based solver (which diverges wildly depending on the order contact forces are summed) against SunaEngine's integer approach, which reproduces the same 256-bit state hash every time.
The determinism claim is the article's real news hook, since it solves a longstanding problem in networked and reproducible physics simulation. Floating-point arithmetic is notoriously non-associative — the order in which numbers are added or multiplied can change the result at the level of the last bit, and those tiny discrepancies compound rapidly in chaotic systems like fluid or granular particle simulations. This is why multiplayer games that simulate physics locally on each client typically need to constantly stream corrected state over the network (or use client-server authoritative models) rather than simply syncing inputs. By replacing floats with integer math and enforcing order-independent summation, SunaEngine sidesteps that problem entirely, meaning only user inputs — quantized to physics substeps — need to be transmitted for two machines to stay in perfect sync. The developer backs this up with verified replay hashes matching across Apple Silicon, an iPhone, a Steam Deck, Intel integrated graphics, and an Nvidia RTX 2070S, a fairly rigorous cross-hardware validation for a solo/small-team project.
What makes this notable in the context of Claude and Anthropic coverage is the disclosed workflow: the human contributor describes their own role as generating ideas and providing "qualia-type feedback" — essentially taste, feel, and design judgment — while Claude is credited with doing "the math, the code, and the adversarial review." That framing positions Claude not as a simple autocomplete tool but as the primary engineer executing on a nontrivial numerical-methods problem (designing a fixed-point/integer physics solver with guaranteed determinism, a domain that normally requires specialized expertise in game engine architecture, numerical stability, and GPU compute via WebGPU). The "adversarial review" detail is particularly telling, suggesting Claude was also used to stress-test and critique its own or the human's design decisions, rather than just generate code on request.
This project fits into a broader pattern of AI-assisted development enabling individuals or small teams to ship technically sophisticated software — in this case, a novel physics engine with cross-platform determinism, WebGPU integration, and a polished nostalgia-driven consumer product — that would traditionally require a larger specialized team. It also reflects a growing trend of hobbyist and indie developers using coding agents like Claude Code to tackle systems-level problems (deterministic simulation, networking architecture, GPU compute) rather than just web app scaffolding, suggesting increasing developer confidence in AI models' ability to reason about hard numerical and architectural constraints. The planned Steam release, with expanded particle caps and workshop features, indicates the team intends to commercialize the project further, which will serve as an ongoing test of whether AI-assisted small teams can sustain and scale a live product against community expectations.
Read original article →