Bilingual

Show HN: Small World – The "Preact" of 3D Web Engines (TypeScript, WebGPU)
20 days ago
- The text describes a hybrid rendering system with optimized Cook-Torrance PBR shading that compiles dynamically for WebGL2 and WebGPU.
- It includes a zero-allocation Finite State Machine utility executing natively in a recursive update loop for behaviors.
- Post-processing filters feature specialized shader passes for effects like night vision, VHS tracking, glitches, and Noir film.
- Conformal reflections involve mirror geometries, procedural checkerboard maps, and real-time sphere inversion math.
- Octree interactions provide DOM-like pointer events supported by optimized O(log n) raycasting.
From Muon to Gradient Clipping: Some Thoughts on QK Stability
12 days ago
- Muon optimizer works in function space by constraining the spectral norm of weight updates, aiming to limit the functional change for any input, contrasting with parameter-space optimizers like Adam.
- Applying standard Muon to Transformer's Q and K matrices leads to training instability because it separately constrains updates to W_Q and W_K, failing to control their coupled product in attention scores, which can cause spectral norm explosion.
- A principled attempt to modify Muon for QK involves directly constraining the change in attention score matrix S, leading to an optimization problem with a constraint coupling ΔW_Q and ΔW_K.
- Decoupling this constraint via the triangle inequality and solving subproblems results in a theoretically clean update formula, but it requires expensive pseudoinverse computations, making it impractical.
- A mathematical simplification reveals that the gradient G_Q naturally contains W_K, leading to a geometric interpretation involving a projection matrix, yet the external pseudoinverse remains, not solving the computational bottleneck.
- Viewing attention through a unified bilinear matrix B = W_Q W_K^T simplifies the Muon update to a single-matrix constraint, but this approach is impractical due to parameter explosion (d^2 vs. 2dh parameters).
- Multi-head attention can be seen as a low-rank approximation of full bilinear attention, connecting to LoRA's structure, suggesting potential for a Muon-LoRA optimizer for parameter-efficient fine-tuning.
- Due to computational hurdles, practical solutions shift to empirical approximations like gradient clipping or Kimi's MuonClip, which use batch information to approximate worst-case behavior and stabilize training.
AttoChess, a complete, playable chess program for 16-bit x86 DOS in 278 bytes
9 days ago
- AttoChess optimizes by rethinking board display, move decoding, and search loop, saving 10 bytes.
- Board display draws directly to console using int 29h, eliminating render buffer and DOS function 09h.
- Startup removes BIOS mode-set, using deterministic assumptions and cld for smaller code.
- Input decoder folds constants into base address, using 16-bit pointer math and imul for efficiency.
- Search loop frees CX by comparing pointers, avoiding depth reload from stack.
- Pawn direction is folded into color bit with xor, handling both colors in one path.
Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?
7 days ago
- Claude Fable 5 outperformed GPT-5.6 Sol on an NP-hard optimization problem, demonstrating superior consistency and producing the best overall solution.
- The /goal mode had mixed results; it improved outcomes in some runs but caused significant regressions in others, leading to worse average performance despite winning most individual trials.
- The study highlights that /goal is not a generic 'try harder' switch but alters the control loop and search path, which can amplify both good and bad decisions.
GNU C Library 2.44 Released With /etc./tunables.conf, More Optimizations
6 hours ago
- Introduced /etc/tunables.conf for system-wide tunables, more persistent than environment variable
- Implemented FMA-enhanced cosh function for ~35% performance improvement
- Imported optimized and correctly-rounded math functions from CORE-MATH project
- Added optimizations for LoongArch CPUs and introduced LoongArch32 32-bit support
- Optimized various functions on AArch64 and RISC-V

optimization

Show HN: Small World – The "Preact" of 3D Web Engines (TypeScript, WebGPU)

From Muon to Gradient Clipping: Some Thoughts on QK Stability

AttoChess, a complete, playable chess program for 16-bit x86 DOS in 278 bytes

Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?

GNU C Library 2.44 Released With /etc./tunables.conf, More Optimizations