Chinese Startup CosmosMind Claims Agent-Designed Photonics Chip and Faster Ascend Kernels
jwt0625 · x · 2026-11-17
A long thread contrasts OpenAI's admission that GPT-6 Astra failed to respect PCB manufacturing tolerances on its TamaGPT voice toy (37% on RWE-Bench) with Chinese startup CosmosMind, which claims its MetaRSI agents plugged directly into fab process tolerances and real NPU racks:
- Agent-designed photonics chip taped out: 31 SEM scans confirmed line-width variation under 5nm (half their 10nm limit), -0.66 dB coupling loss; agents wired 3 variants of 1.6 Tbps (8x200G) optical links at 1.46 pJ/bit wall-to-wall
- Ascend 910C kernels: DeepSeek's official V4.1 Ascend C port doesn't compile on 910C (targets the newer 950); CosmosMind's agents wrote PyPTO-FlashDSA from scratch and beat the official fp32 reference decode kernel (54.45µs vs 56.49µs), with some Qwen3 patches merged upstream into Huawei's pypto-lib
- MiniMax H3 port: fused 8,000 kernel ops into 1 and pushed the 33B video+audio model to 450% of MiniMax's own release after it lost 71% of speed off CUDA
The author cites domestic chips passing 41% of China's AI accelerator shipments in 2025, and notes the 0-star GitHub repo is being ignored by Western AI Twitter. Some claims are unverified.
Related event: Chinese Startup CosmosMind Reportedly Tapes Out AI-Designed Photonic Chip(3 posts)→
More from coding & agent
- IBM's Tejas Kumar on Building Agent Harnesses From Scratch: Guardrails and Daiku — TejasKumar_ · 2026-11-17
- Cohere Labs-backed FeatureGraph standardizes scientific time-series analysis into a single source of truth — Cohere_Labs · 2026-11-17
- Idea: Magic Effects That Generate Code at Training Time, Run It at Execution — SubtleGradient · 2026-11-17
- Same Weights, 5x Cost Gap: HF Engineer's Agent Harness Workshops at AI Dev NYC — DeepLearningAI · 2026-11-17
- A practical CI prompt: make your tests and builds use every CPU core on the runner — DanielLockyer · 2026-11-17
- The boring work behind successful AI agents: data and process cleanup decide outcomes — bigdata · 2026-11-17