Show HN: Infinate –O(k)constant-time spatial attention for unlimited LLM context

1 point

5 months ago

I built infinate, an open-source (Apache 2.0) attention mechanism that places tokens in a 3D semantic space and limits attention to ~50 nearest neighbors via exponential decay + hard cutoff. This gives true O(k) constant complexity: latency ~7–14 ms and memory fixed (1.50 MB in production) even at 10M+ tokens on CPU.

Latest update (Jan 20, 2026): Milestone 1.11 – Strafe Jumping Navigation. As a longtime Quake player, I applied real game physics exploits (7/9 validated) — bunny hop, circle jump, warp lanes, momentum accumulation, LOD hopping, etc. — to make semantic traversal ultra-fast, like speedrunning in 3D space.

Benchmarks (all CPU, Qdrant vector store): - In-memory mode: - CodeQA (100K): 3.57 ms vs MIT RLM 15,000 ms → 4,198× faster - OOLONG (500K): 4.06 ms vs 35,000 ms → 8,628× faster - BrowseComp+ (10M): 7.18 ms vs 120,000 ms → 16,722× faster - Average speedup: 10,317× vs MIT Recursive Language Models (arXiv:2512.24601) - Production (Docker/Qdrant): Average 533× faster - Memory: Constant 1.50 MB in container mode (62.2% less than in-memory; 10× tokens → 0.96× memory) - O(k) scaling: 20× tokens → 2.85× time increase (vs 400× for O(n²)) - Cost: 1,330× cheaper than MIT RLM

Repo: https://github.com/ch1pu/infinate (Python/PyTorch backend, Qdrant/pgvector adapters, 369 tests @99.2% pass rate, 89.58% coverage). GPU-native design (local neighborhoods suit warp parallelism) — Blackwell sm_120 kernels planned next.

I'm a solo dev (Navy vet, Uber driver, Quake player). The spatial 3D + physics-inspired navigation feels like an unusual but effective combo. Does it hold up technically? Any obvious flaws, better ways to exploit the 3D space, or integration ideas with existing LLMs?

Curious for HN feedback — questions, critiques, suggestions welcome.