Researchers introduce δ-mem, a lightweight memory mechanism that compresses historical information into a fixed-size 8x8 state matrix, delivering 1.31x improvement on memory-intensive benchmarks without fine-tuning or expanding the context window. The method works alongside a frozen full-attention backbone using delta-rule learning and low-rank corrections, offering a practical solution to the persistent memory limitations of current LLMs.
NVIDIA Labs has released SANA-WM, an open-source world model with just 2.6 billion parameters that can generate up to one minute of 720p video. The model pushes the frontier of efficient video generation, achieving minute-scale coherent world modeling at a fraction of the parameter count of competing approaches.
Noah Golmant has released version 1.0 of pytorch-hessian-eigenthings, a library for efficiently computing Hessian eigendecompositions of neural networks without quadratic memory costs. The rewrite adds Lanczos, Hutch++ trace estimation, spectral density analysis, and fused Triton kernels for language models, incorporating ideas from PyHessian, curvlinops, and HessFormer.
Accelerate is a Haskell library for expressing parallel array computations that compile and run on multicore CPUs via LLVM and NVIDIA GPUs via CUDA. It provides a rich ecosystem including FFT, BLAS, and graphics operations with type-safe parallel programming and online compilation targeting specific hardware.
Futhark is a purely functional data-parallel array programming language designed for GPU-accelerated high-performance computing. Its examples page showcases 50+ documented programs from basic language features through automatic differentiation and literate programming, demonstrating real-world uses in ray tracers, particle simulators, and cryptographic implementations.
Charles Stross's 2005 novel Accelerando — available free online — is a landmark science fiction work exploring the technological singularity, post-human economics, and runaway AI development across three generations. It's resurfaced on HN, likely because its themes of recursive self-improvement and machine intelligence feel increasingly prescient in 2026.
Project Gutenberg, the volunteer-driven digital library founded in 1971, now hosts over 75,000 free eBooks and recently expanded with nearly 5,000 computer-generated audiobook titles produced through collaboration with Microsoft and MIT. The AI-generated narration initiative represents a compelling intersection of open-source principles and modern text-to-speech technology.
A developer created illustrated flashcards for children learning the Greek alphabet, where each card's image both starts with the letter and visually resembles the letter's shape. The project used a 35,000-word Greek dictionary, AI-assisted visual matching, and AI-generated Eric Carle-style illustrations to produce the final card deck.
Julia Evans details her migration from Tailwind to semantic HTML with vanilla CSS, motivated by build system dependencies, 2.8MB file bloat, and a desire for more control. She developed a nine-part organizational framework including component-scoped CSS, centralized color variables, CSS Grid for responsive layouts, and utility classes — effectively recreating Tailwind's best constraints without the framework.
A collaborative editor would silently stop saving when users edited emoji in specific ways — the root cause being that Yjs's splice method could split emoji into orphaned UTF-16 surrogate pairs, crashing encodeURIComponent. The fix and broader lesson: JavaScript's default string operations work on 16-bit code units, not characters, so any code doing str[0] on emoji is potentially broken.
