How I Fight AI Brain Fry
Article URL: https://sergeykuhne.com/blog/fighting-ai-brain-fry Comments URL: https://news.ycombinator.com/item?id=50416363 Points: 83 # Comments: 56
AI NEWS JOURNAL
Search AI news
Search headlines, summaries and articles to connect the dots.
Show HN: SafeScript, a safe version of JavaScript for executing untrusted code
Article URL: https://github.com/SafeScript/safescript Comments URL: https://news.ycombinator.com/item?id=50417478 Points: 1 # Comments: 0
b12208
The ggml library has been updated to make tied selections in the radix Top-K algorithm deterministic on CUDA and Metal backends by replacing atomic compaction loops with a deterministic scheme using simdgroup scans and coalesced memory accesses. This change ensures byte-identical outputs across runs, improves performance by 20-45% on M5 Ultra hardware, and enables deterministic testing for Metal and ROCm. Redundant CUDA kernel resets were removed, and Vulkan/Metal changes were reduced or skipped.
Show HN: LLM weight editing tool made by a 15-year-old
Article URL: https://github.com/Miyamoto-Renya/LaplacesScissors Comments URL: https://news.ycombinator.com/item?id=50409986 Points: 1 # Comments: 0
Show HN: Nino – Fastest C# serializer, safe on untrusted input, also TS/Go/Rust
Nino is a source-generated binary serializer for C#, and as far as we know the fastest one. In the same BenchmarkDotNet run on GitHub runners against MemoryPack 1.21.4, it serializes classes and polymorphic types 3.5–7.2x faster and deserializes them 1.7–2.1x faster (arrays of unmanaged structs are on par). Deserializing into existing objects allocates nothing. Reproduce: https://nino.xgamedev.net/en/docs/v4/benchmarksIt is built to read bytes you do not trust: lengths are checked before allocating, depth is bounded, hash flooding is limited, and the read paths are fuzzed in CI.The bytes are a spec, not a memory dump, so CoreCLR, Mono, IL2CPP and 32-bit hosts write the same bytes, and runtimes for TypeScript, Go and Rust pass the same conformance corpus.Trade-offs: positional by default (VersionTolerant or Keyed for stored data), a 4-byte type id on polymorphic values, no System.Type, needs C# 9 and Roslyn 4.3+.Docs: https://nino.xgamedev.net Comments URL: https://news.ycombinator.com/item?id=50413601 Points: 1 # Comments: 0
Show HN: TheySay – Turn Your Social Media Feeds into a Dashboard
Article URL: https://github.com/laike9m/TheySay Comments URL: https://news.ycombinator.com/item?id=50407700 Points: 1 # Comments: 0
b12175
A fix was implemented to override the hardware_concurrency value in CI tests, addressing an issue where test-llama-context and test-backend-ops hung due to incorrect thread count estimation on 8-CPU runners. The solution sets default values of 8 for TEST_ARG_HARDWARE_CONCURRENCY and 4 for TEST_ARG_PARALLEL_WORKERS specifically for the cpu-x64-high-perf configuration.
b12179
The update introduces support for 3D convolution (conv_3d) on Hexagon hardware by generalizing the existing conv_2d implementation. It leverages the HMX tiled path, folds kernel dimensions into a virtual row, and uses a shared solver with an is3d flag. The host computes unfold divisors to avoid device-side division, prevents DMA queue overflow, and improves tile size selection based on VTCM usage. Additional changes include handling arbitrary positive strides, preserving load/shuffle paths for small strides, and fixing compiler warnings.
b12184
The llama.cpp release b12184 includes a fix to add a missing link to pthread and privately link Threads::Threads, addressing potential runtime issues. The release provides precompiled binaries for multiple platforms including macOS, Linux, Windows, Android, and openEuler, supporting various backends such as CUDA, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon-specific hardware. Additional UI components are also available.
b12189
The metal backend's kernel_group_norm_f32 was updated to correctly handle tensors with ne03 > 1 and non-contiguous inputs by dispatching one threadgroup per (group, i03) pair, computing per-batch statistics, clamping group element count for incomplete groups, and supporting strided access via separate strides for source and destination. The fix also restricts GROUP_NORM to f32 inputs and improves performance through specialized pipeline variants.
Show HN: Combining Bursts of Images
An approach of taking many photos to enhance one final image is a technique called "stacking". Stacking alone doesn't account for everything so this is a take that helps get better alignment of images, allows comparisons, and drop bad images that dont match. Its not super resolution but it feels like when you look at it. it reduces the noise and makes it look more organic shot with less expensive hardware. enjoy! Comments URL: https://news.ycombinator.com/item?id=50401218 Points: 1 # Comments: 0
Show HN: I taught my Fujifilm X-T30 II to print on an instax Link 3 over BLE
Article URL: https://ronangaillard.github.io/posts/fujifilm-x-t30-ii-instax-link-3/ Comments URL: https://news.ycombinator.com/item?id=50403007 Points: 3 # Comments: 0
Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster
Zyphra released ZYFA, open-source FP8 and BF16 prefill attention kernels for AMD Instinct MI355X. On real model activations, its FP8 kernels run 1.16-1.76x faster than AMD’s AITER, with 62-78% lower error at head dimension 128. The post Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster appeared first on MarkTechPost.
Agents Are Going Async
Async subagents let an agent stay responsive while background workers run long tasks. How the orchestrator, notifications and approvals work in Deep Agents.
NUMA-Torch: Compiler-Driven NUMA-Aware LLM Prefill on Multi-Socket CPUs
arXiv:2611.07105v1 Announce Type: new Abstract: For sufficiently large transformer-based Large Language Models (LLMs), whose parameters cannot be retained in on-chip caches, prefill throughput on multi-socket CPUs is often limited by sustainable DRAM bandwidth rather than peak floating-point throughput. In the standard PyTorch compilation (i.e., TorchInductor) path, model weights are allocated without considering the Non-Uniform Memory Access (NUMA) topology: specifically, the affinity between memory banks and compute cores is disregarded. This causes a substantial fraction of weight accesses to traverse the inter-socket interconnect, severely degrading the effective memory bandwidth available to each socket. We address this bottleneck through a compiler-level solution that makes TorchInductor NUMA-aware. Treating each NUMA socket as an independent tensor-parallel partition, we introduce NUMA-aware Tensor Parallelism (NUMA-TP). This approach transparently shards weights across sockets and schedules compute kernels such that thread pools and memory allocations remain strictly local to their respective sockets using CPU affinity. We evaluate the approach on GPT-2 (117M, 345M, 762M) and BERT (Small, Large), which exercise standard multi-head attention and MLP, as well as Gemma 3 270M, which combines Multi-Query Attention with a GeGLU-gated MLP. By strictly enforcing data locality, our approach successfully mitigates cross-socket traffic, achieving prefill speedups of up to 2.63x on a dual-socket Haswell platform, up to 1.87x on a dual-socket Cascade Lake platform, and up to 2.27x on a four-socket Haswell-EX platform.
Show HN: Eski – an in-memory store in Odin, measured against Redis on one core
Article URL: https://github.com/vighnesh-bheed/eski Comments URL: https://news.ycombinator.com/item?id=50389748 Points: 1 # Comments: 0
Show HN: Repairer – close the breach when a Linux flaw goes public
As we've seen with the Ubuntu critical CVE-2026-98432 (CVSS 9.8), AI-dicovered zero day exploit are being reported at an exponential rate and we don't have the process for dealing with the flood of exploits in real time.What we need is an AI-based system for autopatching systems based on real time information. The vendor's fix can take days or weeks to be out (is the Ubuntu thing patched yet?) and meanwhile the grid is vulnerable to exploitation, especially AI worms.Repairer is a small program that runs on each of your servers. The day an exploit goes public it puts up a stopgap that blocks it, and once the real fix is up it takes the stopgap back off by itself.A stopgap only takes something away — drop certain packets, turn off a kernel module or service, remove a setuid bit. It never opens a port, installs or downloads anything, nor runs code. It arms an undo playbook before applying the stopgap, so the change reverts if affects any services.It only ever acts on rules you sign with a (hardware) crypto wallet.Let's go further: stopgaps are written and checked by a swarm of AI agents (gnosh.ai sessions) running on gno.land, each with bonded stake. Each stopgap is run on fresh machines, and whoever is caught lying loses their stake to whoever caught them.The agent is built and tested on Ubuntu and Debian, all open source, standard-library Go. More distros coming. Let's build repairer together.https://gno.security/repairer/ https://github.com/gnosecurity/repairer Comments URL: https://news.ycombinator.com/item?id=50388417 Points: 1 # Comments: 1
NVIDIA CUDA 13.5 Brings NVIDIA Vera Rubin GA Support
As artificial intelligence and accelerated computing continue to push the boundaries of what's possible, the latest NVIDIA CUDA release arrives to unlock...
Show HN: Edgepad – Turn an Android phone into a trackpad for your laptop
Article URL: https://github.com/akshit-bansal11/edgepad Comments URL: https://news.ycombinator.com/item?id=50374433 Points: 1 # Comments: 1
Show HN: Homotope, vectorize images to SVG in the browser, MIT engine on NPM
Article URL: https://github.com/stevenroman91/homotope-engine Comments URL: https://news.ycombinator.com/item?id=50370248 Points: 1 # Comments: 0