Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-29 · source-backed
Build b10677 fixes ggml_vk_graph_optimize, where is_src_of didn't treat two views of one tensor as dependent, so the optimizer reordered nodes across aliased reads and writes. Maintainers describe the result as silently wrong tokens under greedy decoding, different output on every server start, and invalid speculative-decoding acceptance, hitting Qwen3.8's recurrent state and any model with view-aliased state. CUDA was unaffected. Nothing logged. Anyone benchmarking Vulkan inference on those models in the last stretch has results that need re-running. (GitHub)
Each link below shares sources, entities, or timing with this story.
June local-inference benchmarks across the 128GB class put NVIDIA's DGX Spark (~$4k), AMD's Strix Halo / Ryzen AI Max+ 395 (~$2 to 3k), and the M5 Max 128GB (~$5k) head to head (Hardware Corner). Prompt processing favors CUDA hard. But token generation lands at a surprisingly...
PR #19378 landed in llama.cpp this week, and I think most people are underselling what it means. Backend-agnostic tensor parallelism via --split-mode tensor makes multi-GPU inference work across AMD, Intel, and Apple Silicon. Not just CUDA. Everything. For context: llama.cpp h...
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
Huang used his inaugural X post on July 24 to publish "Open Weights and American AI Leadership," a three-page letter on Nvidia's own servers signed by 25 companies including Meta, Microsoft, IBM, Mistral, Mozilla, Hugging Face, a16z, Palantir and the Linux Foundation. Within a...
Before v4.1.15 on August 23, it only applied to tools also opted in individually, so flipping it appeared to do nothing. Anyone who left it on assuming it was inert has just widened their unattended tool surface across every connected server. (GitHub) Same release train fixed...
DeepClaude hit 470 points on Hacker News. It swaps Claude Code's API backend to DeepSeek V4 Pro while preserving the full agent loop: file editing, bash execution, git tooling, the whole workflow. DeepSeek V4 Pro scores 96.4% on LiveCodeBench at a fraction of Anthropic's prici...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.