A three-link chain of mmap → MTLBuffer(bytesNoCopy) → Wasmtime MemoryCreator that makes a Wasm linear memory share the same physical bytes as a Metal GPU buffer. Llama 3.2 1B runs at 9ms/token on M1.
Two simultaneous announcements from Cloudflare Agents Week 2026: Agent Memory manages agent recall via Durable Objects, Vectorize, and Workers AI, while isitagentready.com scores how well sites are prepared for agents.
Alibaba's Qwen team released Qwen3.6-35B-A3B as open weights. A 40-layer hybrid of Gated DeltaNet, Gated Attention, and MoE hits 73.4 on SWE-bench Verified, 37.0 on MCPMark, and 1397 on QwenWebBench.
Running local models and still getting refusals? LLM safety stacks five layers: input filter, system prompt, RLHF, Constitutional AI, output filter. Where abliterated and uncensored variants cut, and which blocks survive on Gemini, Claude, and local LLMs.
Google DeepMind's AI writing tool Fabula was demoed at CHI 2026 by Piotr Mirowski. Co-designed with 42 professional writers using convergent iteration for story structure. But the timeline shows Fabula was first demoed in May 2025 and entered early access in September 2025 — still a research prototype with no general availability.
Bryan Cantrill's 'The Peril of Laziness Lost' argues that LLMs have zero cost to write code and no motivation to abstract. Humans must serve as the 'deletion engine' or systems will bloat endlessly.
Four Japanese tech giants form a new company backed by mega-banks and Nippon Steel to build a trillion-parameter foundation model for physical AI, with roughly ¥3 trillion in combined public-private funding.
MegaTrain flips the GPU-centric paradigm by treating CPU memory as primary storage and the GPU as a transient compute device, enabling full-precision training of 100B+ LLMs on a single GPU with up to 12.2x throughput over DeepSpeed ZeRO-3.
Meta unveils Muse Spark, the first model from its new Meta Superintelligence Labs. The Scale AI acquisition, the shift from open-weight to proprietary, multi-agent reasoning via Contemplating mode, and the evaluation awareness problem.
Zhipu AI's GLM-5.1 is a 744B MoE (40B active, 200K context, MIT) targeting long-horizon agent tasks. Hits 58.4% SOTA on SWE-Bench Pro (edging out GPT-5.4 and Claude Opus 4.6) and sustains performance across 8-hour sessions with 6,000+ tool calls without degradation.
9 Japanese-specialized LLMs as of April 2026 — LLM-jp-4 (11.7T tokens from scratch), PLaMo, Nemotron Nano 9B JP (#1 sub-10B on Nejumi 4), Swallow 30B-A3B, Namazu — broken down by whether they were scratch-trained, continued pre-trained, or post-trained, with size, license, benchmark scores.
Benchmarking NII's LLM-jp-4-32B-A3B-thinking on EVO-X2 (Ryzen AI Max+ 395) with ROCm. 62.9 t/s vs Qwen3.5-35B-A3B's 44.7 t/s. Covers thinking control issues, KV cache trade-offs, knowledge cutoff, Japanese quality comparisons, code generation tests, and training data composition.
