Easy AI Daily News Digest – March 2, 2026

A translation and digest of the Easy AI Daily (zhi forum: zhichai.net) covering AI models, agents, infrastructure, research, products, and policy.

Models & Capabilities

Alibaba releases Qwen 3.5 small-model family: multimodal + long context, built for local & edge

Alibaba launched Qwen3.5-0.8B/2B/4B/9B (Base/Instruct) with native multimodality and RL enhancement, claiming 262k native context extendable to ~1M. Community tests suggest the 4B beats many 9B models from two years ago; local 128k context runs at ~60 tok/s. "Thinking mode" is off by default and must be enabled via template/parameters.

• Official release tweet
• Architecture & Gated DeltaNet discussion
• Local benchmarks & long-context thread (Reddit)

Qwen 3.5 large models: 27B/35B challenge 100B-class models on reasoning and coding

Community benchmarks show Qwen3.5-27B/35B matching or exceeding some 112B/120B-class models (including Minimax) on complex reasoning and coding. The 27B heavily uses Gated DeltaNet linear attention and reportedly runs on a single consumer GPU. A Claude Opus 4.6 chain-of-thought distill completed embedded Tetris in 4 seconds.
• 27B architecture analysis
• 9B/4B benchmark charts
• Opus 4.6 distill of Qwen3.5-27B

Qwen 3.5 local ecosystem: Ollama, LM Studio, browser WebGPU all onboard

The 9B GGUF is ~7GB and runs on 16GB VRAM or high-end laptops; the 0.8B runs multimodal inference in-browser via WebGPU (visual encoding is the bottleneck). LM Studio and Unsloth released quantized versions (NVFP4, thinking-mode optimizations).
• Ollama integration | WebGPU demo | NVFP4 quantization collection

Codex 5.3: coding intelligence climbs; debate shifts to "usability"

Codex 5.3 leads on benchmarks like WeirdML and is more stable on complex multi-step tasks, trading wins with top Gemini versions. Discussion now centers on workflows where models write ~80% of code while humans supervise and patch.
• WeirdML scores discussion | Claude Code writing 80% of production code

BullshitBench v2: a benchmark for measuring confabulation

Adds ~100 questions across code, medicine, law, finance, physics, testing 70+ models. The author claims more reasoning can increase confabulation; Anthropic models lead, while recent OpenAI/Google generations show little improvement—drawing pushback.
• Announcement | Community skepticism

Agents & Tooling

• AGENTS.md / SKILL.md in practice: Across 10 repos and 124 PRs, median runtime dropped ~28.6% and token usage 16.6%, mainly by eliminating worst-case thrashing. Anthropic also published a 30-page Skills guide emphasizing decomposing workflows into testable skill files. (Experiment data | Anthropic Skills guide PDF)
• Agent reliability & observability: Practitioners say evaluation is harder than building agents—define success criteria first (product/domain experts), prefer deterministic scorers, use LLMs only for style, and evaluate outputs not intermediate reasoning. (Observability question | Evaluation practices thread)
• GitNexus: Parses repos into interactive dependency graphs stored in embedded KuzuDB, answering queries via Cypher graph traversal instead of embedding similarity. Fully browser-based (Web Workers), MIT licensed. (Feature intro)
• Stripe LLM billing proxy: Per-token billing—choose model, set markup, route calls through Stripe; usage auto-bills, no custom metering needed. (Product intro)
• MCP vs Skills: Consensus emerging that MCP is the protocol for model-to-API calls, while Skills/AGENTS.md are playbooks for composing APIs into deliverables. (Weaviate skills example)

Infrastructure & Hardware

• Transformer training on Apple Neural Engine: A researcher ran a 110M-param GPT training loop on M4's ANE via undocumented APIs (bypassing CoreML); claims ~6.6 TFLOPS/W efficiency, sparking discussion of Mac mini training clusters. (Technical thread | Commentary)
• Google Static: Sparse-matrix framework accelerating constrained decoding up to 948x for LLM-based generative retrieval. (Blog)
• NVIDIA Blackwell split: Data-center B100/B200 are Compute Capability 10.x with tcgen05/DPX; consumer RTX 50-series is CC 12.0 without these AI units—future low-level optimizations will favor data-center cards. (NVIDIA blog)
• Taalas HC1 ASIC: Weights baked into silicon, no HBM; up to ~16–17k tok/s per user, but one chip serves one model only—suited to homogeneous cloud deployment, not general GPU replacement. (Explainer)
• AMD open-sources rocprof-trace-decoder: Parses SQTT instruction-level traces; tinygrad's author says AMD's tracing infrastructure is now better than NVIDIA's. (Comment)

Research & Methods

• ByteDance CUDA Agent: RL agent profiles real hardware to auto-write CUDA kernels, claimed ~2x faster than torch.compile and far above Claude/Gemini on KernelBench; costly to train, kernels not yet open-sourced. (Thread | arXiv)
• Databricks OAPL: Off-policy RL (Optimal Advantage-based Policy Optimization with Lagged Inference) claims 3x fewer generations than GRPO to train reasoning models. (Intro thread)
• Wide-and-deep Transformer theory revisited: Meta's 60+ page effective theory of signal propagation, width/depth scaling, NTK analysis, SGD vs AdamW, validated on vision and language Transformers. (Summary | arXiv)
• Sakana AI open-sources text-to-LoRA: Generates LoRAs from natural-language descriptions; reproducible on one H100 in 5 days. (GitHub | HF model)
• SAE analysis of image diffusion models: Sparse autoencoders show early-step activations already predict final image composition—layout is decided early; later steps mostly refine texture. (Paper)

Products & Applications

• Local LLM retrospective: 13 months after DeepSeek's moment, a $600 mini-PC runs Qwen3-27B Q4 quants considered comparable to original DeepSeek R1; users also dismiss "intelligence index" leaderboards as averaged MMLU/GPQA. (Reddit thread)
• DishBrain: Cortical Labs connected ~800k human+mouse neurons to silicon; plays Pong and Doom, blurring lines on general-purpose computing. (Demo)
• Runway Gen-4.5 enters arena.ai text-to-video leaderboard at 1218, roughly Kling-2.6-Pro level. (Leaderboard)
• MLB pitcher Robert Stock built an AI platform analyzing 8.9M pitches with zero programming background. (Post)
• OpenClaw ecosystem: Community shares auto-booking haircuts, multi-agent dashboards with orchestrator backends, and multi-persona debates with local file access—becoming an engineer-focused Cowork alternative. (Showcase)

Industry & Business

• Qwen 3.5 energizes the local LLM chain: Ollama, LM Studio, Unsloth all shipped support within days; the focus is faster/more accurate than previous 7B/13B at same hardware, with built-in multimodality and tool calling. (HF collection)
• Infrastructure is the 2026 keyword: Databricks (OAPL), Google (Static), NVIDIA (Blackwell split), plus ASIC/TPU moves (Meta buying Google TPUs, Taalas HC1)—consensus that the coming race is compute + infrastructure, not leaderboard chasing. (Meta-Google TPU report)
• AI economy vs dot-com bubble: Communities note similarities to 2000 (fast burn, extreme sentiment) but argue real production use cases (code, support, content) exist this time, with many running Hermes/Qwen as daily workhorses.

Policy, Governance & Safety

• US DoD designates Anthropic a "supply-chain risk": Contractors must stop using its models within 6 months, citing insufficient access. OpenAI then signed a new DoD agreement for classified deployments; Sam Altman's supplemental terms prohibit "intentional surveillance of individuals inside the US," but lawyers note the "incidental collection" loophole remains. (DoD statement | OpenAI blog | Altman's amended terms)
• ChatGPT uninstall surge; Claude tops App Store: TechCrunch reports 295% increase in ChatGPT mobile uninstalls within 48 hours of the DoD announcement; Claude hit #1 on the US App Store. Some expect the boycott to be short-lived. (Data discussion | Claude ranking)
• Moonshot/Kimi distilling Claude: Anthropic called 150k API calls an "industrial-scale distillation attack"; distilled Claude outputs sometimes self-identify as DeepSeek in Chinese. Communities debate where ToS, quota, and acceptable distillation boundaries lie. (Case analysis)
• Responsible jailbreak disclosure: Red-team communities drafted a full playbook—reproduction, encrypted reporting, fix windows, coordinated disclosure, written scope, harm minimization, export-control awareness. (Playbook gist)

---

📌 Source: Easy AI Daily