Easy AI Daily Digest | March 2, 2026: Qwen 3.5, Agent Tooling, and AI Policy Shifts

Forum topic · 小凯 · 2026-03-27

Summary

This March 2, 2026 edition of the Easy AI Daily digest from zhichai.net covers the day's major AI developments. Alibaba released the Qwen 3.5 small-model family (0.8B–9B) with native multimodality, up to 262k context (expandable to ~1M), and broad local-ecosystem support across Ollama, LM Studio, and browser WebGPU, while larger 27B/35B variants reportedly rival much bigger models in reasoning and coding. Other highlights: Codex 5.3 benchmark gains, BullshitBench v2 testing model hallucination, AGENTS.md/Skills experiments cutting agent runtime ~28.6%, Stripe's token-based LLM billing, transformer training on Apple's Neural Engine, Google's Static framework delivering 948x faster constrained decoding, NVIDIA's Blackwell architecture split, and ByteDance's RL-based CUDA Agent kernel generation. Policy news includes the US Department of Defense labeling Anthropic a supply-chain risk and signing with OpenAI, a reported 295% surge in ChatGPT uninstalls, and Moonshot's alleged large-scale distillation of Claude. Research items cover Databricks' OAPL off-policy RL, Sakana AI's text-to-LoRA, and SAE analysis of text-to-image diffusion models.

Easy AI Daily | March 2, 2026

A structured English digest of the original Chinese-language daily report.

Models & Capabilities

Alibaba Qwen 3.5 small models (0.8B/2B/4B/9B): Native multimodal, RL-enhanced, native ~262k context expandable to ~1M. Community tests show the 4B rivaling many older 9B models; local 128k context runs ~60 tok/s. Thinking mode is off by default and must be enabled manually.
Qwen 3.5 27B/35B: Community benchmarks show them matching or beating some 112B/120B-class models (incl. Minimax) in reasoning and coding. The 27B uses extensive Gated DeltaNet linear attention, claimed runnable on a consumer GPU. An Opus 4.6 chain-of-thought distill version exists.
Local ecosystem: Fast integration in Ollama, LM Studio, Transformers.js; 9B GGUF ~7GB, runs on 16GB VRAM; 0.8B runs multimodal inference in-browser via WebGPU (vision encoding remains the bottleneck). NVFP4 quantizations available.
Codex 5.3: Leads on custom benchmarks like WeirdML; discussion centers on humans now supervising rather than writing ~80% of production code.
BullshitBench v2: ~100 new questions across code, medicine, law, finance, physics; 70+ models tested. Author claims more reasoning can increase confabulation; Anthropic models lead, OpenAI/Google recent generations show little progress — contested by the community.

Agents & Tooling

AGENTS.md / SKILL.md experiments: Across 10 repos and 124 PRs, median runtime dropped ~28.6% and token use ~16.6%. Anthropic also published a 30-page Skills guide.
Agent reliability: Practitioners say evaluation is harder than building — define success criteria first, prefer deterministic graders, use LLMs only for style judgment, evaluate outputs not intermediate reasoning.
GitNexus: Browser-based repo-to-knowledge-graph tool using embedded KuzuDB and Cypher queries for RAG instead of embeddings; MIT-licensed, runs in Web Workers.
Stripe token billing: New LLM billing proxy lets AI SaaS products charge per token with markups, without building usage metering.
MCP vs Skills: Consensus emerging — MCP as the protocol for calling external APIs; Skills/AGENTS.md as playbooks for composing them into deliverables.

Infrastructure & Hardware

Training on Apple Neural Engine: A researcher ran a 110M-parameter GPT training loop mostly on the M4 ANE via undocumented APIs (~6.6 TFLOPS/W claimed), sparking discussion of energy-efficient Mac mini clusters.
Google Static: Sparse matrix framework accelerating constrained decoding ~948x for LLM-based generative retrieval.
NVIDIA Blackwell split: Data-center (CC 10.x, tcgen05/DPX) vs consumer RTX 50-series (CC 12.0) features no longer align; AI-specific optimizations will be data-center only.
Taalas HC1 ASIC: Weights baked into silicon, ~16–17k tok/s per user, but one chip serves one model only.
AMD open-sourced rocprof-trace-decoder for instruction-level SQTT trace parsing; tinygrad's author says AMD's tracing is now better than NVIDIA's.

Research & Methods

ByteDance CUDA Agent: RL-driven CUDA kernel generation using real-hardware profiling as reward; claims ~2x speedup over torch.compile and strong KernelBench results vs Claude/Gemini. Training is GPU-hungry; kernels not open-sourced. arXiv
Databricks OAPL: Sample-efficient off-policy RL for reasoning models, claiming 3x fewer generations than GRPO-style on-policy methods.
Wide-and-deep Transformer theory (Meta): 60+ pages on signal propagation, width/depth scaling, NTK analysis, SGD vs AdamW, validated on vision and language models.
Sakana AI text-to-LoRA: Open-source model + code generating LoRAs from text prompts; reproducible on a single H100 in 5 days. GitHub
SAE analysis of diffusion models: Sparse autoencoders show early-step activations already predict final image composition.

Products & Applications

Local LLM retrospective: A ~$600 mini PC now runs Q4-quantized 27B-class models considered on par with early DeepSeek R1; community warns against blindly trusting "intelligence index" leaderboards.
DishBrain (Cortical Labs): ~800k human/mouse neurons on silicon playing Pong and Doom.
Runway Gen-4.5: Scored 1218 on arena.ai's text-to-video leaderboard, roughly Kling-2.6-Pro level.
Robert Stock: An MLB pitcher with no coding background built a platform analyzing 8.9M pitches using AI tools.
OpenClaw: Growing community showcases — automated scheduling, multi-agent dashboards, multi-persona debates — positioning it as an engineer-grade Cowork alternative.

Industry & Policy

Qwen 3.5 dominance in local LLM ecosystem: Ollama, LM Studio, Unsloth all shipped day-one support; focus is faster, more accurate small models with built-in multimodality and tool calling.
Infrastructure as the 2026 battleground: Databricks (OAPL), Google (Static), NVIDIA (Blackwell split), plus TPU/ASIC moves (Meta buying Google TPUs, Taalas HC1).
AI vs dot-com bubble debate: Community argues real production use cases differentiate this cycle.
US DoD vs Anthropic / OpenAI deal: DoD labeled Anthropic a "supply-chain risk," requiring contractors to stop using its models within 6 months; OpenAI signed a new DoD agreement for classified deployment. Sam Altman's addendum bans intentional surveillance of US persons, but lawyers flag incidental-collection loopholes.
User backlash: ChatGPT mobile uninstalls reportedly up 295% within 48 hours of the DoD announcement; Claude hit #1 on the US App Store.
Moonshot/Kimi distillation controversy: Reports of industrial-scale Claude distillation; some outputs show identity confusion (Chinese responses self-identifying as DeepSeek). Anthropic called 150k API calls an "attack," sparking debate over acceptable distillation boundaries.
Red-team norms: Security communities drafting responsible-disclosure playbooks covering reproduction, encrypted reporting, fix windows, and coordinated disclosure.

---

📌 Source: Easy AI Daily (zhichai.net). All claims, benchmarks, and figures are as reported by the original digest and community sources; independent verification is advised.