Easy AI Daily News Digest – March 2, 2026: Qwen 3.5, Codex 5.3, Agent Tooling, Hardware, and Policy

Forum topic · 小凯 · 2026-03-27

Summary

A digest of AI industry news for March 2, 2026. Alibaba released the Qwen 3.5 small-model family (0.8B–9B) with native multimodality and up to ~1M token context, which was rapidly integrated into Ollama, LM Studio, and browser-based WebGPU demos; community benchmarks show the 27B/35B models rivaling 112B+ class models in reasoning and coding. Other highlights: Codex 5.3 tops custom coding benchmarks; AGENTS.md experiments cut agent runtime by ~28.6%; Stripe launched token-based LLM billing; GitNexus turns repos into browser-side knowledge graphs. On infrastructure, researchers ran a 110M-parameter GPT training loop on Apple's Neural Engine, Google announced its Static framework claiming 948x faster constrained decoding, and NVIDIA's Blackwell datacenter/consumer split became explicit. Research news includes ByteDance's RL-based CUDA Agent, Databricks' sample-efficient OAPL, and Sakana AI's open-source text-to-LoRA. Policy sections cover the US Department of Defense dropping Anthropic for OpenAI, a 295% ChatGPT uninstall spike, and Moonshot's large-scale Claude distillation controversy.

Models & Capabilities

Alibaba releases Qwen 3.5 small models: Qwen3.5-0.8B/2B/4B/9B (Base/Instruct) with native multimodality, RL enhancement, and 262K native context (expandable to ~1M). Community tests show the 4B beating many two-year-old 9B models; local 128K context runs at ~60 tok/s. Thinking mode is off by default.
Official release | Architecture discussion | Reddit benchmarks
Qwen 3.5 27B/35B punch above their weight: Community benchmarks show them matching or beating some 112B/120B-class models on reasoning and coding. The 27B uses Gated DeltaNet linear attention and reportedly runs on a single consumer GPU.
27B analysis | Opus 4.6 distill
Local ecosystem: Ollama, LM Studio, Transformers.js all integrated; 9B GGUF ~7GB, 0.8B runs multimodal inference in-browser via WebGPU.
WebGPU demo
Codex 5.3: Leads WeirdML-style benchmarks, more stable on multi-step tasks; discussion shifts from "can it code" to "80% of code is now model-written, humans supervise."
BullshitBench v2: ~100 new questions across code/medicine/law/finance/physics across 70+ models; author claims more reasoning can increase confabulation; Anthropic models lead, OpenAI/Google reportedly flat — contested by the community.

Agents & Tooling

AGENTS.md / SKILL.md in practice: Across 10 repos and 124 PRs, median runtime dropped ~28.6% and token use 16.6%. Anthropic published a 30-page Skills guide.
Skills guide PDF
Agent evaluation is harder than building agents: Advice — define success criteria first (domain experts lead), prefer deterministic scorers, use LLM judges only for style, evaluate outputs not reasoning traces.
GitNexus: Parses repos into interactive dependency graphs stored in embedded KuzuDB, using Cypher graph queries instead of embedding similarity; fully in-browser, MIT licensed.
Stripe token-based LLM billing: Route calls through Stripe's LLM proxy to bill usage with markup automatically.
MCP vs Skills: Emerging consensus — MCP as the unified protocol for tool/API calls; Skills/AGENTS.md as playbooks for composing them.

Infrastructure & Hardware

Training on Apple Neural Engine: A researcher trained a 110M-param GPT on M4's ANE via undocumented APIs (~6.6 TFLOPS/W claimed), sparking debate about Mac mini training clusters. Details
Google Static: Sparse-matrix framework claiming 948x faster constrained decoding for generative retrieval. Blog
Blackwell split: Datacenter (CC 10.x, tcgen05/DPX) vs consumer RTX 50-series (CC 12.0) no longer feature-parity.
Taalas HC1 ASIC: Weights baked into silicon, ~16–17k tok/s per user, but one chip serves one model only.
AMD rocprof-trace-decoder open-sourced: Instruction-level SQTT trace decoding; tinygrad calls AMD's tracing infrastructure better than NVIDIA's.

Research & Methods

ByteDance CUDA Agent: RL agents profiling on real hardware to write CUDA kernels; claims ~2x over torch.compile, strong on KernelBench. Paper: https://arxiv.org/abs/2602.24286
Databricks OAPL: Sample-efficient off-policy RL for reasoning models — 3x fewer generations than GRPO-style on-policy methods.
Meta's wide-and-deep Transformer effective theory resurfaces: 60+ pages on signal propagation, scaling, NTK, SGD vs AdamW.
Sakana AI text-to-LoRA: Open-source model generating LoRAs from natural-language prompts; reproducible on one H100 in 5 days. GitHub
SAE analysis of text-to-image diffusion: Early-step activations already predict final image composition — layout is decided early, later steps refine textures. Paper: https://arxiv.org/abs/2504.15473

Products & Applications

Local LLMs one year after DeepSeek: A $600 mini PC now runs Qwen3-27B Q4 comparably to the original R1; skepticism toward "intelligence index" leaderboards grows.
DishBrain: ~800K living human/mouse neurons on a chip play Pong and Doom.
Runway Gen-4.5 enters the text-to-video arena at 1218 Elo, near Kling-2.6-Pro. Leaderboard
MLB pitcher Robert Stock built an 8.9M-pitch analysis platform with no coding background using AI tools.
OpenClaw ecosystem: Real-world automations from barbershop booking to multi-agent dashboards; positioning as an engineer-friendly Cowork alternative.

Industry & Policy

US DoD drops Anthropic, signs with OpenAI: Anthropic labeled a "supply chain risk"; OpenAI deploys in classified environments with Altman-published clauses barring intentional surveillance of US persons — critics note the "incidental collection" loophole.
User backlash: ChatGPT mobile uninstalls up 295% in 48 hours; Claude hits #1 on the US App Store.
Moonshot/Kimi distillation controversy: Industrial-scale Claude distillation alleged; distilled Claude reportedly self-identifies as DeepSeek in Chinese. Anthropic calling 150K API calls an "attack" draws pushback about where acceptable distillation boundaries lie.
Responsible red-teaming: Security communities draft full disclosure SOPs — reproduction, encrypted reporting, fix windows, coordinated disclosure, export-control awareness.

---

📌 Source: Easy AI Daily