Easy AI Daily – March 2, 2026

A digest of AI industry news from the Easy AI Daily (zhichai.net), covering models, agents, infrastructure, research, products, and policy.

Key points

• Alibaba Qwen 3.5 small models (0.8B/2B/4B/9B, Base/Instruct): native multimodal, RL-enhanced, native 262k context extendable to ~1M. Community tests show the 4B beating many older 9B models; local 128k context runs ~60 tok/s. Thinking mode is off by default and must be enabled via templates/parameters. Official release
• Qwen 3.5 27B/35B: benchmarks suggest they match or beat some 112B/120B-class models (including Minimax) on reasoning and coding; 27B heavily uses Gated DeltaNet linear attention and reportedly runs on a consumer GPU. An Opus 4.6 chain-of-thought distill also exists. 27B analysis
• Local ecosystem: Ollama, LM Studio, and Transformers.js integrations landed quickly; 9B GGUF is ~7GB, and the 0.8B runs multimodal inference in-browser via WebGPU (visual encoding is the bottleneck). NVFP4 and thinking-mode quantizations available. WebGPU demo
• Codex 5.3: leads on custom benchmarks like WeirdML; community discussion shifts from capability to usability, with some workflows delegating ~80% of code writing to models.
• BullshitBench v2: ~100 new questions across code, medicine, law, finance, physics, testing 70+ models; author claims more reasoning can increase confabulation; Anthropic models lead, OpenAI/Google generations show little progress (disputed).

Agents & tooling

• AGENTS.md experiments: across 10 repos and 124 PRs, median runtime dropped ~28.6% and token use fell 16.6%. Anthropic released a 30-page Skills guide advocating testable skill files over long prompts. AGENTS.md data
• Agent reliability: practitioners argue evaluation is harder than building agents — define success criteria first (product/domain experts), prefer deterministic scorers, use LLMs only for style, and evaluate outputs rather than reasoning traces.
• GitNexus: parses repos into interactive dependency graphs stored in embedded KuzuDB, answering queries via Cypher graph traversal instead of embeddings; fully browser-based (Web Workers), MIT-licensed.
• Stripe LLM billing proxy: pay-per-token metering with model choice and markup, feeding usage directly into Stripe billing.
• MCP vs Skills: MCP is increasingly seen as the protocol for models calling external APIs; Skills/AGENTS.md are playbooks for composing those APIs into deliverables.

Infrastructure & hardware

• Apple Neural Engine training: a researcher ran a 110M-parameter GPT training loop on an M4 ANE via undisclosed APIs, bypassing CoreML; claims ~6.6 TFLOPS/W efficiency, sparking discussion of Mac mini training clusters.
• Google Static: a sparse-matrix framework claiming 948x faster constrained decoding for LLM-based generative retrieval. Tech blog
• NVIDIA Blackwell split: data-center B100/B200 (CC 10.x, tcgen05/DPX) vs consumer RTX 50 series (CC 12.0) lack AI-specific units — future low-level optimizations will favor data-center cards.
• Taalas HC1 ASIC: weights baked into silicon, ~16–17k tok/s per user, but one chip per model — suited to large homogeneous cloud deployments.
• AMD rocprof-trace-decoder open-sourced for SQTT instruction-level traces; tinygrad's author says AMD's tracing is now better than NVIDIA's.

Research & methods

• ByteDance CUDA Agent: RL on real hardware profiling to write CUDA kernels; claims ~2x speedup over torch.compile and strong KernelBench results. Trained at large GPU cost, kernels not open-sourced. arXiv
• Databricks OAPL: sample-efficient off-policy RL reportedly needing 3x fewer generations than GRPO to train reasoning models.
• Wide-and-deep Transformer theory (Meta): 60+ pages on signal propagation, scaling, NTK analysis, and optimizer behavior, validated on vision and language Transformers.
• Sakana AI text-to-LoRA open-sourced: generates LoRAs from natural-language prompts; reproducible on one H100 in ~5 days. GitHub
• SAE analysis of diffusion models: sparse-autoencoder study shows early-step activations already predict final image composition. arXiv

Products & applications

• Local LLM progress: a ~$600 mini PC now runs Qwen3-27B Q4 quants considered comparable to early DeepSeek R1; community cautions against averaged "intelligence index" leaderboards.
• DishBrain (Cortical Labs): ~800k human/mouse neurons on silicon playing Pong and Doom.
• Runway Gen-4.5 enters the arena.ai text-to-video leaderboard at 1218, comparable to Kling-2.6-Pro.
• MLB pitcher Robert Stock built an 8.9-million-pitch analysis platform with AI tools despite no programming background.
• OpenClaw community shares: automated haircut bookings, multi-agent dashboards, persona-debate plugins — emerging as an engineer-oriented alternative to Cowork.

Policy, governance & safety

• US DoD drops Anthropic, signs with OpenAI: Anthropic labeled a "supply chain risk"; OpenAI will deploy models in classified environments. Altman's addendum bans intentional surveillance of US persons, though lawyers note incidental-collection loopholes remain. OpenAI blog
• ChatGPT backlash: uninstall volume reportedly up 295% in 48 hours post-announcement; Claude hit #1 on the US App Store.
• Moonshot/Kimi distillation controversy: reports of industrial-scale Claude distillation (150k API calls) to train Kimi, with instances of the distilled model self-identifying as DeepSeek in Chinese; communities debate where acceptable distillation boundaries lie.
• Red-team disclosure SOP: security communities circulated a playbook covering reproduction, encrypted reporting, fix windows, and coordinated disclosure.

📌 Source: Easy AI Daily