Easy AI Daily | 2026-03-02

English translation of the March 2, 2026 AI industry roundup.

Models & Capabilities

Alibaba releases Qwen 3.5 small model family: multimodal + ultra-long context for local and edge

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

• Official release tweet
• Gated DeltaNet architecture discussion
• Local benchmarks (Reddit)

Qwen 3.5 27B/35B: reasoning and coding punch above weight

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

Local ecosystem: Ollama, LM Studio, browser WebGPU

Qwen 3.5 small models were rapidly integrated into Ollama, LM Studio, and Transformers.js. The 9B GGUF is ~7GB and runs on 16GB VRAM or high-end laptops; the 0.8B runs multimodal inference in-browser via WebGPU (vision encoder remains the bottleneck). LM Studio and Unsloth shipped NVFP4 and thinking-mode-optimized quantizations.
• Ollama integration | WebGPU demo | NVFP4 quants

Codex 5.3: focus shifts from "can it code" to "is it usable"

Codex 5.3 leads on custom benchmarks like WeirdML, handles complex multi-step tasks more reliably, and trades wins with top Gemini versions. Discussion centers on workflows where models now write ~80% of code, with humans supervising and patching.
• WeirdML results | 80% production code discussion

BullshitBench v2: benchmark for hallucination updated

Adds ~100 questions spanning code, medicine, law, finance, physics across 70+ models. Author claims more reasoning can increase confabulation; Anthropic models lead clearly, while recent OpenAI/Google generations show little progress — sparking pushback.
• Announcement | Community criticism

Agents & Tooling

• AGENTS.md/SKILL.md in practice: Across 10 repos and 124 PRs, median runtime dropped ~28.6% and token use fell 16.6%, mainly by eliminating the worst flailing attempts. Anthropic also published a 30-page Skills guide emphasizing testable skill files over long prompts. (Data | Skills guide PDF)
• Agent reliability & observability: Practitioners say evaluation — not building agents — is the hard part: define success criteria first (domain experts lead), prefer deterministic scorers, use LLM judges only for style, and evaluate outputs not reasoning traces. (Thread)
• GitNexus: Parses repos into interactive dependency graphs stored in embedded KuzuDB, answering queries via Cypher graph traversal instead of embedding similarity. Fully in-browser via Web Worker, MIT licensed. (Intro)
• Stripe LLM billing proxy: Token-based billing — pick a model, set markup, route calls through Stripe; usage is invoiced automatically. (Product intro)
• MCP vs Skills: Emerging division of labor — MCP as the protocol for calling external APIs, Skills/AGENTS.md as playbooks for composing them into deliverables. Weaviate shared Skills examples for vector search workflows. (Example)

Infrastructure & Hardware

• Transformer training on Apple Neural Engine: A researcher bypassed CoreML, using undocumented APIs to run a 110M-parameter GPT training loop on an M4 ANE (most ops on ANE, some gradients on CPU). Claimed ~6.6 TFLOPS/W efficiency, sparking discussion of Mac mini training clusters. (Details)
• Google Static: Sparse matrix framework accelerating constrained decoding up to 948x for LLM-based generative retrieval — ideal for RAG where outputs must be valid KB entries. (Blog)
• NVIDIA Blackwell split: Datacenter B100/B200 are Compute Capability 10.x with tcgen05/DPX; consumer RTX 50-series is CC 12.0 without those AI units — deep optimizations will be datacenter-only. (NVIDIA blog)
• Taalas HC1 ASIC: Weights baked into the mask, no HBM pipeline — up to ~16–17k tok/s per user, but one chip serves exactly one model. A cloud-scale deployment play, not a 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 on real hardware and writes CUDA kernels, claimed ~2x faster than torch.compile on standard kernels and far ahead of Claude/Gemini on KernelBench. Costs huge GPU training; kernels not yet open-sourced. (Paper)
• Databricks OAPL: Off-policy RL that reportedly trains reasoning models with 3x fewer generation runs than on-policy methods like GRPO. (Intro thread)
• Wide-and-deep Transformer theory (Meta): 60+ pages on signal propagation, width/depth scaling, NTK analysis, SGD vs AdamW, validated on vision and language Transformers — worth reading before hyperparameter guessing. (arXiv)
• Sakana AI text-to-LoRA open-sourced: Generates LoRAs from natural-language descriptions; reproducible on a single H100 in 5 days. (GitHub)
• SAE analysis of text-to-image diffusion: Sparse autoencoder analysis shows early-step activations already predict final image composition — models "decide what to draw" early, then refine texture. (Paper)

Products & Applications

• Local LLM progress: 13 months after the DeepSeek moment, a $600 mini PC runs Qwen3-27B Q4 quantization, considered competitive with the original DeepSeek R1; skepticism grows toward averaged "intelligence index" leaderboards. (Reddit)
• DishBrain: Cortical Labs connected ~800k human+mouse neurons to a chip; it plays Pong and Doom, blurring lines on what counts as general computation. (Video)
• Runway Gen-4.5 enters arena.ai's text-to-video leaderboard at 1218, roughly on par with Kling-2.6-Pro. (Leaderboard)
• MLB pitcher Robert Stock built an 8.9-million-pitch analysis platform with AI tools despite no programming background — "non-programmers building serious apps with AI" is now real. (Post)
• OpenClaw showcase: Auto-booking haircuts, multi-agent dashboards with an orchestrator backend, multi-persona debates with local file access — becoming an engineer-focused Cowork alternative.

Industry & Business

• Qwen 3.5 fires up the local LLM supply chain: Hugging Face, Ollama, LM Studio, WebGPU, GGUF quants all shipped day one; the community consensus is faster and more accurate than prior 7B/13B models at equal hardware, with built-in multimodality and tool calling. (HF collection)
• Infrastructure is the 2026 keyword: Databricks RL efficiency, Google Static, NVIDIA's Blackwell segmentation, plus TPU/ASIC moves (Meta buying Google TPUs, Taalas HC1) — the consensus: the coming years are about compute + infrastructure, not leaderboard scores.
• AI economy vs dot-com bubble analogy: Similar burn rates and mood swings, but unlike 2000, real production use cases (code, support, content) are already deployed locally and in the cloud.

Policy, Governance & Safety

• US DoD drops Anthropic, signs with OpenAI: Anthropic labeled a "supply chain risk"; contractors must stop using its models within 6 months, officially for lack of "sufficient access." OpenAI announced a classified-environment deployment; Sam Altman's amendment bans intentional surveillance of US persons, but lawyers note the incidental-collection loophole persists. (OpenAI blog | Altman's terms)
• User backlash: TechCrunch-reported ChatGPT mobile uninstalls rose 295% within 48 hours; Claude hit #1 on the US App Store. (Reddit)
• Moonshot/Kimi distillation controversy: Moonshot allegedly extracted Claude capabilities via industrial-scale distillation; distilled Claude sometimes identified itself in Chinese as DeepSeek. Anthropic calling 150k API calls an "attack" drew pushback — where is the acceptable distillation line? (Case analysis)
• Responsible jailbreak disclosure: Red-team communities chasing Claude 4.6/Gemini/GPT-5.x bypasses are also drafting SOPs — reproduction, encrypted reporting, fix windows, coordinated disclosure, written scope, export-control awareness. (Playbook)

---

📌 Source: Easy AI Daily