Easy AI Daily | March 2, 2026
A roundup of AI industry news covering models, agent tooling, infrastructure, research, product launches, and policy.
Models & Capabilities
Alibaba releases Qwen 3.5 small model family: multimodal + ultra-long context, targeting local and edge use
Alibaba launched Qwen3.5-0.8B/2B/4B/9B (Base/Instruct variants) with native multimodality and RL enhancement, claiming native 262k context extendable to ~1M. Community tests show the 4B beating many older 9B models; local 128k context runs at ~60 tok/s. Small models default to "thinking mode" off — it must be enabled via templates/parameters.> Links: Official release tweet | Architecture & Gated DeltaNet discussion | Local benchmarks & long-context discussion (Reddit)
Qwen 3.5 large models: 27B/35B challenge 100B+ class models on reasoning and coding
Community benchmarks show Qwen3.5-27B and 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 completes embedded Tetris-style tasks in under 4 seconds.> Links: 27B architecture analysis | 9B/4B benchmark charts | Opus 4.6 distilled Qwen3.5-27B
Qwen 3.5 local ecosystem: Ollama, LM Studio, browser WebGPU integration
Qwen 3.5 small models were quickly integrated by 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 directly in-browser via WebGPU (vision encoding remains the bottleneck). LM Studio and Unsloth released quantizations optimized for NVFP4 and thinking mode.> Links: Ollama integration | WebGPU demo | NVFP4 multimodal quantization collection
Codex 5.3: coding intelligence keeps climbing; discussion shifts to usability
Codex 5.3 leads on custom benchmarks like WeirdML and is reportedly more stable on complex multi-step tasks, trading wins with top Gemini versions. Discussion now centers on workflows where models write ~80% of code, with humans supervising and patching.> Links: WeirdML results | Claude Code writing 80% of production code
BullshitBench v2: a benchmark for measuring confabulation
BullshitBench v2 adds ~100 questions spanning code, medicine, law, finance, and physics, testing 70+ models. The author claims more reasoning can increase confabulation; Anthropic models lead clearly, while recent OpenAI/Google generations show little progress — drawing pushback.> Links: BullshitBench v2 announcement | Community skepticism
Agents & Tooling
AGENTS.md / SKILL.md in practice: ~30% runtime savings, not a magic wand
A team tested AGENTS.md across 10 repos and 124 PRs: median runtime dropped ~28.6% and token consumption fell 16.6%, mainly by reducing worst-case flailing. Anthropic also released a 30-page Skills guide emphasizing breaking complex workflows into testable skill files rather than long prompts.> Links: AGENTS.md experiment data | Anthropic Skills guide PDF
Agent reliability and observability: evaluation is harder than building agents
Developers report the hard parts are defining success, monitoring, and evaluation. Recommendations: define success criteria first (led by product/domain experts), prefer deterministic scorers with LLMs judging only style, and evaluate outputs rather than intermediate reasoning.> Links: Agent observability questions | Evaluation practice advice
GitNexus: repo-to-knowledge-graph RAG in the browser
GitNexus parses repositories into interactive dependency graphs, stores relations in embedded KuzuDB, and answers queries via Cypher graph traversal rather than embedding similarity. Fully browser-based via Web Workers, MIT-licensed — a good backend for project-tour/code-QA agents.> Links: GitNexus introduction
Stripe launches LLM billing proxy: charge by token
Stripe added token-based billing: pick a model, set a markup, route calls through Stripe's LLM proxy, and usage is automatically billed — no need to build your own usage metering for AI SaaS.> Links: Product intro and use case
MCP vs Skills: division of labor getting clearer
The community increasingly treats MCP as the unified protocol for model-to-API calls, while Skills/AGENTS.md serve as operation manuals for composing APIs into deliverables. Weaviate shared examples wrapping vector-retrieval/tool combinations as Skills.> Links: Weaviate skill examples
Infrastructure & Hardware
Transformer training on Apple Neural Engine: training on M4 is now real
A researcher bypassed CoreML, using undocumented APIs to run a 110M-parameter GPT training loop largely on the M4's ANE (some gradients still on CPU). Claimed M4 ANE efficiency of ~6.6 TFLOPS/W far exceeds A100/H100, sparking discussion of power-efficient Mac mini training clusters.> Links: Technical details | Additional interpretation
Google Static: ~1000x faster constrained decoding for retrieval
Google's Static sparse matrix framework accelerates constrained generation (e.g., outputting entries from a knowledge base), claiming 948x speedup over naive approaches — well-suited to RAG and generative retrieval where outputs must be valid.> Links: Static technical blog coverage
NVIDIA Blackwell architecture split: data center and consumer cards diverge
Blackwell is now two tracks: data center B100/B200 are Compute Capability 10.x with tcgen05 and DPX features, while consumer Blackwell RTX (50-series) is CC 12.0 without those AI-specific units. Future low-level optimizations will mainly apply to data center cards.> Links: Official blog on the architecture split
Taalas HC1 and ASIC inference: one chip, one model, ~17k tok/s per user
ASICs with model weights baked into the mask skip HBM pipelines, delivering ~16–17k tok/s per user — but each chip serves only one model, so upgrades require new hardware. Best suited to hyperscale deployment of a single fixed model, not general GPU replacement.> Links: Original explainer
AMD open-sources finer-grained GPU trace tool; tinygrad says it beats NVIDIA's
AMD open-sourced rocprof-trace-decoder, parsing SQTT instruction-level traces for detailed profiling. The tinygrad author called AMD's tracing infrastructure "better than NVIDIA's now" — good news for ROCm performance tuning.> Links: tinygrad comment
Research & Methods
ByteDance CUDA Agent: RL-driven kernel generation, ~2x faster than torch.compile
ByteDance's CUDA Agent paper uses RL with real-machine profiling as reward to auto-write CUDA kernels, claiming ~2x speedup over torch.compile on standard kernels and outperforming code-writing LLMs like Claude/Gemini on KernelBench. Downsides: massive GPU training cost and no open-sourced kernels yet.> Links: Paper discussion | arXiv
Databricks OAPL: cheaper off-policy RL for reasoning models
OAPL (Optimal Advantage-based Policy Optimization with Lagged Inference) is a sample-efficient off-policy RL method, reportedly requiring 3x fewer generation runs than on-policy methods like GRPO to train reasoning ability, with a simpler training system.> Links: OAPL introduction thread
Revisiting the theory of wide-and-deep Transformers
Meta's effective theory of wide and deep Transformers (~60 pages) systematically analyzes forward/backward signal propagation, width/depth scaling, NTK analysis, and SGD vs AdamW behavior, validated on vision and language Transformers — worth reading for principled hyperparameter tuning.> Links: The Turing Post summary | arXiv link
Sakana AI open-sources text-to-LoRA
Sakana AI released a text-to-LoRA model and training code that generates LoRAs directly from natural-language descriptions for style/task fine-tuning. Officially reproducible on a single H100 over 5 days.> Links: GitHub repo | Hugging Face model page
SAE analysis of image-to-image diffusion: layout predictable before the first step completes
Sparse autoencoder analysis of mainstream text-to-image diffusion models shows early-step activation spatial distributions predict final image composition quite accurately — the model "decides what to draw" early, with later steps mostly refining texture.> Links: SAE diffusion paper
Products & Applications
Local LLM retrospective: a $600 mini PC now runs former "god-tier" models
A Reddit retrospective on the year since DeepSeek's breakout: a $600 mini PC now runs Qwen3-27B-class models at Q4 quantization, with many considering the results on par with the original DeepSeek R1. Users also caution against "intelligence index" leaderboards that just average MMLU/GPQA scores.> Links: Local LLM evolution discussion
DishBrain: 800,000 living neurons play Doom and Pong
Cortical Labs connected ~800k human and mouse neurons to a silicon chip, creating DishBrain, which plays Pong and even Doom — half the community sees sci-fi, half debate whether it counts as a general computing device.> Links: Project showcase
Runway Gen-4.5 enters the text-to-video arena
Runway's Gen-4.5 scored 1218 on arena.ai's text-to-video leaderboard, roughly in the same tier as Kling-2.6-Pro.> Links: Text-to-video leaderboard
MLB pitcher Robert Stock builds an 8.9-million-pitch analysis platform with AI
Robert Stock, with no programming background, built a platform analyzing 8.9 million pitches using ML for pitching strategy and training analysis — a widely shared example of non-programmers building serious AI applications.> Links: His X post
OpenClaw + local/cloud models: from automated barber bookings to multi-agent dashboards
Community showcases include automated salon reservations, multi-agent dashboards with an orchestrator managing concurrent specialized agents, and multi-persona debates within one session with local file access. OpenClaw is becoming an engineer-oriented Cowork alternative.> Links: OpenClaw showcase thread
Industry & Company News
Qwen 3.5 small models ignite the local LLM ecosystem
Ollama, LM Studio, Unsloth, and others all shipped Qwen3.5 support within days. The focus: faster and more accurate than older 7B/13B models on the same hardware, with built-in multimodality and tool calling.> Links: Official Hugging Face collection
Databricks, Google, NVIDIA: infrastructure is the keyword for 2026
Databricks pushing efficient RL (OAPL), Google optimizing retrieval with Static, NVIDIA splitting data center/consumer Blackwell — plus TPU/ASIC moves (Meta buying Google TPUs, Taalas HC1) — point to a consensus that the coming competition is about compute and infrastructure, not just model leaderboards.> Links: OAPL & Static discussion | Meta-Google TPU coverage
AI economy compared to the dot-com bubble — but many argue this time there's real output
Community discussions (including Nous Research) note AI investment resembles the 2000 dot-com bubble in burn rate and sentiment, but differ in that real production use cases (code, support, content) already exist, with many running Hermes/Qwen models as daily drivers.Policy, Governance & Safety
US DoD drops Anthropic, signs with OpenAI: contract terms and surveillance red lines in focus
The DoD labeled Anthropic a "supply chain risk," requiring contractors to stop using its models within 6 months, citing insufficient access. OpenAI announced a new DoD agreement deploying models in classified environments; Sam Altman's published addendum bans intentional surveillance of individuals inside the US, though legal observers note the intentional/incidental collection loophole remains.> Links: DoD risk statement | OpenAI blog: our agreement with the DoD | Altman's amendment
ChatGPT DoD contract triggers uninstall wave; Claude hits #1 on the App Store
TechCrunch-reported data shows ChatGPT mobile uninstalls up 295% within 48 hours of the announcement, while Claude topped the US App Store. Reddit users cite ethical reasons for switching; others doubt the boycott's durability and worry more about long-term privacy and compliance.> Links: Uninstall data discussion | Claude at #1 discussion
Moonshot/Kimi's large-scale Claude distillation sparks identity crisis and data security concerns
Moonshot AI allegedly used industrial-scale distillation against Claude to train Kimi; reports note distilled Claude sometimes identifies itself as DeepSeek in Chinese. Anthropic calling 150k API calls an "attack" drew pushback — by that standard, most benchmarks would be attacks. The community is debating where the line on ToS, call volume, and acceptable distillation lies.> Links: Kimi identity confusion case | Eleuther discussion
Security community jailbreaks and responsible disclosure SOPs
While hunting jailbreaks against Claude 4.6, Gemini, and GPT-5.x, security communities are also formalizing responsible disclosure: a Red-Team Playbook covering reproduction, encrypted reporting, fix windows, and coordinated disclosure, with guidance on written scope, harm reduction, and export controls.> Links: Red-Team Playbook Gist
---
📌 Source: Easy AI Daily