Easy AI Daily Digest | March 4, 2026

A roundup of AI industry news compiled by zhichai.net.

Models & Capabilities

Google launches Gemini 3.1 Flash-Lite: faster, pricier high-throughput endpoint

• Preview of Gemini 3.1 Flash-Lite positions it as the lowest-latency, highest-throughput multimodal model in the Gemini 3 family: 1M context, ~360+ tokens/s, ~5.1s average response.
• Jeff Dean quoted ~$0.25/M input, $1.5/M output tokens; LMArena Elo 1432 — but pricing is ~2.5–3.75x higher than 2.5 Flash-Lite, sparking debate over value.
• Links: DeepMind thread | Jeff Dean pricing | Arena leaderboard | Cost analysis discussion

OpenAI ships GPT-5.3 Instant, teases GPT-5.4

• Rolled out to all ChatGPT users; claimed more natural, less preachy, fewer unnecessary refusals. Hallucination rates down 26.8% with search, 19.7% without.
• gpt-5.3-chat-latest appeared in the API and Arena comparisons. Official teaser says GPT-5.4 is coming "sooner than you think"; some read this as distraction from the DoD/NSA contract controversy.
• Links: Announcement | GPT-5.4 teaser

Alibaba Qwen 3.5: small models, local deployment, reasoning

• Reddit buzz: the 0.8B model includes a vision encoder and runs in-browser via WebGPU and on phones (~12 tokens/s on old hardware). 27B/35B versions perform close to larger models on reasoning, research, and long-context tasks; linear attention improves efficiency. Community still warns of notable hallucination risk.
• Links: Qwen 3.5 vs 3 benchmarks | 0.8B WebGPU demo

Apple M5 Pro / M5 Max: up to 4x faster LLM prompt processing

• M5 Pro: 64GB unified memory, 307GB/s bandwidth; M5 Max: 128GB, 614GB/s. Apple claims up to 4x LLM prompt processing vs M4 Pro/Max. SSD throughput up to 14.5GB/s; integrated N1 chip adds Wi-Fi 7. Community awaits Mac Studio / local LLM results.
• Link: LocalLLaMA discussion

Infrastructure & Hardware

Long-context training and memory optimization

• Together: Context Parallel + sequence parallelism enables training an 8B model at 5M context on a single 8×H100 node, cutting attention memory up to 87%.
• Databricks: open-sourced FlashOptim, reducing AdamW state from ~16 to 7 bytes/param; 8B fine-tuning peak memory drops from 175GiB to 113GiB.
• SkyPilot: Job Groups split RL training across high-end GPUs, cheap GPUs, and large-memory CPUs.
• Links: Together | FlashOptim | SkyPilot

NVIDIA Blackwell splits into datacenter vs consumer lines

• Blackwell is split into datacenter CC 10.0 and consumer CC 12.0; some features exist only on sm_100a / sm_100f, creating forward-compatibility issues for CUDA/kernel developers.
• Link: NVIDIA blog

ByteDance CUDA Agent and RL kernel generation

• CUDA Agent auto-generates high-performance CUDA kernels, reportedly ~2x faster than torch.compile on small/medium kernels and well ahead of mainstream LLMs on complex benchmarks. An academic RL-CUDA agent shows similar performance but requires large GPU pools and process isolation.
• Links: CUDA Agent site | RL Agent paper

Specialized inference chips and Apple ANE

• Taalas HC1 hardwires Llama-3.1-8B for ~17,000 tokens/s but runs only that fixed model. On M4's Neural Engine, Llama2 110M reportedly achieves ~80x better energy efficiency than an A100. Analysis projects a $255B inference infrastructure market by 2030.
• Links: Taalas paper | M4 ANE benchmarks

Agents & Tooling

"MCP is dead"? The ecosystem keeps expanding

• Notion integrated MCP/API (Meeting Notes usable in Claude Code in one command); Cursor launched MCP Apps rendering interactive UIs in agent chats. Meanwhile, security researchers flagged a messy MCP security model with 5 easily exploited attack patterns.
• Links: Notion MCP | Cursor MCP Apps | MCP security analysis

ShadowClaw: a minimal local agent in a single C executable

• Single-file C agent calling local LLMs (e.g., Ollama) via curl; built-in shell execution, file I/O, HTTP GET, simple math, and automatic state persistence. A controllable baseline for personal local agents.
• Link: ShadowClaw repo

RLM (Recursive Language Modeling): agent paradigm debate

• DSPy circles are debating giving models a REPL to write and run their own code instead of pre-written tools. Supporters see more flexibility than ReAct; critics worry about controllability, debugging, and safety.

Perplexity Computer and Cursor cloud agents

• Both run agents in isolated VMs/sandboxes that operate browsers, terminals, and IDEs, producing PRs or documents — no user-managed API keys, but less debugging transparency and higher cost.

Research & Methods

• Agent benchmarks vs real work: new work argues current agent benchmarks over-index on math/coding, mismatching real labor distribution. Arena launched Document Arena (PDF-based reasoning); Claude Opus 4.6 leads. Document Arena
• Multi-agent consensus & Theory of Mind: Byzantine consensus experiments show LLM multi-agent systems struggle to reach agreement even without malicious nodes (timeouts, deadlock, worse with more agents). ToM/BDI + formal verification helps only depending on the base model.
• Spectral norm scaling and muP: EleutherAI discussion of a 2023 paper explaining when networks move from kernel-method behavior to true feature learning via √(fan-out/fan-in) scaling — an intuitive account of maximal update parametrization. Paper | Modula
• SAEs on text-to-image diffusion: sparse autoencoders reveal image composition is largely predictable early in reverse diffusion, style mid-way, texture last; targeted interventions demonstrated. Paper

Products & Applications

• Claude & Claude Code surge: Anthropic reports traffic far above expectations and is scaling capacity; seen as beneficiary of OpenAI's DoD controversy, with reported rise to #1 US enterprise market share. Claude Code is gradually rolling out a voice mode (hold space to speak, no extra charge). Discussion
• Cursor and "AI coworkers": Cursor refreshed its IDE (Zen mode, cloud agents incl. Android WebApp); "Viktor," an AI coworker in Slack, handles marketing audits, ad management, and lead research with 3,000+ SaaS integrations and persistent memory.
• Local LLM tooling: LM Studio fixed LM Link device discovery, reopened its waitlist, and discusses Vulkan multi-GPU balancing. OpenClaw opened community rules/architecture; Manus.im shares credit-usage optimization tips.
• easytranscriber: KBLab released ASR with accurate timestamps, a WhisperX-like interface, 35%–102% faster depending on hardware, with HF models as backends. Blog

Industry & Companies

Alibaba Qwen team exodus

• Tech lead Justin Lin and other key members departed; community worries about future open-source/licensing strategy, even as Qwen 3.5 training scripts, LoRA guides, and Int4 GPTQ releases continue. Justin Lin farewell

OpenAI post-training VP jumps to Anthropic

• Max Schwarzer, VP overseeing RLHF/post-training (worked on GPT-5 series), left for Anthropic to do hands-on RL research — a major talent win for Anthropic and fuel for OpenAI stability speculation.

Anthropic vs Pentagon/Palantir

• Reports say the DoD threatened to label Anthropic a "supply chain risk," affecting Palantir's use of its models; Anthropic wants guardrails against mass surveillance and autonomous weapons uses.

Policy, Governance & Safety

• OpenAI-DoD/NSA contract trust crisis: scholars demanded contract transparency; Sam Altman said terms were amended to prohibit domestic surveillance of US citizens, but many call for independent legal review. Altman statement
• "295% surge" in ChatGPT uninstalls: media cited the figure post-DoD news, but skeptics note a small base could inflate the percentage; revenue impact depends on long-term retention. Claude downloads reportedly rose alongside.
• MCP security "is a mess": a detailed write-up lists 5 attack patterns (prompt injection via tool descriptions, file/network overreach, third-party abuse), arguing MCP exposes powerful APIs to an easily fooled model without adequate sandboxing.
• BASI community: the Jailbreaking Discord shares both legitimate red-team frameworks (MITRE ATLAS) and paid jailbreak prompts alongside harmful synthesis content — useful but gray-zone.

📌 Source: Easy AI Daily