Easy AI Daily Digest | March 14, 2026

📅 AI Industry Digest — March 14, 2026

A curated daily roundup of AI industry news, originally published on zhichai.net's Easy AI Daily.

Key points

Models & Capabilities

• Anthropic ships 1M-context Opus 4.6 as default. 1M context is now the default for Max/Team/Enterprise, long-context surcharges removed, no beta header needed, and image/PDF limits raised to ~600 pages per request. Scores 78.3% on MRCR v2 1M tokens. Source: Latent Space
• OmniCoder-9B released. Tesslate's open-source (Apache 2.0) coding agent model fine-tuned from Qwen3.5-9B on 425k+ agentic coding trajectories (data generated with Claude Opus 4.6, GPT-5.4). Native 262k context, extendable to 1M+.
• Qwen3.5-9B praised as a strong small model. Community users report it runs on a 12GB RTX 3060 and approaches larger models like gpt120b on agentic coding tasks.
• Qwen 3.5 fine-tunes called 'noticeably stronger.' A community post highlights 33 Qwen 3.5-based fine-tunes, including a 40B dense variant and Claude Opus-style models.

Agents & Tooling

• MCP debate continues. Engineer consensus: demand exists, but onboarding cost and ergonomics are the problem. LlamaIndex: MCP suits stable-API/real-time-data needs; local skills are lighter but more fragile. Sources: LlamaIndex, Pamela Fox
• Chrome v146 adds Web MCP support. Demos show a LangChain Deep Agent continuously browsing X and auto-generating daily digests — a step toward 'browser as agent host.'
• Hermes Agent discussed as a self-hosted agent with long-term memory and self-improvement over time.
• Multi-agent coding pipelines emerging. Engineers share setups with 5 agents handling review, testing, security, and performance, plus 2 agents merging PRs and running regression — closer to an automated CI factory than autocomplete.
• Automated research heats up. Karpathy's autoresearch sparks debate (veterans note lineage from DSPy, GEPA, Bayesian optimization); Together AI open-sources Open Deep Research v2 app, evals, and code.

Infrastructure & Hardware

• 'Context drought.' Latent Space argues context windows have grown less than an order of magnitude since 2024; the bottleneck is HBM/DRAM supply, with predictions of future 'context rationing.' Source: Context Drought, podcast
• IndexCache reuses sparse-attention indices across layers in DeepSeek Sparse Attention: ~1.2x end-to-end speedup on GLM-5 744B; 1.82x prefill / 1.48x decode on 30B-class models at 200k context, with ~75% fewer index computations.
• Klein KV (Black Forest Labs) injects reference-image KV caches into DiT denoising steps, speeding multi-reference image editing up to ~2.5x.
• Vera Rubin era begins. Satya Nadella says Azure is first to validate NVIDIA Vera Rubin NVL72; Lambda advocates bare-metal (non-virtualized) deployments for Rubin clusters.
• tinygrad's 'exabox' vision: a Python-driven box exposing as one giant GPU by 2027, hiding all distributed details.

Research & Methods

• RandOpt / Neural Thickets (MIT et al.): adding Gaussian noise to pretrained weights plus ensembling can approach or exceed GRPO/PPO across reasoning, coding, writing, chemistry, and VLM tasks.
• Stanford on general data replay: fine-tuning gains ~1.87x, mid-training ~2.06x; downstream: +4.5 points on web-navigation agents, ~2% on Basque QA.
• Multi-agent memory as a computer-architecture problem: a paper frames shared agent memory as cache/memory hierarchies with consistency and permission issues.
• BrokenArXiv benchmark: GPT-5.4 rejects only ~40% of subtly falsified math claims from recent papers.
• OpenFold3 Preview 2: claims to nearly close the gap with AlphaFold3 across modalities; uniquely releases weights, training data, and configs for full from-scratch reproduction. Source: Mo AlQuraishi
• WAXAL dataset: 2,400+ hours of speech covering 27 Sub-Saharan languages — TTS for 17, ASR for 19, serving ~100M+ speakers.

Products & Applications

• Personal agent UX goes always-on, cross-device: Perplexity Computer launches on iOS (monitor/take over desktop browser tasks from phone); Claude Code demos phone-initiated desktop coding sessions; Genspark's Claw runs as a persistent cloud 'AI employee.'
• Gemini task automation: The Verge demos include auto-hailing an Uber and selecting dishes from a menu — acting, not just advising. Source: The Verge
• Gemini UI/UX 2.0: personalized, interactive redesign with heavy promotion of the $250/month Google AI Ultra tier, drawing pricing criticism.
• Nano Banana Pro quality complaints: users report pixelation and blurring after March 10, suspecting model or safety-policy changes.
• Claude's interactive chart UI — manipulating charts and data inside chat — draws wide attention as a promising analytics-in-chat direction.

Industry & Business

• xAI restarts hiring: Elon Musk says xAI is re-reviewing past interviews and re-contacting strong candidates previously rejected.
• Altman: intelligence as a utility. Sam Altman likens OpenAI's future business to metered electricity/water — selling intelligence by usage.
• John Carmack on training data: open-source code is a gift; AI training amplifies its value rather than stealing it.
• AINews joins Latent Space as a column with searchable archives; the Discord channel won't reopen in its original form.

Policy, Governance & Safety

• Palantir CEO Alex Karp claimed AI will reduce the political influence of highly educated, Democratic-leaning voters while empowering technically skilled working-class men — drawing criticism for entangling AI in US political polarization.
• Bernie Sanders proposes a bill to ban construction of all new AI data centers, citing AI as an existential threat; the blanket approach sparked strong controversy in mainstream policy circles.

📌 Source: Easy AI Daily (zhichai.net)