Easy AI Daily Digest | March 18, 2026: GPT-5.4 mini, Mistral Small 4, Nemotron 3 Ultra & More

Key points

Products & Applications

• Anthropic Claude Cowork remote control: Anthropic added true remote-control capabilities to Claude Cowork, letting it operate directly on your computer rather than just issuing instructions. Commentators compare it to OpenClaw as Anthropic's formal entry into computer-control agents for office and development workflows.
• Perplexity Comet Enterprise: An enterprise AI browser for team-level search/Q&A with admin-controlled rollout, auditing, and CrowdStrike Falcon integration for security-compliant deployment.
• Hugging Face hf CLI coding agent plugin: Automatically selects the best local model and quantization for your hardware, launching a local coding assistant in one command.
• Ollama OpenClaw workflow support: New web search/scrape plugins and headless mode; also appears in CodexBar as a unified model entry point.
• LTX 2.3 game-style LoRA: Trained on 440 clips from the game *Dispatch*, supporting 6+ characters and styles via trigger words—suited to game/film previsualization.
• OldNokia UltraReal LoRA: Replicates 2000s phone camera aesthetics (soft focus, washed-out colors, JPEG artifacts, noise) from a Nokia E61i photo dataset.

Models & Capabilities

• OpenAI GPT-5.4 mini / nano: Available across API, ChatGPT, and Codex. mini is 2x faster than GPT-5 mini with 400K context, near-frontier scores on SWE-Bench Pro and OSWorld, but consumes 30% of the 5.4 Codex quota. Higher prices; weaker on sycophancy/deference tests.
• Mistral Small 4 (119B MoE): 128 experts, 6.5B active per token, 256k context, multimodal, Apache 2.0. Community questions focus on tool-calling stability vs. Devstral 2 and long-context usability.
• Qwen3.5-9B: Scores 77 on document AI benchmarks (rank 9), slightly below GPT-5.4's 81; strong at key information extraction, table understanding, and OmniOCR. Also noted as energy-efficient for long reasoning tasks.
• NVIDIA Nemotron 3 Ultra Base (~500B): Claims leading open-base results on MMLU Pro, HumanEval, GSM8K and 5x throughput efficiency—though the community criticized unclear baselines and charts starting at 60%.
• Holotron-12B: H Company + NVIDIA's open multimodal model for computer-use agents (screen reading, clicking, form filling).

Agents & Tooling

• LangChain: Launched LangSmith Sandboxes for safe one-off code execution and open-sourced Open SWE, an engineering agent system with Slack/Linear/GitHub integration, sub-agents, middleware, and validation.
• Converging agent stacks: OpenAI Codex adds sub-agents (positioning GPT-5.4 mini as the sub-agent pick); Hermes Agent v0.3.0 ships plugins, Chrome control, IDE plugins, local voice, and PII redaction; LangChain's Deep Agents offers a MIT-licensed, inspectable Claude Code-style harness.
• Unsloth Studio: Fully open-source Web UI for local training + inference across 500+ models, claiming 2x training speed and 70% less VRAM; seen as an open LM Studio alternative for advanced users.
• Cursor RL context compression: A reinforcement-learned self-summarization policy for Composer reportedly cuts compression error by ~50%.

Infrastructure & Hardware

• NVIDIA GTC: Jensen Huang framed future computers as "token factories," emphasizing inference and agents. LangChain announced 1B+ framework downloads and joined the Nemotron alliance; llama.cpp added Nemotron 3 Nano 4B support.
• DGX Station: Now shipping via OEMs at roughly $85–90K, a local AI supercomputing node with unified memory and no default video output.

Research

• Moonshot Attention Residuals: "Vertical attention" lets each layer query states from previous layers—inter-layer memory at near-zero added latency, similar work at ByteDance.
• Mamba-3: From Albert Gu and Tri Dao, a stronger MIMO variant claiming fastest prefill+decode at 1.5B scale—positioned for long-trajectory RL and heavy inference workloads.

Industry & Policy

• $1T is "just the first half": Huang argues the often-cited $1T AI infrastructure opportunity covers only part of the stack through 2027.
• Anthropic CEO prediction: Half of entry-level white-collar jobs could be automated within 3 years; commenters worried about accountability for AI errors.
• NBC poll: Only 26% of US voters view AI positively, 46% negatively.