Good morning. It’s Wednesday, March 18th.

Here’s something I’ve been testing that’s pretty useful:

You can export your entire Gmail inbox using Google Takeout—it comes down as a .mbox file. From there, have Claude write a quick script to extract just the useful parts (sender, metadata, message text) and strip out things like quoted replies. That compresses a year or two of emails into a small, clean text file.

Once you’ve got that, drop it back into Claude and start asking questions (missed business opportunities, recurring contacts, patterns in conversations, etc.) It’s surprisingly good at surfacing things you wouldn’t notice.

Takes a couple extra steps upfront, but it’s 100x more powerful than just using Gemini inside Gmail.

-Jeff
AI Breakfast

You read. We listen. Let us know what you think by replying to this email.

Anthropic has launched their own version of OpenClaw - Dispatch. A research preview for its Claude Cowork platform that enables remote control of Mac-based desktop sessions via mobile devices. The feature uses a QR-pairing system to bridge the desktop app with the iPhone, mirroring the Remote Control functionality recently introduced for the terminal-based Claude Code.

Dispatch leverages Claude's agentic architecture and Connectors to summarize emails, fetch Notion database entries, and locate local files using OCR, though it currently operates within a sandboxed virtual machine for security. Read more.

Mistral AI is shifting to a sovereign infrastructure strategy with the launch of Mistral Forge, a platform for full-cycle model development including proprietary pre-training and reinforcement learning. Unveiled at NVIDIA GTC 2026, Forge provides enterprises with internal training recipes and forward-deployed scientists to bypass third-party provider lock-in for high-compliance sectors.

The ecosystem centers on Mistral Small 4, a 119B-parameter MoE model that maintains an 8B active footprint, delivering 40 percent faster completion via a new reasoning_effort toggle. This Apache 2.0-licensed model supports a 256k context window and integrates with Leanstral for formal code verification. Combined with its lead in the NVIDIA Nemotron Coalition, Mistral expects to surpass $1 billion in ARR by giving organizations total control over data and model behavior. Read more.

NVIDIA has become more aggressive in its move to own the inference era. At GTC 2026, Jensen Huang laid out a roadmap to $1.25 trillion in revenue, anchored by the Vera Rubin platform’s 60-exaflop compute power and the Groq 3 LPX dedicated inference pipeline. Technical breakthroughs like KV Cache Transform Coding (KVTC) now shrink LLM memory overhead by 20x, while DLSS 5 leverages generative AI to reconstruct photorealistic lighting in real time.

The strategy extends deep into physical AI, using the Cosmos world model and a new Data Factory Blueprint to replace scarce real-world data with synthetic simulation for GR00T N2 humanoids and a 2027 Uber robotaxi launch. NVIDIA also launched NemoClaw and OpenShell for agent sandboxing, supported by the Nemotron Coalition's partnership with Mistral AI. The rollout is further bolstered by native CloudXR 6.0 on Apple Vision Pro, DLSS 5 generative lighting, and the strategic restart of H200 sales in China. Read more.

OpenAI is consolidating its focus on enterprise infrastructure and coding, killing off "side quests" like the standalone Sora app to preserve compute for its Frontier agent platform. The company reorganized its Stargate initiative into three specialized groups under new leadership, pivoting from self-built data centers to a $600 billion rental strategy with AWS and Google Cloud. This shift coincides with a 20 percent surge in API usage driven by Codex subagents and a $10 billion joint venture with private equity firms.

While the new GPT-5.4 mini and nano models offer near-flagship performance, they arrive at a significantly higher price point, costing up to 4x more than their predecessors. GPT-5.4 Pro pushes research boundaries by proposing solutions to unsolved problems in the HorizonMath benchmark. Read more.

MuleRun is an always-on AI workforce that automates end-to-end workflows 24/7. Unlike chatbots, it uses dedicated computers to complete tasks, monitor data, and act proactively while you sleep.

Lightning Rod is an SDK that converts real-world data into production-ready training datasets in hours. Use Python to bypass manual labeling and turn news or documents into verified data.

OpenViktor is a free, open-source AI employee that integrates into your tools to automate roles 24/7. It masters your team's context from day one, proactively executing tasks without prompts.

Perceptis is an AI consulting OS that automates business development workflows end-to-end from identifying leads to generating proposals. It ingests your firm’s data, past work, and methodology to produce client-ready deliverables that match your exact style.

JetBrains Air is an agent-driven development platform that runs Codex, Claude, Gemini, and Junie side-by-side. Isolate tasks, review results, and manage multiple coding agents within a single, coherent workflow

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.