AI Signals — March 12
Two things happened this week that, taken together, define the current moment in AI better than any single announcement could.
First: Anthropic opened the “Agent Skills” standard — making agent capabilities portable across platforms, not locked to Claude. On the same day, Microsoft announced Copilot Cowork, a desktop-native agentic knowledge work tool built in partnership with Anthropic. The message is clear. Anthropic is not building a product company. It’s building infrastructure for the agentic era, and it wants enterprise distribution yesterday.
Second: Apple’s new Siri is launching this month, powered by Google’s 1.2 trillion parameter Gemini model running through Private Cloud Compute. $1-5 billion per year deal. 2+ billion devices. The most significant distribution event in personal AI history, and it’s arriving quietly — as an OS update.
These two moves define the poles. On one end: enterprise infrastructure, orchestrated agents, Microsoft distribution. On the other: consumer ambient AI, baked into the OS, in everyone’s pocket by default.
The space in between — deep, personal, proactive AI that actually knows you — is still wide open.
The model pace is not slowing down. March 2026 has already seen more than a dozen significant releases. Alibaba’s Qwen 3.5 2B model runs offline on any recent iPhone with 4GB of RAM. DeepSeek V4 is incoming with 1 trillion parameters and only 32B active per token — more efficient than V3, built to run on Chinese chips. Microsoft’s Phi-4-reasoning-vision-15B is open-weight, Apache 2.0, and punches well above its weight class. Ai2’s OLMo Hybrid achieves the same accuracy with 49% fewer training tokens. The efficiency curve is steepening.
The practical signal: on-device inference is no longer a research demo. It’s a deployment target. Privacy-first architecture is about to become a competitive differentiator, not a positioning claim.
On the enterprise side, the language has settled. The problem is called “agent sprawl.” The solution is called “orchestration layer.” Gartner reports 61% of enterprises have started agentic AI development. Deloitte projects 50% of enterprises using gen AI will deploy autonomous agents by 2027. The companies winning are redesigning operations from the ground up — not layering agents onto broken workflows.
What’s quietly getting confirmed is the core premise behind proactive AI: 60% of a knowledge worker’s time goes to “work about work” — status updates, searching for context, coordinating between tools. AI isn’t solving that yet. The tools are smarter, but the overhead remains. The next real win is reducing coordination cost, not just answering questions faster.
In the personal AI / second brain space, the feature tier is consolidating fast. Notion shipped agents in v3.1 — not Q&A, but actions across the workspace, pulling from Slack and Google Drive. Saner.AI is explicitly positioning against manual organization overhead. Obsidian added Siri integration. AFFiNE bridges whiteboard and structured doc natively.
Every serious tool is converging on the same three things: multimodal capture, AI that acts rather than just answers, and local-first privacy. The ones doing all three are pulling ahead. None of them, however, are operating above the tools. They are still tools the user operates. The coordination layer — the thing that understands priorities, tracks open loops, and acts across context without being asked — doesn’t exist yet in this space.
That’s the wedge. It isn’t closing.
The week’s clearest lesson: the platforms are moving fast to own ambient AI. Apple + Google own the device layer. Microsoft + Anthropic own the enterprise layer. What neither owns is the personal operating layer — the thing that knows what you’re trying to do, not just what you just said.
That’s the bet worth making.
