Perspectives
Writing on Physical AI, agentic infrastructure, AI governance, and venture strategy — originally published on LinkedIn.
The AI Read the Whole Document and Skipped the Exhibits
A 32-page case study went through my document pipeline. It filed a clean summary and skipped all 17 exhibits, the pages with the financial tables. What I changed, and what I learned testing vision models on numbers.
Disrupt Takeaways: Three Questions, Answered
Back from three days at TechCrunch Disrupt: how the panels answered my three questions on robots, networks and timing, and the one nobody answered.
The AI Race Has Two Finish Lines
Meg Rithmire argues the US is paying for the frontier while China pays for diffusion. Her sharper question is who owns the stack the world relies on next.
Three Questions I Will Ask Every Robotics Founder at Disrupt
Before Disrupt, the three questions I planned to ask every robotics founder, starting with what a robot does when the network drops.
The Robot's Brain Is Moving Onto the Machine. Who Owns the Timing?
On October 1, SEMIFIVE signed a deal with Mobilint to build an AI chip for robots that must keep working when the network goes down.
Intel Physical AI at MassRobotics: Who Owns the Timing, Intel's Slide
Notes from Intel's session at MassRobotics on open infrastructure for physical AI, and the question it left open: who owns the timing?
Meta Shipped the Execution Layer. Nine Days Later, Someone Shipped Its Missing Trust Layer.
Muse (by Meta) proved people will run an AI agent that keeps working after they close the app. That raises a governance question Meta didn't answer. Nine days later, TypeSafe AI shipped a model built to answer exactly that question, and it lines up with a discipline I'd already built into my own system the hard way.
AI Is Easy to Verify When Its Behavior Is Bounded. Robotics and the UN Just Hit the Same Wall Where It Isn't.
UNDP's chief digital officer explained how he measures AI programs, and his answer split along a line robotics safety engineers know well: bounded behavior can be verified, open-ended behavior cannot.
HUMAIN's Week
At LEAP 2026 in Riyadh this week, HUMAIN put five separate pieces of Saudi Arabia's AI strategy into public view at once.
Watch the Bond Market, Not the Stock Market
In April a data center bond backed by a Microsoft lease priced at 5.7%. In August the same issuer's next bond, same tenant, priced at 7.2%. Credit markets are pricing a risk that equity markets are not.
I've Walked the Floor at LEAP Before. This Year I'm Watching From Here, and I'm Paying Close Attention
I've attended LEAP before. A scheduling conflict is keeping me from this year's edition, but the story underneath it, five editions, $42 billion in cumulative investment, a conference that got rescheduled rather than skipped, is worth watching regardless of whether I'm in the room.
One Line in Sahil Bhaiwala's Data Thesis Points at the Next $100B Question
Handshake AI's Sahil Bhaiwala made the sharpest case I've read this year for the data-authoring economy, and scoped it carefully enough to set physical intelligence aside for a future post. That's exactly the layer I've spent this year tracking.
The Silence Was the Signal
Jensen Huang posted on X for the first time ever this week, sharing a letter signed by 25 companies arguing for open AI models. Three of the most important names in AI didn't sign it, and didn't fight it either. That silence tells you more than the letter does.
The Physics Problem Has a Geography
Navin Chaddha's latest post lays out why AI's real constraint is now physics, not model capability. I went looking for the stress test: three sovereign AI bets in the Gulf, three very different answers to the same power and cooling problem.
What I Learned Watching Ants Outperform Humans at MIT's Agentic AI Conference
Notes from the NANDA Summit at MIT Media Lab: why intelligence lives in the interaction between agents, and why leadership has no visibility into what agents already run inside their companies.
Performance Gets You the Demo. Certifiability Gets You the Factory Floor.
NVIDIA just sent its robotics safety stack to two separate inspection labs. Agility is doing the same for Digit. That's not a coincidence, it's the industry quietly admitting that the real gate on humanoid deployment isn't model capability. It's whether anyone can certify the thing.
250 Years of Betting on What's Next
250 years ago, a small group of people bet on a system built to make risky bets, over and over. Right now, AI, robotics, and space are all running that same experiment at once, faster than any prior American platform shift. Here's the pattern connecting a chatbot, a warehouse robot, and a reused rocket booster.
The Model That Beat Its Teacher Has a Speed Problem
A new report scored 40 humanoid foundation models against each other. One number in it quietly settles an argument I've been watching for two years: Vision-Language-Action models, the backbone of nearly every humanoid robot today, just got beaten by more than double on tasks they were never trained for.
The Gap Nobody at Automate Has Named Yet
I spent a couple of days earlier this week walking the floor at Automate, the largest robotics trade show in North America, looking for one specific thing. What I found wasn't a gap in anyone's effort. It was a question almost nobody is asking yet, and a market-sized opportunity sitting underneath it.
The Missing Robot ECU - and Who Will Build It
McKinsey just published a supply chain analysis of humanoid robotics. Everyone quoted the actuator cost data. I think the most important sentence in the report was buried three sections in. And it names a gap that no one has filled yet.
The Wrong Race
I spent part of last week at Boston Tech Week, sitting in on technical sessions with founders building in the agentic AI space. One presentation stopped me cold — not because of the product, but because of the framing. The CTO walked through why the entire field is racing toward a solution that is, information-theoretically, the wrong one. This is my attempt to share that framing.
Physical AI Has a Data Problem That Software AI Never Had
Last post I described the architectural gap: robots need two incompatible things simultaneously - a Slow Brain for reasoning, and a Fast Brain for real-time control. That's the infrastructure problem. But there's a second constraint running in parallel, and it's just as structural. It's the data problem. And it's not the kind of problem you solve by waiting for the next generation of GPUs.
Memory Is the Moat
Memory Is the Moat - Why Stateful AI Beats Stateless AI Every Time Following up on my recent series on AI agents, the harness layer, and Systems of Action... One thread I've touched on in every post but haven't fully developed: Persistent memory and context. Let me fix that.
The Fast Brain, the Slow Brain, and the Missing Layer
In my last post, I described two jobs every production robot has to do simultaneously: think, and act. Job 1: Cognition - is what large AI models do well: interpret the environment, reason, plan, adapt. It tolerates variance. It can take a fraction of a second longer and still be useful. Job 2: Control - is different in kind. Motor signals, sensor fusion, safety loops. It needs guaranteed response times, every time, with no exceptions. Not fast on average. Guaranteed. I said these two jobs have fundamentally incompatible compute requirements. Today I want to name the paradigm, because it turns out the research community already has a name for it and the name matters.
The Two Jobs Every Robot Has to Do at Once
Following up on my last post on Physical AI and the next S-curve in autonomous systems... I want to go one level deeper into something that I think is under appreciated: the core architectural tension that sits at the heart of deploying Physical AI at production scale. It comes down to two jobs every robot has to do simultaneously - and the fact that those two jobs have fundamentally incompatible compute requirements.
The Next S-Curve Is Physical
I've been writing about the autonomous AI wave unfolding in software: agents, harnesses, systems of execution, and where the moat forms in the digital layer. The thesis across those posts: when AI stops recommending and starts executing, the value doesn't sit in the model. It sits in the integration layer - in the harness that connects AI capability to real-world workflows. There's a parallel wave building that follows exactly the same structural pattern - but in the physical world. Physical AI. And I think it's the next S-curve.
When the AI is the Executor
I closed my last post with a question: when the AI is the executor, what does accountable governance actually look like? I spent the last couple of days at the YPO Corporate Governance Symposium in New York. Serious board members. Senior governance practitioners. Some of the most credentialed corporate governance faculty in the country. Here's what I found, and what it clarified...
The Trust Reckoning - When Accountability Can't Scale the Way Intelligence Can
Later this week, I'll be heading to New York for a YPO Corporate Governance Symposium. It's an event I've been genuinely looking forward to: an intimate gathering of leaders wrestling with governance under real pressure. The agenda includes a full session on AI governance and emerging technologies, and I'm going in ready to listen, learn, and participate. Most governance conversations I've been part of frame the challenge as: "How do boards oversee AI risk?" I'm bringing a different question...
From Systems of Record to Systems of Action - Where the New MOAT Is Forming
A broader shift is starting to take shape in AI: We are moving from systems that store and inform... to systems that actually act...
The Moat in AI Isn't Where You Think It Is - It's in the Harness
Following up on my recent post on Autonomous AI... Across a series of conversations over the last few months, with founders, investors, and CVC teams, one question continues to resurface: Where is the REAL MOAT in AI agents?
Autonomous Experts and the Next S-Curve: Rethinking Where Value Will Accrue
Last evening, I had the opportunity to dine with a small group of corporate venture and innovation leaders (thanks to Shahid Azim and C10 Labs) to discuss a question that is becoming increasingly urgent: Where should corporate venture invest as AI moves from copilots to autonomous operators? I've outlined my key thoughts and takeaways in the attached writeup and welcome perspectives from others thinking about this transition; particularly how you're seeing autonomy reshape investment strategy and enterprise innovation. Where should corporate venture invest as AI moves from copilots to autonomous operators?