Ivan Poupyrev

Grateful to @ethanjb and @Venrock for the conversation onstage at @StartupGrind, and for backing our bet on AI for the physical world. We're surrounded by sensors we understand in isolation, but together they represent physics today's generative models can't reason about. Recent example: a customer asked us to find anomalies in their wind turbine data. Instead of the failure modes they already knew, we had to find the ones they'd never seen. Newton was able to identify these failures because it has a foundational understanding of physical behavior in complex systems.

New demo: Invisible Lost Time in offshore drilling costs $10K–$100K+ per instance and is often caused by gaps between rig sensors and real-time visibility. Our new Drilling Monitor, built on Newton World Model, uses open data from Equinor's Volve field (14 North Sea wells, 2007–2009) to address this challenge. How it works: 🔹 9 sensor channels (ROP, RPM, hookload, flow rate, etc.) stream into Newton in 25-sample windows 🔹 Newton's Process Monitoring Agent embeds each window and runs few-shot classification against 1,000 reference examples to return rig state live: DRILLING or NOT_DRILLING 🔹 No per-well training (one foundation model and live rig state across every well) Traditional ML needs one model per asset, retrained when conditions drift. Newton fuses multichannel data into a single representation that transfers across wells.

We benchmarked Newton World Model across a number of smaller devices including CPU-only. The main finding: Newton is deployable across the entire stack. Mac M4 configurations deliver the highest throughput, as expected. The logarithmic chart on the left brings components with very different throughput regimes — including the MAC-class components — onto a single plot. The chart on the right also shows that CPU-plus-GPU edge servers deliver more than enough capacity for sustained, multimodal monitoring at the site level. On a constrained CPU like a Raspberry Pi 4, Newton still produces meaningful inference for scoped tasks; that mean it can be leveraged for use cases that previously had no realistic AI path at all.

To put this in perspective: there are three fundamental physical mechanisms for practical sound generation: mechanical vibration, unstable airflow, and electromagnetic actuation, i.e. speakers. All the music and sound you hear today is made using one of these three. Here is a fourth: plasma. When modulated at audio frequencies, it heats and expands the air rapidly generating audible pressure waves, aka thermoacoustic sound generation. Most of us have heard plasma during a thunderstorm 🌩️. Also you can buy small plasma speakers online, sometimes called ionophones. Historically, Nikola Tesla never explicitly talked about making sound with his coils and arcs, though in 1891 he patented a design that some of the ionophones are based on — a light bulb that uses an electrical arc discharge instead of a regular filament developed by Edison. Technically, Tesla's light bulb design does point toward an “eternal life” light bulb concept — a favorite topic of conspiracy theorists arguing about "planned technological obsolescence", science fiction writers and once-popular job interview question at McKinsey. Fun stuff. Turn the sound on to hear plasma.

They're modulating electrical arc produced by Tesla coil with an audio wave — the arc itself becomes a plasma speaker so you can play music on it. First time I saw it IRL. Ran into this a couple weeks ago when I was visiting @MIT. Sound On.

Great panel at @StartupGrind with @ethanjb from @venrock on being the first Physical AI company. Hard part: most people won't believe you or understand it. The best decision: getting the @PhysicalAI handle early. Documented proof matters when everyone jumps on the bandwagon.

Running inference locally can: - Cut cloud egress costs - Reduce dependence on always-on connectivity - Keep proprietary process data protected. It also makes the financial case for AI more predictable. You size the hardware once, and the marginal cost of running another inference doesn't scale linearly with usage.

It was ahead of its time: youtube.com/watch?v=pRuNsc…

Went to @StartupGrind to speak on a panel and ran into Alex Pashintsev after many years. Back in the ’90s at Paragraph, he led development of the first real-time pen-based cursive recognition system shipped on Apple Newton. He gave it to me for my VR work back then. Legendary.

In one famous episode of The Simpsons, Homer finds a counterexample to Fermat’s last theorem spklr.io/6014EKlL6

Sovereign AI was the other big theme at @Semafor World Economy this month. Industrial leaders treat their operational data (vibration, telemetry, process signals) as core IP, and nobody wants to hand it to a hyperscaler where it might train models their competitors run. That's the challenge we're tackling at Archetype AI. We're building a foundation model and platform organizations can deploy in their own environment, fine-tune on their own data, and own outright.

This Wednesday at @StartupGrind 2026, I'm doing a fireside chat on what it takes to be a category pioneer. Starting Archetype AI meant inventing the language for talking about Physical AI and building the product simultaneously. April 29, 2:30 PM, Breakout Stage. Come find me if you'd like to talk about building AI systems for the real world. #SGC2026

Every deck, document or image these days looks like it was made by Claude to the point that when a deck doesn’t, your first thought is: “Who made this?” I think I will stop spel cheking my posts to make them look rel.

“When Newton published [his theory of gravitation], quite a lot of other scientists were disappointed. They wanted to know what gravity was, and Newton apparently didn't tell them what it was; he told them what it did. Then Hume came along and said that actually, the kind of thing Newton did was the only kind of thing the human mind could ever do.” —Simon Blackburn on the philosophy of science

New demo - Grid Monitor powered by Archetype AI Newton. California's grid moves hundreds of megawatts in minutes. Solar surges at noon and collapses at sunset; batteries have to catch the swing before evening peak. A missed forecast or late curtailment call can result in operators loosing millions. Grid Monitor visualizes CAISO supply and demand in real time, then lets Newton explain what the operator is seeing. How it works: 🔹 10 fuel sources rendered at 5-minute resolution 🔹 Actual demand overlaid against day-ahead and hour-ahead forecasts 🔹 Newton queries the live feed and answers in natural language: why batteries are charging, where the duck curve is bending, when renewables are being curtailed Traditional analytics can only offer a static chart. Newton can generalize across use cases and facilities and create a dynamic representation of what's happening in the real world.

Introducing Newton Fine-Tuning: your operations are unique and your AI should be too. What does it mean for a machine to understand the world it's in? Not the world in general, but the specific one it belongs to. Train a model on enough physical signals — vibration, current, acoustics, thermal — and strong zero-shot generalization emerges across assets and industries. That’s what Newton is: our Physical AI foundation model. For most deployments, a handful of examples is all it takes to put it to work. For @PhysicalAI customers with years of their own operational data, fine-tuning opens another door. A foundation model gives you a general prior over physical signals; fine-tuning turns those weights into a compressed encoding of how a specific facility actually behaves, as a world on its own. Local world models are one of the more underexplored ideas in AI right now, and fine-tuning is the mechanism behind them.

For 40+ years, building a robot that could rally with an elite human table tennis player at full speed was an unsolved problem. Sony AI's Ace research project set out to change that—and the results are now accepted for publication in @Nature and featured on the cover.

Happy to join @FutureEnergy_VC and @Cathayinnov on a panel at SF Climate Week — April 21 in San Francisco. We're seeing a shift in where compute happens and that makes the cost matter as much as the model. A Physical Agent on a turbine can't always round-trip to a hyperscaler. Latency, bandwidth, and data sovereignty push inference to the edge. The energy implications are just starting to land.

Thanks to @Politico for including @PhysicalAI in this week's piece on the world models and the key use cases for Newton. Read on: politico.com/newsletters/di…