Hub · everything on one subject · 8 articles · 3 repos & apps
Physical AI and World Models
Physical AI is AI that lives in the real world: it sees through cameras, moves through robots, and rides on your face as glasses. A world model is the part that lets it imagine what happens next before it acts, the way you picture a ball's path before you catch it.
In engineering terms: perception of video and motion, simulation (NVIDIA Isaac Sim) and synthetic data, world models that predict the next state from an action and are calibrated from simulation to reality, and head-worn devices as the interface. The hard problems sit at the edges: sim-to-real transfer, latency, and what a device can actually do today.
It is my current work: technical lead on a Physical AI team, first author of a world-model paper on sim-to-real calibration (under review at NeurIPS 2026), and a documented audit of which of my real tasks AI glasses can already do.
Physical AI and robotics (4)
Why a physical-AI company is a fleet of learning loops, not a robot demo.
- The AI-Native Company Has a Nervous System And Physical AI Gives It Hands, Wheels, and Very Expensive Ways to Learn From Mistakes
2026-07-05 · LinkedIn
Most companies claiming to be “AI-native” are not AI-native. They are regular companies wearing an AI hat.
- The AI-Native Physical AI Company Build a Fleet of Learning Loops—Not a Robot Demo With a Big GPU Bill
2026-07-05 · LinkedIn
Most companies think the path to Physical AI looks like this: Buy robot. Add foundation model.
- Why Your AI + Hardware Project Feels Like Groundhog Day
2026-02-23 · LinkedIn
A practical guide to the closed-loop problem in physical AI development — and how to fix it before it kills your momentum. I
- Machines Can Finally Watch Video. The Hard Part Is Teaching Them to Understand Motion.
2026-07-17
A Vision-Language Model that captions a frame is a tourist with a camera; one that understands motion is a witness who can testify — and the gap between them is where most of the value in video AI lives. A mental model a 15-year-old can run: a VLM watching video is a brand-new student driver, and three things turn the student into a driver, none of them a bigger model. (1) Time-glue — spatiotemporal reasoning that binds the same object across frames so 30 snapshots become one scene with a start and end, not per-frame captions. (2) A strict examiner — an agentic-eval harness that grades localization + temporal + narrative consistency against reality and refuses to pass a claim it can't back with evidence (no evidence ⇒ no claim; a benchmark you can't reproduce is a vibe, not a score). (3) A practice-test factory that can't cheat — the 'AI training AI' curation flywheel where the model proposes labels on raw footage, only clips where independent passes agree survive, and a human gates promotion, governed by maker ≠ checker so the system can't launder its own hallucinations into ground truth. The punchline: everyone rents the same multimodal brain, so the edge is the loop — fine-tune → evidence-gated eval → human-gated curation → fine-tune — not the model. The model is the cheapest part; the loop is the moat. The same lesson recurs across driving video, protein folding, and aging biology. Tooling open at github.com/wjlgatech.
World models (1)
From six lines of Stanford's CS229 to a demo you can play: learn the wolf, guard the flock.
- World Models Start in Six Lines of CS229: Learn the Wolf, Guard the Flock (for Age 20)
2026-10-16
A complete world model small enough to read in one sitting is six lines in CS229 section 19.3: count what the world does, plan with value iteration on what you counted, act, repeat. Play it as a shepherd learning a wolf (learning lifts it from 17–37% of wolves turned away to 87–100%, seed 1), then climb to Kalman, Dyna, World Models, Dreamer, MuZero and Genie, with every famous claim's hedge kept.
Wearables: AI glasses (3)
The Wearable Ladder: what glasses can do now, why 19 of 23 head-worn computers died of money, and dated predictions.
- AI Glasses, by the Documentation: 11 of My 18 Laptop Tasks Should Work Hands-Free
2026-09-30
Eighteen laptop tasks, written down first, each judged against Meta's own documentation for Ray-Ban Meta glasses: eleven documented as working hands-free, four not supported, three undocumented. Messages, calls and the calendar are on the yes side; every work app — WeChat, Feishu, GitHub, a coding agent — is on the no side. The limit is who the platform lets in, not the hardware. The same tasks are pre-registered for a road-trip day that will test the documentation, logged live.
- Head-Worn Computers Died 23 Times — 19 of Them Died of Money, Not the Computer
2026-09-30
A coded table of every head-worn computer sold to buyers since 1989 — smart glasses, visors and headsets, with a source on every row. Twenty-three died a documented death. Nineteen died of money, two of the computer, one of social backlash, one of the display, and none of battery or heat. A fighter's career rarely ends on a knockout; it ends when the purse stops covering the camp. The pair that finally sold in millions removed the most expensive part and sold through an eyewear company.
- AI Glasses Replace the Desk Only When Five Things Come True — Here Are the Dates
2026-09-30
Every AI-glasses keynote says the desk is going away and none says when, so none can be wrong. This ledger can: eight predictions, each with a resolution date between 2027 and 2030, the public evidence that settles it, and my probability today. Five are the conditions that must all come true before glasses replace a desk — all-day power, fast non-keyboard writing, a real display, open app publishing, and the laptop becoming the server an agent drives on your behalf. Three are tripwires for the market and the law. The build fails if any goes unscored thirty days after its date.
Repos and live apps (3)
- Shepherd's World Model (live demo)
A world model you can play: a shepherd learns a wolf's hidden habit by counting, plans with value iteration, and guards two flocks. Kid and engineer modes, seeded so every number reproduces.
- AI glasses audit (live)
The task-by-task audit behind the Wearable Ladder: which laptop tasks should work hands-free on today's glasses, by the documentation.
- allin-anything
All-in interaction with the digital and physical world: one super-repo composing the -anything/-os agent family.
Open the live app →