Spotify Podcast Link

Today’s guest is Will McDonald, co-founder and Chief Product Officer of Miris.

Miris is rethinking how high-fidelity 3D content gets delivered at scale. Their thesis: the future of 3D depends entirely on simplifying distribution.

In that vein… we talk about why pixel streaming falls short, how Miris uses volumetric data instead, and why that matters for everything from e-commerce and gaming to robotics and simulation.

Will and I worked together for many years at AWS, so we also dive into lessons learned from Amazon’s customer obsessed culture, what it gets right, and experience of going from big tech back into startups.

If you care about spatial computing and simulation, and where the 3D internet is heading, this one’s worth the time.

With that I bring you, Will McDonald.

That said, I hadn’t thought about Sesame Street or its wonderful characters in a long time. Not until this conversation with Pasquale D’Silva.

Pasquale is on a mission to recreate the magic of Sesame Street, but with the power of AI and robotics. In our convo, he waxed poetic about Jim Henson; his vision, his philosophy, and his blueprint for creating characters that deliver joy.

Pasquale is following Jim’s blueprint, but he’s adding a unique spin: the art of clowning.

Yes, you heard me. Clowning.

Sounds strange, but bare with me… It connects back to robotics; particularly, building robots we’ll actually want.

Pasquale’s mission sparked something in me. Perhaps because my daughter Chloe is turning one and becoming a sponge. Or perhaps, because this mission is as important as any other in the AI race.

So I wrote a letter to good ol’ Jim to explain why.

Dear Jim,

I hope all is well in Heaven and that you’re finding good company, including with all your closest friends: Kermit, Big Bird, Miss Piggy, and the whole crew.

Down here on earth? Boy, are things a changin’.

You’d be both thrilled and horrified by the possibilities.

On the plus side, AI is finally real and an entirely new form of puppet is on the horizon. Ones that will be in our homes, raising our children, supporting our elderly, and keeping us company.

On the down side, the engineers are at the wheel, not the artists.

While the result is mind bending technology, the felt experience is soulless. There’s no character, there’s no emotion, there’s no wonder, and as a result… there’s no connection.

These companies are wildly powerful and well resourced, but at times, misguided; driven by a race towards what we call AGI. Their priorities? More data, more compute, more agents.

What these companies really need is more Jim Hensons.

Fortunately, I found a sci-fi Jim in the making. His name is Pasquale and he’s following in your footsteps. His project is small, yet profound.

The goal: build characters and personalities into robots to make them interesting, fun, warm, and worthy of trust.

The proof of concept: via partnership with Boston Dynamics, he’s taken the Spot robot dog and turned it into a character with a real personality, called Spark.

In developing this character, your principles are a guiding light. But Pasquale is bringing in his own flavor that I think you’d appreciate.

It’s the art of clowning.

When I first heard this, it was a head scratcher. But as Pasquale explained, it made complete sense and I realized… Heck, all Sesame Street characters are little clowns in their own right. And damn… we could all use a bit more clowning in our lives, both personally and creatively; especially if we want to design intelligent, embodied systems that humans actually want to engage with.

But not by the mainstream definition… aka a silly (scary?), colorful, red-nosed, goofball.

Rather, by the more by the classical/academic definition. The one you’d find at ‘L’Ecole Philliippe Gaulier’, the famous clown school near Paris.

At Gaulier, the clown is simply the part of you that plays — without agenda, without technique, without the need to be impressive.

In this tradition, the clown is not a costume or a performance style. It’s a state of being. Specifically, it’s your inner child.

Said another way, to study the clown is to study oneself. And needless to say, no two selves (clowns) are alike. The clown is your clown — singular, unrepeatable, drawn from the most vulnerable version of you.

The French clown tradition also relishes the full range of emotions. People tend to think clowns are just one-note funny. But they can also be tragedy, despair, longing.

These elements, these states of being: vulnerability, emotional range, individuality… They all lie at the heart of the characters we love most.

In imbuing robots with character, Pasquale takes the definition a step further. He considers the clown a ‘social technology’ for breaking the frame.

It’s a character device, whose function is to enter a scene and make everything around it visible in a new way. By introducing the absurd, the clown recontextualizes the normal.

Suddenly people see the script they’ve been running, the NPC loop they’ve been stuck in, and they have permission to step out of it.

More specifically, Pasquale would say the clown is the thing that reaches the inner child before the adult has time to intercept. It bypasses the thousand layers of social programming — the system prompt of adulthood — and connects directly with whoever the person actually is underneath it all

My favorite idea from our chat? The clown as a presence, not a performance.

You don’t need to do anything elaborate. You just show up — slightly absurd, fully open, entirely present — and the field opens around you. People approach. They drop their guard. They get weird and real and human in ways they wouldn’t otherwise allow themselves. Kind of like how people behave with a cute and cuddly dog.

This is why Pasquale started with Spark. The clown is a dog in human form.

Dogs are radically present, they wear their hearts on their sleeves, they play without agenda, they connect without pretense. That’s the spirit of the clown. Not a joke machine. Not a performer trying to be funny. Something closer to a being that simply... is, fully, in front of you. And in doing so, gives you permission to do the same.

In our convo, Pasquale gave a powerful example of these effects in action.

You might have seen the movie Patch Adams. It stars Robin Williams, playing a ‘clown doctor’.

Turns out, Patch is a real person (news to me), who becomes a clown when interacting with patients. He’s well into his 80’s now and still practicing!

Patch’s approach was simple. When he first walked into a hospital, he noticed something the doctors couldn’t see from behind their clipboards: the patients weren’t just physically sick. They were psychologically abandoned.

The medical system treats the body like a biological computer — identify the problem, run the fix, send them along. It’s efficient and transactional. And for someone lying in a cancer ward or spending their final weeks in hospice, it’s utterly insufficient.

So Patch showed up differently. In clown. With warmth. Meeting people where they actually were, not where the chart said they should be.

The exemplar scene in the film involves a Vietnam veteran. He’s prickly, guarded, and possibly dying; the kind of man who’d sooner stare at the wall than let a stranger in. When Patch first visits, he shows up as a clown. The guy screams and shoves him out the door.

Most people would leave it there. Not Patch. He went away and asked himself a different question: what does this person actually want to talk about that nobody else will talk to him about?

The answer was death.

So Patch came back dressed as an angel — halo, wings, the whole bit — and started reading him poetry about heaven. And something cracked open. Because that was exactly where the veteran’s head already was. Patch didn’t try to cheer him up or distract him. He just got on his level. Fully. Without flinching.

That’s the thing Pasquale is pointing at; the clown, done right, isn’t performing for the audience. It’s finding the other person’s frequency and meeting them there. It’s presence over performance. And in a world with AI beings in our hospitals, our homes, our most vulnerable spaces… that distinction is everything.

So, Jim… despite the risks with robotics, fret not. There are people like Pasquale who are on the right track; artists who understand the full range of human emotion, the centrality of characters, and their ability to make us feel something. And this, this idea of feeling something… this is the greatest lesson we can take from your work. Your characters have such an internal coherence, emotional range, and distinctive personality that the audience forgets they’re watching a piece of foam, they experience empathy, and they start rooting for them like they’re real.

Kermit wasn’t a prop. He had anxieties. He had hope. He had a particular way of carrying himself through a world that was constantly overwhelming him. We love him because we recognize something true in him.

The body of the puppet, the aesthetics, the movement, all of it matters. But what animates it, what makes it feel alive, is the inner life. Strip that away and you have a very well-crafted object. Keep it and you have a being. Beings that elicit joy.

But joy wasn’t just an experience for you. It was a mechanism.

You believed deeply that joy, wrapped around warmth, humor, and delight, was the most efficient delivery system for everything important: empathy, learning, difficult truths, social values.

People don’t realize: Sesame Street did rigorous research and designed the show around measurable educational outcomes. It just never felt like it, because joy was the wrapper around everything. The lesson arrived before the audience knew they were being taught.

And in these lessons, you never talked down to them. Children especially. You believed they were capable of grappling with real complexity — grief, conflict, difference, fear — if it was given to them through characters they trusted and in a container that felt safe.

The clown and the puppet both serve this function. The absurdity creates permission. Once you’re laughing, your defenses drop, and the real thing can get in.

So, dear Jim, here’s to hoping robots can do the same. I’ll keep you posted on the progress.

Sincerely,

Evan

Hope you enjoyed this reflection on social robots. If so, subscribe below! More to come.

It also followed a day-in-the-life of a robotics developer (named Maya) and a stressful one at that; fumbling and bumbling her way through the lifecycle of robotics data, i.e. a physical AI ‘data flywheel’.

Today, we’re blasting Maya into the future.

This is what it feels like when the data flywheel runs itself.

Enjoy.

PS: When I first wrote this, it was before the release of agentic tools like OpenClaw. Suddenly, this stuff is not nearly as sci-fi as you might think. Much of ‘2033’ will likely be pulled into 2028-29, if not sooner.

Robots-as-a-Service

Maya was an entrepreneur and the kind the world needed to see.

She set the standard for what’s possible in the age of AI. Rather, the age of embodied-AI. The age of near-infinite leverage.

Turns out, companies didn’t make huge headcount cuts. They empowered existing headcount to do more. Much, much more. This IKEA example from 2026 set the stage.

Entrepreneurs did the same, and in robotics, Maya became the shining example.

A roboticist by trade, an agentic tinkerer by night, Maya caught one of AI’s largest waves: owning and managing her own fleet of robots for industrial tasks. Similar to the trucking or large equipment rental business.

Of course, most Fortune 500’s owned their own fleets, with the in-house capability to post-train, deploy, and orchestrate across their infrastructure (factories, warehouses, refineries, you name it).

But the SMB market? A much different story. They needed to outsource and Maya was primed for the opportunity. Texas was a honey pot of small manufacturers and logistics firms. She seized it.

But Maya was also a mom and the kind who refused to succumb to the status quo of choosing between family or work.

With infinite leverage machines, she could do both. Here’s how.

8am: Observe & Collect Experience

Maya’s day doesn’t start with code. And it certainly doesn’t involve alarming Slack messages from her client’s warehouse team.

It starts with the warm embrace of her daughter and giggling laughter in bed.

It then follows with a friend; as Maya steps into her office, a cute little droid rolls in behind her, hand delivering a fresh and foamy latte.

She takes a sip and slips on her glasses. A small display lights up inside the lenses.

“Thanks Bill. Can you show me last nights report?”

Maya’s AR glasses come to life, displaying an immersive array of insights.

It’s not rows of raw robot logs and file names. It’s a clean list of actual events; various failures, anomalies, and edge cases, all detected automatically; a stumble here, a dropped package there, and a surprising number of wrong turns.

Her main KPI in the top right corner is yellow.

The average package transport time has dipped a few seconds. Not good, especially with the holidays right around the corner.

“Bill, show me the failure events with the longest time to recovery and then replay the top three scenes.”

A timestamp floats out of the report: 2:17am Event: Package dropped. Recovery time: 14.3 seconds. Type: Rare and emerging.

Immediate context, on demand.

“Play event.”

Her glasses switch into VR mode. Suddenly, Maya is the robot, reliving its experience.

As expected, each scene showed the same thing; a stack of boxes that were left in the rain, all slippery as can be.

The good old fashion sim-2-real gap strikes again.

9:30am: Understand the Gaps

“Bill, what are you seeing here? Where is the model falling short?”

The scene collapses and Maya is back in her office. Three holographic panels appear, each one replaying a different version of the failure.

“I’ve clustered the incidents,” he says, his voice chirpy and calm. “Eighty-two percent correlate with low-friction surface conditions after rainfall. This scenario is underrepresented in both your real-world dataset and simulation corpus.”

A 3D heatmap blooms into her field of view. It’s not just a spatial dashboard. It’s a diagnosis.

Red islands mark blind spots in the model’s experience. Yellow gradients show where performance degrades gracefully. Green shows confidence.

“So what changed?” Maya asks.

“Humidity,” Bill replies. “The model has insufficient exposure to wet-surface interactions at this humidity level, with these payload weights. And looking ahead... These conditions are going to persist on and off for the next week.”

Maya sat back and smiled.

She remembered the days of manually searching a data lake, knee-deep in SQL queries and half-broken scripts, trying to guess what the robot hadn’t seen enough of.

Now for the fun part…

10am: Creating the Right Experience

“Okay Bill, let’s fix this. Do we have the right 3D assets and physics parameters in our sim?”

A rotating saucer appears.

Numerous 3D models of cardboard boxes are displayed on top, each slightly soggy and deformed. Quite the desert plate.

“Here are the current assets in our library,” Bill says

Something is off. The boxes in the videos were a different shape, size, and color.

“Ah, forgot we got a new box vendor! Bill, extract images of the boxes and let’s GSplat ‘em”

“On it.”

Maya looks over to the video panels. A shimmering line traces the new boxes and pops them out like a cookie cutter. The images blur, spin, and voila: what was once a 2D image is now a perfect 3D model, with the exact same texture and lighting as the boxes in the video.

“Great. Now let’s optimize the physics engine and create the scene,” Maya commands.

“A step ahead of you. Here’s the plan.”

The 3D models and videos disappear. In their place, a workflow diagram, showing all the detected gaps and steps Bill took to fill them; new box material with new friction coefficients, new payload mass, new actuator latency to match the warm, damp motors.

“Great. Now randomize the domains across 10,000 different variations.”

A new report surfaces. Each scenario is parameterized, ranked, and tagged with an estimated learning yield.

These aren’t one-off virtual worlds anymore. They’re flexible experiences the system can probe and stretch. Now, the digital robot doesn’t just repeat the same failure; it explores all the variations around it.

In the past, Maya had to sit with a team of 3D developers, hand-authoring edge cases based on intuition and experience, hoping she’d imagined the right failures.

Now, simulation isn’t driven by imagination. It’s driven by evidence.

“Run the top tier of scenarios,” she says. “Stop when marginal gains flatten.”

“Already planned,” Bill replies. “Simulation will terminate once gradient contribution drops below threshold. Should take 127 minutes”

The parallelized simulations spin up automatically, along with all of the supporting cloud infrastructure; not as a bespoke experiment, but as a service the system knows how to use.

10:30am: Data Evaluation

Maya steps out for a walk with her husband; a refreshing stroll along the lake near their house. They grab a lite bite and toss bread to a family of waddling ducks.

As they wrap and head home, her device buzzes. She throws on her glasses.

“Hi Maya, beautiful day isn’t it? A quick update: the simulations and data parsing are complete.”

A 3D graph materializes on her horizon. Thousands of simulated trajectories appear, and just as quickly, most of them vanish.

Maya watches a thin curve chart downward as the system discards the majority of simulation runs. Only four percent survive.

“Redundant trajectories removed,” Bill narrates. “Retaining only samples that expand the AI model policy boundary or introduce novel recovery behavior.”

Each clip of the simulations has a short annotation. Things like ‘boundary condition’, ‘compounding failure’, ‘rare recovery’.

In the old world, Maya had to hoard all synthetic data because curation was painfully manual. Now, the curation is automatic and ruthless.

For the first time in her career, she has less data and more confidence.

She also has more time. When she gets home, she takes her daughter to music class and joins in on the musical fray, reliving her first-grade glory days on the recorder.

1pm: Time to Learn

When Maya gets home, she sifts through the curated data samples and summons Bill.

“Good stuff Bill. The data looks good. Let’s update the model.”

A soft confirmation pulse appears in her view.

“Training underway,” Bill says. “I’m blending last night’s real-world success data with the highest-signal simulation runs. I’ll focus the learning on low-friction recovery and weight transfer. Estimated time: 92 minutes”

Real-world data teaches the system what truth looks like. The simulation gives it the reps.

In the old days, this part always felt like a gamble. One bad assumption, one misaligned dataset, and the whole run was toast; discovered only after burning a night of (very expensive) compute.

Worse, real and synthetic data never played well together. They lived in different systems, followed different rules, and told slightly different stories about the same moment.

One came from a messy world full of quirks. The other from a clean digital one that always thought it was right.

Getting them to agree used to take weeks of human glue work. Now, the system treats both as the same thing: pure experience.

Maya leans back as learning continues quietly in the background.

2:30pm: Reality as the Judge

“Maya, training is complete. I’ve deployed a single canary back into the wet environment for validation. Do you want to observe?”

“Yep, give me a side by side”.

Two scenes appear.

On the left, a replay of the simulation: what the system expected the robot to do.

On the right, the real world; what actually happened in the warehouse.

The real robot stopped and stuttered for a split second. Not a huge deal. But could compound over time.

“Why the pause?” Maya asks.

“Motor response latency,” Bill replies. “Today’s humidity is higher than last night, reducing latency in the actuators by about 8%. I’ll adjust the simulation parameters and do one more quick roll out.”

Maya smiled. This is her favorite part.

Simulation is no longer guessing reality. It’s mirroring it.

3:00pm: Quiet Deployment

“The second training run is complete. Rolling out a new model to the first 10 units.”

“Thanks Bill, you’re a gem.”

Maya watches as the first squad of robots revisit the scene of the crime.

“Early indicators are positive. Recovery times are improving. No regressions detected,” Bill says.

Maya now watches the average delivery time KPI. It slowly drifts from yellow to green.

Success. No emergency calls. No one even notices.

3:30pm: Freedom

Maya’s days no longer drag on into the wee hours of the night. Her mission is complete and she can go enjoy the afternoon with family.

As she packs up, Bill chirps.

“Maya, one more thing.”

She pauses. “Whatya got?”

“There’s a cold front coming in three weeks,” Bill says. “Morning temperatures will slow actuator response.”

Maya sighs. “Of course.”

“I’ve already pulled real world data from our Canadian fleet,” he continues. “They saw this exact pattern last month. Cold starts, slight delays. Nothing dramatic, but enough to matter.”

A simple flow appears in her view: real footage from snowy warehouses feeding directly into new practice scenarios.

“Simulations will run overnight,” Bill says. “By the time the cold arrives here, the model will be ready.”

Maya smiles.

When she started this job, surprises like this meant late nights and rushed fixes. Now, the system learns ahead of time.

She heads for the door. Outside, it’s still warm. Inside the data flywheel, winter has already passed.

Hope you enjoyed this glimpse into the future.

Our next essay will break down which parts of this vision are real today, and what will be required to bring it all to life tomorrow.

Until then, subscribe below and stay tuned!

Spotify Podcast Link

This was the most unique and fun episode I’ve done in a while, and it starts in a very unexpected place…

With the art of clowning.

Strange I know, but stick with us, because we quickly connect that world to technology, robotics, and why it holds surprisingly powerful lessons for this moment in time; especially if we want to design intelligent, embodied systems that humans actually want to engage with.

Our guest today is Pasquale D’Silva, CEO of The Illusion of Life, a company building characters and personalities into robots to make them interesting, fun, warm, and worthy of trust.

This all starts with their partnership with Boston Dynamics, where they’ve taken the Spot robot dog and turned it into a character with a real personality, called Spark.

In this chat, we talk about the relationship between humans and machines, and explore all kinds of crevasses, including trust, presence, emotional intelligence, and why artists—not just engineers—may be the key to building joyful, human-centered robots.

If you’re curious about the future of human–machine interaction, you’re gonna love this one. It will stretch your imagination in all the best ways.

So with that, enjoy this episode with Pasquale D’Silva.

This is not just MLOps applied to robots.

It’s an entirely new discipline for running systems that move through the physical world all day, every day.

For the non-technical: MLOps is how dev teams reliably train, deploy, and operate machine learning models at scale.

It was built for the internet and it assumes AI lives in a safe, predictable digital bubble; where nothing can explode and data changes slowly.

If something breaks, it’s usually a math problem, not a physics problem; you just check the dashboard, teach the AI model a new trick, and push a software update.

Over the last decade, we’ve perfected the assembly line for this, with standard tools to organize data, track experiments, and manage versions, all optimized for things like chatbots, Netflix movie recommendations, and spam filters.

These systems live safely behind an API in the cloud, far away from the chaos of reality.

Physical AI breaks those assumptions, creating an entirely different paradigm.

The difference begins with what is being operated.

In MLOps, the primary artifacts are models, datasets, features, and training code.

Data is typically tabular, textual, or image-based, collected passively from digital systems. Models are trained offline, evaluated against static benchmarks, and deployed into well-instrumented digital environments.

In RobotOps, the artifacts are far more interdependent and dynamic: perception and control models, sensor configurations, calibration parameters, 2D/3D maps and world representations, simulation environments, digital assets, and massive streams of multimodal sensor data captured during real-world operation.

Code and models still matter, but they are no longer the center of gravity. Behavior, data, and interactive environments are.

The feedback loop also differs.

MLOps closes the loop through metrics, predictions, and downstream business signals inside digital systems. This includes:

• Click-through rates (did the model’s prediction cause a user to take the intended action?)

• Classification accuracy (how many predictions are technically “correct”)

• Drift statistics (is the data is slowly changing?)

• Loss curves (is a certain math score during training is going up or down?).

From there, data is logged, analyzed, and fed back into retraining pipelines on a human-defined cadence.

In contrast, RobotOps closes the loop through the physical world itself.

A deployed model produces behavior; that behavior interacts with an unpredictable environment; sensors capture the consequences; and those sensor logs must be ingested, indexed, graded, and transformed into new training data and simulation scenarios.

This loop is continuous, not episodic.

Training, validation, and operations collapse into a single, always-on learning system (or at least, that’s the goal...)

Failure also carries a very different weight.

In traditional MLOps, a failure is annoying: a model becomes biased, a user sees the wrong ad, an irrelevant search result is surfaced.

In RobotOps, failure is physical. It means damaged hardware, safety incidents, regulatory nightmares, or the silent poisoning of future training data.

A bad model doesn’t just output a wrong number; it creates a dangerous event.

Because of this, RobotOps has to treat provenance (knowing exactly where data came from) and determinism as survival mechanisms, not just best practices.

Knowing exactly which model, environment, and scenario caused a robot to twitch isn’t a debugging convenience. It’s a safety requirement.

And then, we have the most fickle of beasts: simulation.

This is where the gap between MLOps and RobotOps becomes a canyon.

In MLOps, you look backward.

You validate models by replaying historical data or shadow-deploying alongside existing systems. The model sits safely behind an API, observing the world without touching it.

In RobotOps, you must look forward.

You can’t just ask how a model performs on past data; you have to ask how it behaves when the world pushes back.

To answer that, you need simulation. You need to run candidate models through thousands of procedural scenarios (rare edge cases, lighting changes, sudden obstacles) before you ever let code touch a physical machine.

But here is the hard truth about simulation today: For most teams, it’s a pipe dream.

Ignore the shiny visuals you see in keynote demos. While a handful of advanced, deeply funded companies have achieved full-fidelity digital twins, the reality for the other 95% is starkly different.

Building a photorealistic, physics-accurate virtual world isn’t just an engineering task; it effectively requires building an in-house AAA game studio.

Today, most teams use simulation sparingly, mostly for synthetic data augmentation or simple component testing. It lives on the periphery.

Over the next few years, this will invert.

Eventually, simulation will move from a supporting role to the center of the development loop, becoming the primary environment for validation, regression, and learning.

But first, some core challenges must be addressed, starting with simulation’s fragility; if one small piece breaks, the entire run fails, so you can’t easily retry or patch part of the job like you would with normal ML training.

On top of that, you have massive 3D asset dependencies; each run depends on huge 3D worlds, physics models, and 3D models, and if you can’t replay thousands of simulations in exactly the same way every time, you can’t trust the results or compare models reliably.

The next MLOps vs. RobotOps contrast is automation.

MLOps automation is largely pipeline-driven and rule-based. Humans decide what data to collect, when to retrain, how to evaluate models, and which versions to promote. Automation accelerates execution, but intent remains human-defined. For now...

In RobotOps, the complexity quickly exceeds what humans can manage manually. Deciding which data is missing, which edge cases matter, which scenarios to simulate next, and which models should evolve becomes an ongoing cognitive bottleneck.

This is where AI-native automation enters the equation, and might be necessary to achieve serious scale.

I’m already seeing early signals of this: vision-language models auto-labeling sensor data, world models grading synthetic scenario quality, and agents proposing new simulation campaigns based on observed failures.

Over time, these agents will operate entire segments of the learning loop autonomously.

This is the true inflection point: when RobotOps systems begin to improve themselves faster than humans could ever direct.

To be clear, existing MLOps tools like model registries, training pipelines, orchestration frameworks, etc... these things still matter. But they operate at too low a level.

RobotOps requires higher-order abstractions: scenarios instead of datasets, behaviors instead of predictions, simulation campaigns instead of experiments, data grading instead of drift detection, and learning loops instead of deployment cycles.

In this sense, RobotOps is not just the next evolution of MLOps. It is the operational layer for embodied intelligence: systems that learn through action, adapt through experience, and operate under physical constraints.

But before this can be true... we must remove a ton of pain at every stage of the data lifecycle.

Other wise, we’ll never reach the RobotOps holy grail: a fully automated, physical AI data flywheel.

The Physical AI Data Flywheel

Every company has their own version of such a flywheel. Most look something like this.

To really feel that pain, let’s walk through a typical ‘day in the life’ of a robotics developer.

Her name is Maya.

08:00 AM - The Wake-Up Call (Observe + Collect)

Maya’s day doesn’t start with code; it starts with a Slack notification from the manufacturing ops team:

“Robot #104 stalled in the Cincinnati facility. It thinks a shadow is a solid object. Again.”

This moment straddles phases 7 & 1: Observe + Collect Experience.

The robot failed in the real world, but Maya has no idea why. The dashboard shows a generic Navigation Error, but the neural network offers no explanation. She is staring at a black box failure. To fix it, she needs to see what the robot saw— she needs the log data.

The underlying challenge: Robots fail silently, and observability is shallow

In robotics, most failures are not clean crashes. They’re subtle misinterpretations of the world: shadows, glare, dust, vibrations.

Yet most fleet monitoring systems surface only coarse error codes, not the perceptual or decision context that caused the failure.

Unlike traditional software, where logs explain behavior, neural policies often provide no causal trace. The “observe” loop exists, but it’s thin: teams know that something failed, not why.

This makes learning reactive, slow, and highly manual.

09:30 AM - The Sneakernet Struggle (Ingest & Organize)

Maya tries to pull the logs from Robot #104 remotely, but she hits the ingest bottleneck immediately.

The robot is deep inside a concrete warehouse with spotty Wi-Fi. It has generated 500GB of LiDAR and camera data in the last shift alone, but the upload speed is crawling at 2MB/s.

She can’t wait three days for the upload. She let’s out a deep sigh and calls the site manager in Cincinnati. He answers and she quickly makes her ask.

“I need you to physically walk out to the unit, pull the SSD drive, and upload it from the main office fiber line,” she says.

In an era of 5G and Starlink, Maya is still relying on the “Sneakernet”, physically moving drives because the bandwidth cost of streaming terabytes of raw sensor data is simply too high.

She is forced to make a blind decision: drop 90% of the data to save bandwidth, hoping she doesn’t accidentally delete the one frame that captures the failure.

The underlying challenge: Physical data is heavy, remote, and bandwidth-constrained

Robotics data is fundamentally different from web or mobile data. High-resolution video, LiDAR point clouds, and dense telemetry quickly reach terabyte scale. And robots often operate in warehouses, factories, fields, or mines with poor connectivity.

As a result, data collection becomes opportunistic and lossy. Teams must choose what not to upload before they know what matters. This breaks the data flywheel at its foundation: learning systems depend on complete context, but physical constraints force premature filtering.

11:00 AM - The Needle in the Haystack (Index & Curate)

The logs finally arrive. Now Maya faces the data curation nightmare.

She watches the replay: a forklift drove by with a reflective safety vest that flared in the camera lens, blinding the robot’s depth sensor.

She knows exactly how to fix this: retrain the vision model on more examples of high-glare reflections.

She opens her company’s S3 bucket, which holds 4 petabytes of video data from the last two years. But it’s a data swamp.

The files are named things like:

“log_2024_10_12_robot_55.bag”.

There is no search bar. She cannot type: “Show me videos with lens flare.”

Instead, she spends the next three hours manually scripting queries and skimming through random videos, hoping to find boring footage of shiny objects.

She finds 200 hours of useless driving for every 10 seconds of relevant glare.

She is drowning in data, but starving for information

The underlying challenge: Robotics data is often unindexed, unlabeled, and semantically blind.

Robotics teams collect massive volumes of data, but lack tools to search it by what happened, not when it was recorded. Failures are semantic—glare, hesitation, near-miss—but storage systems only understand filenames and timestamps.

Labeling is manual and expensive, so the most valuable edge cases stay buried. The real tax on iteration speed isn’t data collection—it’s curation.

The root cause is structural. Robotics data lives in formats like ROS bags or MCAP—black boxes that contain everything, perfectly synchronized. But once stored in the cloud, they become opaque blobs. A system can see a 50GB file, but not what’s inside. Container formats solve storage, not understanding. So teams build fragile pipelines and scripts just to find a single failure—turning weeks into days, and learning into archaeology.

02:00 PM - The Waiting Game (Train)

After painfully assembling a small dataset of “shiny vest” scenarios, Maya kicks off a new training run. This is the training bottleneck.

She is retraining a massive Vision-Language-Action model. Even with a cluster of H100 GPUs, this will take 14 hours. If she made a mistake in the data formatting — if even one timestamp is out of sync between LiDAR and camera — the model will hallucinate, and she won’t know until tomorrow morning.

She hits Run and crosses her fingers, knowing that her iteration speed is capped not by ideas, but by compute latency.

The underlying challenge: Training is fragile, slow, and tightly coupled to data quality

Modern robotics models are compute-hungry and extremely sensitive to data alignment.

A single synchronization bug can invalidate an entire training run, wasting hours or days of GPU time. Unlike web ML pipelines, robotics training must reconcile multiple sensor streams, action spaces, and temporal dependencies.

As models grow larger, iteration speed collapses; not because engineers lack insight, but because feedback loops stretch across days instead of minutes.

04:30 PM - The Crisis of Confidence (Validate)

While the model trains, Maya reviews a candidate model from the night before. It’s time for validation.

In a 3D simulation, the model scores 99.9% success. It navigates the digital warehouse flawlessly. But Maya feels the familiar sim-2-real anxiety, aka the sim-2-real gap.

The simulator is clean; the real world is dusty, vibrates, and smells faintly of oil. Alas, the occasional oil spill, aka robots doing their best ‘Bambi on ice’ impression.

(insert Bambi on ice image)

Maya remembers last month’s incident: a model that passed every sim test but caused robots to oscillate violently because the simulator didn’t model the latency of new motor controllers.

Without a trusted “golden regression set” of real-world scenarios, validation becomes less about proof and more about gut instinct.

The underlying challenge: Simulation confidence does not equal real-world safety

Robotics validation lives in an uncomfortable gray zone. On one hand, simulation is necessary but incomplete. On the other, real-world testing is slow, expensive, and risky.

Most teams lack standardized, trusted regression suites grounded in historical failures.

As a result, validation becomes subjective—an exercise in risk management rather than evidence. This uncertainty compounds as fleets grow, making every deployment decision emotionally and organizationally heavy.

06:00 PM - The Button Press (Deploy)

It’s time to push the patch. This is the deploy fear.

Unlike a web developer who can roll back a bad update in seconds, Maya is shipping code to a 300-pound machine moving near & around humans.

A bug here isn’t a broken UI, it’s a collision or worse.

She initiates a canary rollout, pushing the update to three robots in a low-risk zone (Canary rollout = deploying to only one or two robots to ensure everything’s working properly)

Maya will watch them obsessively for the next 24 hours. The fear of bricking hardware or causing harm slows everything down.

The team could deploy daily, but because of the risk, they deploy monthly.

The underlying challenge: Deployment carries physical, reputational, and regulatory risk

Robotics deployment is software delivery with real-world consequences.

Rollouts must be staged, monitored, and reversible—but many teams lack mature tooling for fine-grained canaries, fleet segmentation, and automated rollback.

Each deployment becomes a high-stakes event, encouraging conservatism. The result is a slow, cautious flywheel where learning is throttled by fear rather than ambition.

The Cycle Continues

Maya leaves work. The model is still training. The canary robots are still running. Tomorrow morning, she’ll open Slack to see whether the flywheel spun forward, or whether a new edge case brought it grinding to a halt again.

The meta-challenge: The flywheel exists, but it doesn’t compound

In theory, robotics should improve continuously: failures generate data, data trains models, models improve behavior.

In practice, each step is fragmented, manual, and fragile. The loop turns, but slowly, unevenly, and with enormous human effort.

Until the data flywheel becomes automatic, semantic, and tightly integrated end-to-end, robotics progress will continue to feel heroic rather than inevitable.

With that... Stay tuned for Part III.

We’ll be re-examining a day in Maya’s life, but this time, in the not-too-distant future; in which these challenges are solved, the fly wheel is automated, and human-robot collaboration is the norm.

We’ll also dive deeper into simulation, and all its fun challenges, traps, and promise.

So if that interests you, be sure to subscribe! More to come on the future of physical AI and being human out on this wild new frontier.

Oh, and in case you missed Part I of this series, here’s the link:

Physical AI Deep Dive: Part I | Market Timing

It answers the ‘why now’ question for physical AI + robotics.

The answer in a sentence: breakthroughs in data collection, vision language action models, and training recipes, i.e. bespoke combos of SL + RL/simulation + IL.

Otherwise... until next time!

I’m leaving AWS and joining Nebius, one of the top AI clouds, as their Head of Physical AI.

I’m also leaving my first professional love — spatial computing. At least in its most direct and advertised form, i.e. AR, VR, 3D, aka ‘XR’.

But in another sense, I’m not leaving at all. I’m going deeper.

As we’ve discussed, spatial computing is far more than XR. This is just one end-point; just one form of a computer relying on sensors to localize, track, and understand the physical world.

Self-driving cars, mobile robots, humanoids, autonomous drones... these are all ‘spatial computers’, each relying on a similar technical foundation, with similar computer vision tech, simulation tools, and ‘spatial’ data types, e.g. lidar, point clouds, 3D, video, IoT/sensors.

The key difference?

Autonomy and their ability to impact/interact with the physical world.

So again, I’m not leaving. I’m just moving downstream; from a ‘spatial tributary’ to a gushing river, surging with capital, progress, and commercial opportunity.

Jensen Huang (NVIDIA’s CEO) popularized the term ‘physical AI’ about a year ago. As it goes… interest and awareness has exploded ever since.

To be sure, Jensen likes to live 10+ years in the future. Physical AI remains early and is in a hype cycle of its own. One that, at first glance, feels like the hype cycle that sucked me into XR (heartbreak and shattered dreams, abound).

But this feels different. The timing feels right.

This essay summarizes why. You’ll leave here with a L200 grasp of this new tech stack and the key innovations worth monitoring.

In Part II of this series, we’ll cover some of the challenges that remain, including the ones that I’ll be trying to solve in my new role at Nebius.

Subscribe now

But first, some context.

What exactly is physical AI?

It’s the next (and I believe most impactful) phase of the AI revolution.

Definitions vary, but I like to keep things simple. I define physical AI as computers/machines that can understand, represent, and/or interact with the physical world.

It’s a broad umbrella with common underpinnings; the big three are real-world data, computer vision, and of course, generative AI.

The result doesn’t have to be as complex as a robot. It can include an AI-powered camera on the factory floor, recognizing anomalies and then adjusting an assembly line in real time.

It can also include AR glasses, with cameras gathering front-line insights and providing work instructions, or informing a soldier on the battlefield.

(This is what I’m most exited about — the combo for AR + robotics to keep humans ‘in the loop’, i.e. allowing both humans and machines to see the same thing and work/sing off the same sheet of music)

More abstractly, Physical AI is the merging of the physical and the digital world; the connection of atoms to bits.

Along the way, this includes the purely virtual, in which the physical world is reconstructed & simulated.

The most exciting examples here are Large World Models like Googles Genie (GenAI that can create virtual worlds with a single prompt) and digital twins: the simulation sandboxes where ‘robots learn to be robots’.

(If you haven’t seen this two minute Genie demo video, stop right now & watch. Trust me)

Most importantly, physical AI holds the key to tech’s ultimate endgame: artificial general intelligence (AGI). That key is the embodiment of AI. Without it, machines will struggle to gain a true understanding of our world, its physics, its randomness, its interconnectedness, our larger role within it all. With physical AI, perhaps the greatest mysteries of the universe will be revealed...

I find it all endlessly exciting and hopeful, and just incredible that we get to experience this phase of human history.

But I digress. Putting my realist cap back on.... We still have a little ways to go.

How much further?

I believe we’re where LLM’s were between 2017 - 2019.

World class teams have been funded, foundation models are in development, compelling demos are sparking the imagination, and VC dollars are flowing.

Most notably, many of the top AI minds from academia are leaving the labs in droves, founding companies and taking the commercial leap (e.g. Chelsea Finn & Sergey Levine @ Physical Intelligence, Ken Goldberg @ Ambi Robotics)

As such, it feels like we’re 3-4 years from the ChatGPT moment and 5-7 years away from this stuff hitting meaningful scale (e.g. robotic manufacturing & supply chains are dialed, millions of units are being deployed in the enterprise, and consumer adoption is starting to ramp).

And this is slightly conservative. After all... humans are terrible at grasping exponential curves.

What exactly is the ChatGPT moment?

It’s when we have a generalizable, foundation model for physical AI. Aka: an API to the real world, that can understand, navigate, and complete general tasks in a ‘zero shot’ fashion. Meaning, it can perform a task it has never been explicitly trained on before.

And when I say generalization, I mean across numerous variables, the most important being hardware. Meaning, it works with not just one form factor, but across many form factors, translating data + context into an action/output on humanoids, quadrupeds, AMRs with four arms, or even AR glasses.

So why now?

Because these foundation models are starting to arrive, giving developers a substrate upon which to further train & fine tune.

The first wave emerged in 2023/24. The most notable being Google’s RT-1 and RT-2, Physical Intelligence’s π0, Covariant’s RFM-1, and NVIDIA’s Project GR00T.

You can liken these to OpenAI’s GPT-1 & 2. Not super useful out of the box, but mind-blowing technical breakthroughs, nonetheless. They’re the result of upstream innovation in three main areas: (1) new architectures, (2) new data regimes, and (3) new training recipes.

(1) New Architectures

You’ve probably heard of the Transformer: the architecture that quietly changed the AI world in 2017. It’s the engine behind ChatGPT, Gemini, and every modern large language model. Transformers work by treating language as a sequence of tokens and predicting what comes next.

Its impact on text was obvious. Its impact on physical movement? Less so.

But roboticists quickly saw the parallel. A robotic action is just another kind of sequence. Instead of words, you have motor commands. Instead of sentences, you have trajectories—reach, grasp, lift, place.

Once we applied Transformers to robotics, something remarkable happened: robots could learn useful skills from dozens of demonstrations instead of thousands. Efficiency jumped by an order of magnitude, and for the first time, robots gained a model that could actually generalize across varied tasks. But moving is only half the battle. To interact with the world, you have to understand it.

This is where the Vision-Language-Action models enter the picture (aka VLAs)

Their predecessor was Vision-Language Models like CLIP. These systems fuse text and imagery into the same conceptual space, teaching machines to recognize “dog” as an idea, not a blob of pixels.

As you probably guessed, VLA’s add the missing piece: action. They “tokenize the world,” weaving together text tokens, image tokens, and motor tokens into a single sequence.

If you give them a high-level instruction like “pick up the apple,” and they can identify the apple, understand the intent, and sketch out a plan. It’s the beginnings of a generalist robotic brain.

But a brain that can imagine a plan is different from a body that can execute it.

For that, the industry started turning to Diffusion Policies. This is the same generative magic behind image creation tools like Midjourney or Nano Banana.

In generative art, diffusion starts with pure noise and gradually “de-noises” it into an image. In robotics, it starts with noisy, meaningless motor twitching and de-noises it into a fluid, coherent motion.

This approach solves one of robotics’ oldest curses: the multimodal problem.

This problem arises in tasks that have more than one ‘answer’. Take opening a drawer for example. It has multiple valid solutions: you can grab the handle from the left, right, or top.

Older models (convolutional neural networks) tried to average these demonstrations and ended up reaching for the center, where no handle exists. With diffusion, it doesn’t average; it chooses. It locks onto a single valid mode — left-grab, right-grab, top-grab — and executes it with a kind of crisp, biological confidence.

In short, this gives robots the decisive body that matches the VLA’s semantic brain.

The result is a dual-system architecture. One that looks/feels a lot like human mind, with System 1 and System 2 modes of thinking. And yes, it’s very similar to the concept you’ve probably heard from the book Thinking, Fast and Slow, by Daniel Kahneman’

System 1 (the body) is the the ‘fast’ diffusion policy. It’s reactive, and precise. It takes a strategy and turns it into the 50-millisecond motor adjustments needed to actually grasp a crinkled wrapper.

Systems 2 is the ‘slow’ brain, powered by VLA. It’s deliberate and semantic. It looks at the room, understands “I need to clean up that trash,” and generates the high-level strategy.

But even with this powerful combo, the robot is still just mimicking. It doesn’t truly know if it succeeded. To get closer to human-like intelligence, you need a closed loop system that can learn from its mistakes and upgrade itself on the fly (we’ll cover this further in Part II).

(2) New Data Regimes

Okay, so we now have this dual-system architecture, giving us the right ‘engine design’ for physical AI. Next we need fuel, aka data.

In the purely digital world of LLMs, fueling AI is relatively easy. You scrape the entire internet — billions of images, trillions of tokens of text — and feed it into a GPU cluster running frameworks like PyTorch or JAX.

But in the physical world, there is no internet to scrape. You can’t just download “how to fold a shirt” or “how to solder a circuit board”. You have to create it (although, heavy research is underway towards making YouTube videos digestible for robotics).

This is currently the single biggest bottleneck in Physical AI.

To solve it, the industry has made exciting progress in three key areas: data-crowdsourcing (The Community Pot), teleoperation at-scale, (The Grind), and simulation (The Infinite Well).

The Community Pot: Open X-Embodiment

For years, robotics was a fragmented discipline. A lab at Stanford would collect data for a Franka Emika arm and a lab at MIT would collect data for a Kuka arm. But the data was siloed, incompatible, and useless to anyone else.

That changed with the Open X-Embodiment Collaboration.

Think of this as the “ImageNet moment” for robotics. Led by Google DeepMind and 20+ other institutions, this project pooled data from 22 different robot types into a massive, standardized dataset.

The hypothesis was simple but powerful: positive transfer. They bet that a robot arm in California could learn better grip strategies by studying the data of a completely different robot arm in Zurich.

And they were right. Models trained on this collective “community pot” of data outperformed those trained on narrow, proprietary datasets. It proved that if we stop hoarding data and start standardizing it, we can bootstrap general intelligence much faster.

The Grind: Teleoperation Fleets

Community data is great for general skills, but for high-precision tasks, you need the “gold standard” data, aka human intuition captured at 50Hz.

Enter teleoperation fleets.

If you walk into the offices of the top Physical AI startups today, you won’t just see engineers coding. You’ll see rows of operators wearing VR headsets or using haptic controllers, “puppeteering” robots to do chores, assemble parts, or sort logistics.

This is the grind. It’s painfully slow and expensive (1 hour of human time = 1 hour of robot data). But it is currently the only way to capture the subtle “English” a human puts on a screwdriver, or the way we instinctively jiggle a key to make it turn. This data creates the kernel of knowledge that models need to imitate.

Fortunately, we now have a wave of companies like Scale AI, Sensei, and Deta who are industrialization this capability, standing up “Data Engines-as-a-Service” to collect thousands of hours of expert demonstrations.

Helpful and promising... but its not enough.

The Shortcut: UMI

To ease the TeleOp grind, a promising new idea surfaced back in 2024. It’s called Universal Manipulation Interfaces (UMI).

Instead of teleoperating a specific robot through complex rigs, UMI flips the model. Humans manipulate the world directly using a lightweight, handheld gripper with a wrist-mounted camera.

This captures rich, first-person visual and motion data—how objects move, how forces are applied, how tasks unfold in real environments—without requiring a robot in the loop at all. In effect, it decouples data collection from robot hardware, turning natural human interaction into training data.

The real unlock is generalization. Because UMI represents actions in a relative, hardware-agnostic way, these demonstrations can be transferred across different robot arms, grippers, and platforms with minimal retraining.

It doesn’t replace teleoperation fleets or simulation, but it meaningfully expands the funnel; lowering cost, increasing diversity, and capturing human intuition at scale.

If teleoperation is the gold standard and simulation is the infinite well, UMI sits in between: a faster, more portable way to turn human skill into robot competence.

The Infinite Well: Solving Sim-to-Real

While teleoperation and UMI capture quality, they both have a hard limit: human time.

You can’t hire a billion humans to teach a billion robots. Eventually, the robots must learn in the virtual world. Especially if we want to capture all the edge cases and complexity of the real world.

This is the promise of what the industry calls ‘Sim-to-Real’ (aka learn in simulation and transfer a policy to reality)

In the past, simulators were essentially video games with bad physics. If a robot learned to walk in code, it would often fall over the second it stepped into the real world because the friction was slightly off or the lighting was different.

The industry has started to solve this with a technique called ‘domain randomization’.

Think of this as the “brute force” approach. Since you can’t match reality perfectly, you just randomized everything, creating chaos across every possible dimension; lighting, floor lay outs, friction levels, temperature, sensor noise, etc.

The ideas is that if a robot can survive the chaos, the real world will feel easy by comparison.

This works to an extent, but it’s inefficient, lacks complete accuracy, and leaves ‘reality’ gaps.

But those gaps are starting to close. Three new technologies are maturing that can hep us clone reality rather than just guess it.

The first is neural rendering (aka gaussian splatting).

Traditionally, building a simulation meant hiring 3D artists to hand-draw every table, screw, and factory wall. This is slow, expensive, and rarely 1:1 with reality.

Gaussian splatting flips the script. It allows us to walk through a factory with a camera, scan the environment, and “splat” a photorealistic digital copy into the simulator. It captures the messiness of the real world: the rust on the metal, the grease on the floor, the weird reflection on the window. This matters because robots are easily confused by visual noise.

Now, they can train on the exact visual chaos they will see on deployment day.

Next we have ‘Differentiable Physics’ within the simulations. Old simulators were “black boxes.” If a robot tried to pick up a cup and failed, the simulator effectively just said, “You failed.” It didn’t explain why.

The robot had to guess randomly to fix it.

Differentiable simulators (like those powering NVIDIA’s Isaac Lab) are “glass boxes.” They provide mathematical gradients. Instead of just saying “Fail,” the simulator says, “You failed because your grip force was 10% too low for this friction coefficient.”

It gives the AI a roadmap to the correct solution, speeding up the learning process exponentially.

With better sim tools in hand, the next breakthrough is “Real-to-Sim-to-Real Loops”, allowing developers to fine-tune the virtual world to match the real world 1:1.

To do so, developers deploy the robot in the real world and wait for it to mess up.

When it fails, they take that log data and feed it back into the simulator to auto-tune the physics until the simulation perfectly matches the failure they just saw. So we aren’t just simulating the world anymore; we are calibrating a ‘physics accurate’ digital twin in real-time.

Now, imagine what happens when we combine these “digital twins” with Generative World Models (like NVIDIA Cosmos or Google’s Genie), which let us generate entirely new 3D training environments from a simple text prompt.

This will be the take off point, moving us from a world where data is scarce to a world where the only limits are the amount of compute we can throw at the simulation.

(3) New Training Recipes

We now have the right blue print and we’re starting to get more data. The final answer to ‘why now’ is all new training recipes, yielding pipelines that look more like a staged learning curriculum than just a single ‘lesson’.

For most of robotics history, you had to choose between two flawed religions: imitation learning (IL) or reinforcement learning (RL)

IL is the “monkey see, monkey do” doctrine. You’d teleoperate a robot through a task a thousand times, and it would replay a fuzzy statistical version of your motions. This has been safe and predictable, but it only scales linearly with human time/effort.

RL is the faith of trial and error. You give a robot a goal, a reward score, and then let it stumble through the problem until it figures out the rules. This is how AlphaGo shocked the world.

But in robotics, “trial and error” isn’t theoretical. It’s metal slamming into metal; motors burning out, drones flying into walls. And as mentioned, simulation doesn’t yet cover every edge case. It’s also just really hard for most teams to do well (requires a unique combo of dev talent across infra, sim tools, 3D dev, etc).

Historically, the industry was stuck between these two extremes. Today, that dichotomy has dissolved. The frontier isn’t about choosing one religion over the other. It’s about stacking every ingredient—imitation, RL, prediction, simulation—into multi-stage pipeline that take a robot from novice to expert.

Two distinct philosophies have emerged from this blending. Both are producing breakthroughs, but along very different paths.

Path 1: Simulation-Heavy

The first path — championed by Google, DeepMind, TRI, and NVIDIA — leans heavily on simulation and RL.

You start with a pre-trained foundation model (usually a VLA) that already understands objects, scenes, and basic motions.

You then add a small number of human demonstrations to shape this knowledge into a safe behavioral prior, aka a rough draft that keeps the robot from flailing.

With this baseline in place, the robot enters the simulator, i.e the “infinite gym.”

In sim, physics can be bent, lighting randomized, textures swapped, and time accelerated. Failure costs nothing. This is where RL thrives: thousands of attempts per hour refine the rough draft into polished skill, uncovering tiny efficiencies no human ever thinks to demonstrate.

Recent advances add one more layer: foresight. Instead of reacting frame-by-frame like a goldfish, World Models are starting to give robots an internal physics engine, letting them imagine how an action will unfold before committing to it; predicting slips, collisions, or ideal grasp points. Much like an athlete visualizing a race, the robot rehearses futures and picks the best one.

Put it all together, you get a system that understands, refines, and anticipates. But not everyone believes simulation is the best approach…

Path 2: The Real-World-First Curriculum

While simulation is powerful, it’s also expensive, brittle, and often suffers from the “sim-to-real” gap.

In response, companies like Physical Intelligence, Skild AI,, and 1X are proving that you can achieve general-purpose manipulation without heavy RL or sim. That is… if you have enough real-world data.

Similar to the simulation path, this journey begins with a Pretrained World Encoder. Whether it is a Vision-Language Model like CLIP or a VLA like π₀ (Pi-Zero), this gives the robot conceptual understanding before it ever sees a teleoperation demo.

The divergence happens next.

While sim-first teams rely on RL to discover skills, real-world teams bet on Massive Supervised Imitation. This isn’t just a few dozen demos. This is brute force at scale. Physical Intelligence utilizes ~50,000 teleop demos; Covariant leverages millions of manipulation attempts. This creates a massive prior: a raw instinct for how humans interact with the world.

From there, the accelerator is Self-Supervised Learning.

These companies treat every single frame of robot data (i.e. logs) as a lesson, mining camera video, gripper-camera views, failed grasps, and even passive background motion, all teaching the robot physics and causality without needing an explicit reward function.

Finally, once deployed, they use online adaptation, aka ‘learning on the job’ by tweaking the robots tactics in real-time based on sensor feedback. This allows them to handle slips, nudges, and new objects without needing a full RL training loop

Two Paths, One Destination

Despite their difference, both paths are converging toward the same ideal: machines that understand their environment, learn from experience, imagine the future, and refine themselves over time.

The future will likely blend the two. Either way, we are no longer teaching robots to copy us. We are teaching them to learn from us, practice without us, and ultimately, surpass us.

The Final Bottleneck

We’ve covered the brains (VLA), the bodies (diffusion), the fuel (data), and the training recipes (hybrid pipelines).

On paper, this is the formula for Physical AI, and after decades of tinkering, robotics is at an inflection point. But there is a final, unsexy truth standing in the way…

The bottleneck is no longer hardware. It’s not sensors, actuators, or even algorithms.

The bottleneck is infrastructure.

Robotics companies are all trying to do the same thing: build systems that learn from experience, improve with data, and become more capable with every iteration. But the workflow is fundamentally broken.

If you spend a week inside any robotics organization (humanoid, warehouse AMR, manipulator, or AV), you’ll see the same story: a bunch of engineers — some of the brightest minds in AI — spending 80% of their time fighting the cloud rather than solving robotics.

They’re building bespoke infrastructure from scratch, consistently reinventing the wheel, and often stuck in DevOps hell.

What they want instead is a fully automated robotics data flywheel — a system where robot logs ingest themselves, simulators spin up on demand, training datasets auto-curate, training pipelines retrigger on new edge cases, and updated models roll out safely across the fleet.

This is the problem space we’ll explore in Part II. Until then, enjoy these robot bloopers — a great reminder that until we can operate data fly wheels with ease, this is the extent of most robot careers: extremely expensive slapstick comedians.

Spotify Podcast Link

What if machines could learn the world the way humans do?

Not by crunching data and predicting tokens, but by understanding the world and how it works?

Imagine a robot that doesn’t just see a chair, but understands it— its purpose, its relationship to the table, the way you might pull it closer before sitting down.

That’s the world Matt Miesnieks is building.

Matt is the CEO of Primate.AI, a new kind of physical AI company rethinking how machines perceive and interact with reality. His team is exploring what happens when perception stops being a math problem and starts becoming an act of intuition.

Before founding Primate, Matt helped shape the modern era of spatial computing. He co-founded 6D.AI, the company that pioneered the AR Cloud—a real-time 3D map of the world that made digital content persistent in physical space. 6D.AI was later acquired by Niantic, creators of Pokémon Go, where its technology now powers their next generation of AR experiences.

Now, with Primate.AI, Matt’s taking his biggest swing yet: building machines that perceive reality the way we do—not through pixels, but through concepts, memory, and innate understanding.

If 6D.AI gave AR its eyes, Primate.AI is giving machines a mind.

In this conversation, we explore the future of embodied intelligence, what comes after language models, and why understanding the world like a human may be the only real path to AGI.

Spotify Podcast Link

Today’s guests are building one of the most ambitious startups in the entire spatial computing industry.

The company is called RP1, and we’re sitting down with co-founders Sean Mann and Dean Abramson.

On its face, RP1 is the world’s first true metaverse browser, trying to do for spatial computing what browsers like Mosaic and Internet Explorer did for the web.

But that’s just the tip of the iceberg.

Beneath the surface, RP1 is so much more... They’re building all the infrastructure and protocols required for a truly spatial internet—one where immersive, persistent digital worlds stream on demand, where physical and virtual environments seamlessly merge, and where millions of users can interact in one, shared world, not a bunch of silo’d virtual experiences.

In this episode, we dive into what RP1 actually is, how it scales to millions of users, and why they believe app stores must die.

We also unpack their vision for the core backbone of this future, a Spatial Fabric—a kind of real-time, distributed protocol for the metaverse that manages everything from identity and maps to avatars and live services.

This one’s technical, philosophical, and wildly ambitious. Let’s get into it.

Spotify Podcast Link

This interview is with Alvin Graylin — one of the boldest voices out there today on AI, reimagining civilization in the age of intelligence.

Alvin is the best-selling author of the book, Our Next Reality, which is a super thought provoking exploration of how immersive tech and artificial intelligence are going to reshape every layer of society.

Prior, Alvin spent over a decade on the front lines of the XR industry—building and leading HTC’s VR business in China, and serving as Chairman of the Virtual World Society.

Today, Alvin's on a new mission: to help humanity navigate the AI transition—researching, writing, and evangelizing ways to ensure AI serves a greater good and unlocks an age of abundance.

And While Alvin is an optimist, he’s also a realist—and he’s clear that this transition won’t be smooth unless we get a few critical things right.

We talk about…

1. The real risks of AGI—not just killer robots, but unaligned incentives, unchecked power, and job displacement.

2. His call for an “AI Bill of Rights,” and why the U.S. and China must shift from competition to cooperation—perhaps through what he calls a CERN for AI. (worth checking out the link to AI Bill of Rights here)

• And finally, we explore XR—not as a form of escape, but as a tool for empathy, education, and rewiring human behavior at scale.

This conversation is expansive, urgent, and radical. Let’s get into it. Enjoy.

On the dusty plains of Idaho, hundreds of scientists erupt in celebration; hugging, clapping, cheering.

They just had a breakthrough and their vision was vindicated. A vision that — even by today's standards — was entirely science fiction.

It was also polarizing. Because when we split the atom, we also split the world…

For some, nuclear fission was the key to human flourishing. For most, it was the ultimate weapon.

But the optimists in Idaho held their ground. They imagined a world in which energy was no longer a constraint. Where abundance wasn't theoretical. Where humanity's growth wasn't tied to a negative impact on the earth.

These dreamers weren't just building nuclear reactors. They were ushering in a new epoch in human history: the first atomic age. And their progress was legendary.

The Idaho National Lab (INL), under the purview of the Department of Energy, went on to produce 52 nuclear test reactors. Along the way, a singular debate echoed through its halls: when would nuclear power 100% of the earth's energy?

The scientists started placing bets, and at the time, most agreed — it'd be around the year 2000.

Needless to say, that's not what happened.

Nuclear remained misunderstood and Idaho National Lab became a shell of its former self.

The Second Atomic Age

The year is 2025 and those dusty plains are stirring, yet again.

Back at Idaho National Lab, the old dream is being reborn, but this time, the headwinds of the past have become the tailwinds of the future.

Regulatory bodies are creating modern frameworks for new reactors. Private capital, once allergic to nuclear energy, is now flowing in. And at long last, the technology is mature, safe, and ready.

A nuclear renaissance is underway and the race is on, with multiple startups setting their sails; each one taking a unique tack.

There’s Kairos Power, who just signed an MOU with Google to deliver 500MW of nuclear power for AI data centers via a small modular reactor (SMR).

There’s Radiant Nuclear, with $225M in funding for its ‘nuclear reactor in a box’, aka a micro modular reactor (MMR).

And then there’s the buzzy, Wall Street darling Oklo, who went public via a Sam Altman owned SPAC in 2024, valuing the company at $400M.

Oklo has since surged over 500% to a $10B+ valuation…

Oklo’s next big milestone? A prototype of a small, configurable nuclear power plant at Idaho National Lab (INL). They’re soon to break ground and the markets are watching with bated breath.

But Oklo isn’t the only game in town.

Less than half a mile away, a newer company has partnered with the INL and just secured its own plot of land; to build not just a prototype, but to invent an entirely new class of nuclear energy product.

It’s not a Small Modular Reactor (SMR). It’s not a Micro Modular Reactor (MMR). And it's not your traditional nuclear power plant. It’s something entirely new, filling multiple gaps between all three.

The company is called Aalo, and despite a later start and fewer resources, they’re vaulting ahead on all major fronts.

The rest of this essay is a deep dive into how and why.

Introducing Aalo

Relative to its competitors, Aalo has moved at a blistering pace.

Since launching in 2023, they’ve moved into a 40,000 sq ft factory space, built a non-nuclear prototype, received permission to pursue Department of Energy (DOE) authorization for their first reactor, secured land to build it at Idaho National Lab, and signed non-binding agreements with some of the world’s largest possible customers (most confidential, for now…)

This all just culminated in their two biggest milestones to date: First, a $100M Series B led by Valor Equity Partners, the world-class hard tech investor behind SpaceX, Tesla, and Anduril.

Second, Aalo’s selection by the U.S. Department of Energy as one of eleven companies in President Trump’s Nuclear Reactor Pilot Program—an initiative aimed at building and operating at least three advanced test reactors by July 4, 2026. This places them on the fast-track of America’s nuclear renaissance, with direct federal support to prove their design in record time and a spotlight as one of the few companies shaping the future of clean baseload power

During this time, Aalo benefited from those industry-wide tailwinds across technology, financing, and regulation. But, when they first launched, the market remained stagnant and speculative. It was still missing a key ingredient: a commercial catalyst, aka the right buyer with enough urgency, capital, and political power to pull projects through.

Fortunately, in an almost kismet fashion, that buyer has suddenly arrived: our new digital companions — artificial intelligence.

As AI surges ahead, energy has become the most critical layer of the AI stack.

No doubt, AI offers endless promise and hope, from growing out of a debt spiral, to curing diseases, to modeling climate change, to redesigning our material world.

But its future depends on how many electrons we can push through a wire. And this digital brain has a voracious, 24/7 appetite.

To give you a sense, the typical data center uses around 25-50 MW of energy, which is enough to power a small town. Over the last 10 years, hyperscalers have required around 100 - 300 MW.

But now with AI? They’re requiring 10-20X that… from 1-2 GW. Hence Elon Musk’s recent purchase of a 2GW gas-turbine power plant for Collosus: the 200k GPU cluster behind xAI.

To put that in perspective, 2GW is enough to power the electricity demand for nearly 40% of New York City and 50% of Los Angeles…

When you drag out this level of demand, AI is on pace to consume the majority of the US electrical grid by 2030.

And to think… AI is just one of many looming sources of enormous demand.

Include the electricity needs for EV charging, robotics, manufacturing, growing cities, the military, etc. and there’s no doubt… The future is arriving faster than we can power it and our aging grid — an ancient system that hasn’t effectively changed in 100+ years — can’t keep up.

Fortunately, the energy market isn’t waiting around. A new paradigm is emerging, characterized as ‘bring your own power’, with customers planning to leap frog the grid entirely.

This is Aalo’s opportunity and the kernel of their strategy.

But to fully appreciate it, we must first set context with the state of nuclear energy and Aalo’s origin story.

Not Your Grandpa’s Nuclear

Over the last decade, nuclear energy has quietly become the safest, most reliable, and most power-dense source of clean energy ever created.

Just a single, 1-inch pellet of uranium can replace a ton of coal.

And a single barrel of uranium? It can replace two million barrels of oil/gas.

At scale, for reliable base-load power, nothing else comes close. Which is why nuclear is the answer, not just to climate change, but to many of the world's most complex challenges; from rising costs, to food production, to water scarcity, to geopolitical tension; all of which boil down to a lack of cheap, accessible energy.

Aalo CEO, Matt Loszak, locked on to this problem space back in 2021. After building a successful HR software business (to 150 employees and 4,000+ customers), his roots in physics and engineering took hold; he had mastered the world of bits; it was time to conquer atoms. So as nuclear began to accelerate, Matt began to build.

With a degree in physics, Matt knew enough about nuclear to be dangerous. But he needed to gain an edge. He also needed a technical co-founder. So Matt set off on a global tour, meeting with the most renowned nuclear experts in the world. 100’s of them; gaining knowledge, looking for patterns, and seeking a partner in arms.

One of those experts was Yasir Arafat, the Chief Architect and project lead for one of the most significant and yet underrated nuclear power projects in the US: MARVEL.

MARVEL is a government-funded microreactor at Idaho National Lab – the first new reactor scheduled for construction. It’s also a test-bed built to unlock the next generation of nuclear innovation.

For decades, startups have been stuck in a catch-22: you need operating data to get licensed, but you need a license to collect the data.

MARVEL breaks this cycle, offering a safe, government-backed environment to test components, validate models, and show how advanced reactors connect to real systems like microgrids; all while providing the hard data required by regulators and customers.

For startups like Aalo, MARVEL is a force multiplier, de-risking their tech, speeding up deployment, and signaling that nuclear power is moving forward, yet again.

Upon learning about MARVEL, Matt decided to go all in. His first move?

Recruiting Yasir as his co-founder.

It’s hard to draw up a better profile for a nuclear energy CTO.

Yasir led MARVEL to achieve something no one else had done before: the first-ever DOE authorization for construction of a nuclear reactor.

Prior, he led the eVinci microreactor program at Westinghouse — one of the most influential names in the industry, responsible for the world’s first commercial pressurized water reactor (PWR), a design that still powers over half the world’s nuclear fleet today. It’s also behind the AP1000, one of the safest, most advanced reactors ever built.

In other words, Westinghouse defined nuclear’s past. Now, Matt and Yasir are setting off to define its future, with multiple rock stars joining in the mission, including Richard Williams as Director of Engineering (MARVEL’s former Director of Engineering), Jon Guidroz as SVP of Commercial (former Chief Strategy Officer for Energy at Microsoft), Scott Reynolds as VP Finance (former global head of structured finance at Bloom Energy) and Bryson Gentile as VP of Manufacturing (previously Head of Manufacturing at SpaceX on Falcon 9).

The X Factor

This is an A-List team. But nuclear energy is a modern day gold rush. The world’s best & brightest are flocking to win the prize. Top talent is table stakes.

What differentiates Aalo most is their product strategy. Particularly, their plan for achieving product market fit.

If I had to pick one deciding factor, this is where the race will be won or lost.

That may sound obvious. But the nuclear market is unique. The value chain is in flux and the ultimate buyer isn’t clearly defined— is it the power utility? The data center developer? The co-lo operator? The cloud hyperscaler? All of the above?

Despite this uncertainty, nuclear startups have to pick a design path and bet on a demand signal early.

Aalo’s bet is both clever and sound.

First, they’re working backwards from the most urgent and well-capitalized source of demand; AI hyperscalers (the Microsoft’s, Google's, Meta's, and Amazon’s).

Second, they’re designing for multiple potential buyers, as hyperscalers don’t operate in isolation, but within a complex ecosystem of developers, utilities, and partners.

Which raises an existential question: should Aalo be a power developer or just a reactor OEM? i.e. build & operate the full plant? Or just sell the core primitive to builders/operators?

The reality (and the fun part) is they have to be both, existing in a kind of quantum state, ready to collapse the wave function at the precise moment the market demands clarity.

So the prevailing business question isn’t simply, “What does the customer want?”

It’s “How do we meet demand that is still deciding how to express itself?”

Aalo's answer is two fold:

First, stand on the shoulders of giants.

In partnership with Idaho National Lab, they’re designing an entirely new reactor platform, inspired by learnings from MARVEL, EBR-II, AP-1000, and other well known historical reactors.

This approach helps de-risk the technology, smooth the regulatory path, and build confidence with future customers. And while Aalo-1 is inspired by these past reactors, it adds new engineering innovations, leveraging modern innovations in materials science and more, to make the design uniquely well suited for AI data centers. At the same time, it keeps the proven ingredients, e.g. a proven fuel supply chain with UO2 fuel type, a coolant with deep operational history, well proven safety features, and modern simulation tools, giving regulators and partners much needed comfort and conviction.

The second part of their answer — offer an a la carte menu, from Lego blocks to a full plug-in-play system, aka a nuclear factory-in-a-box.

This approach mirrors NVIDIA’s. While known for GPUs, NVIDIA also sells complete AI server racks, along with CPUs, networking gear, and cooling modules. Their customers can choose to buy a full-on ‘AI factory’, with multiple networked racks, or they can integrate individual parts into existing infrastructure.

Aalo’s offering is similar, and the first of its kind. It’s modular enough for multiple buyers, robust enough to scale, and flexible enough to serve the ‘Bring Your Own Power’ value chain emerging around AI hyperscalers.

It’s also a classic 'blue ocean strategy': avoid a crowded, legacy market (red ocean) by creating an entirely new one. I’m calling this uncontested space ‘the missing middle’.

The Missing Middle

The advanced reactor industry offers two primary options, neither of which is the ideal fit for data centers.

On one end there’s Small Modular Reactors (SMRs).

These are shrunken versions of traditional grid-scale plants, designed for utilities years before the AI boom. While promising modularity, they still follow the old playbook: large projects, long timelines, and high costs per megawatt.

For a hyperscaler who needs power in months, not years, SMRs are too big, too slow, and too tethered to the legacy energy model.

On the other end are Micro Modular Reactors (MMRs).

These are tiny, ~1-megawatt reactors designed for niche, off-grid uses. While highly portable and safe, they are a squirt gun for a wildfire; their power output is too small and their cost per megawatt is too high to be a scalable solution for AI's immense energy needs.

In hindsight, this product gap might seem obvious. But until today, filling it just didn't make commercial sense. A market didn't exist.

Now, an immense opportunity has emerged, to match the right product architecture to the data center's acute, time-sensitive demands.

Subscribe now

Aalo's Answer: The XMR

To seize this opportunity, Aalo is inventing an entirely new category: the XMR (Extra Modular Reactor). The "X" stands for the extra flexibility and adaptability, or offering the X-Factor for AI data centers by designing around their unique needs.

Their core product is called the Aalo Pod, a 50-megawatt power plant composed of five 10-megawatt, sodium-cooled reactors; powerful enough for a small city or medium data center and composable enough for gigawatt scale.

It’s designed around three main vectors:

Optimal Size and Scalability: The Pod delivers the significant power of an SMR but is built with the manufacturing efficiency of a microreactor. This allows data centers to scale power in 50 MW increments, mirroring how they build out their campuses.

Data Center Optimization: With five independent reactors providing N+1 redundancy, the Pod offers the extreme reliability required for mission-critical AI workloads. Its compact, 2.5-acre footprint and air-cooled design means it can be co-located directly on-site, even in arid regions, eliminating reliance on the fragile grid.

Speed and De-Risking: By using proven UO2 and low-enriched uranium (LEU+) fuel instead of waiting for novel fuel types, Aalo sidesteps one of the biggest supply chain and regulatory bottlenecks in the industry. This enables a target deployment cycle of just 52 weeks, from order to operation (a marvel in its own right, which we’ll cover later).

This is energy's version of an AI server rack. After a rapid on-site assembly, you drop it in, turn it on, and voila; the AI factory is humming.

Elegant, indeed, but the real beauty lies in its composability and a hierarchy of innovation that de-risks the technology at every stage.

The fundamental building block of this system is the Aalo-1, the individual 10 MWe reactor module. This is the core technology: a sodium-cooled thermal-spectrum reactor designed for safety and mass manufacturing.

By using liquid sodium as a coolant, it operates at near-atmospheric pressure; a key design choice that eliminates the risk of the high-pressure meltdown events associated with traditional water-cooled reactors.

Each Aalo-1 is a self-contained, factory-built unit. It’s the essential component that makes the entire Pod concept possible.

Before a single Aalo-1 is deployed, however, there is the Aalo-0. This is Aalo's non-nuclear, electrically-heated prototype. It is a full-scale physical twin of the Aalo-1, allowing the team to test, validate, and perfect every aspect of the thermal hydraulics and system performance without touching nuclear fuel.

This "test before you build" approach is critical— it proves the physics and de-risks the technology for regulators and investors, demonstrating a mature and responsible path to market.

The bridge from factory to field is Aalo-X, the first private advanced reactor demonstration being built in partnership with the U.S. Department of Energy at Idaho National Lab. This is the regulator and technical proving ground that unlocks commercial deployment.

This plan is rooted in clear logic: start with the customer problem, deliver a standardized product with the Pod, build it on proven core technology with the Aalo-1, and validate fast to de-risk future deployments with the Aalo-X.

It's also shaped for the urgent timelines of the AI era, while remaining sharp enough to slice through regulatory red tape and stigma.

From Strategy to Reality

To realize its potential, Aalo must compound this momentum across every part of the business, starting with its two strongest points of leverage: a factory that learns and a delivery model that radically compresses timelines.

The factory is inspired by playbooks from Tesla and SpaceX. It’ll begin as a lean “minimum viable factory” to prototype fast. They’ll then certify it to the NQA-1 spec: the gold bar for manufacturing safety and reliability. From there, it will scale into a gigafactory—flipping nuclear energy from decade-long one-offs into a repeatable machine where costs fall, risks shrink, and power rolls off the line like clockwork.

The magic behind it all is a digital twin: a full-stack simulation of the reactor, the factory, and the supply chain. This is not just a virtual model. It’s an orchestration layer; one that can spot bottlenecks, reschedule in real-time, and preempt delays before they surface.

This turns Aalo’s facility from a set of machines into a self-improving organism that can be cloned and deployed anywhere in the world.

Then we have the delivery. This is where Aalo bends reality, starting with a company-wide mantra: “order-to-electrons in 52 weeks”.

Contrast this with nuclear's current decade-long fiascos, and it’s not just a timeline. It’s the kind of maniacal mission behind the greatest companies in the world.

The key is treating deployment like a software-driven manufacturing workflow.

Aalo’s “order-to-electrons” clock starts the moment a customer signs. Three tracks begin in parallel: contracts and financing, AI-assisted regulatory filings, and factory production of reactor modules.

By the time site licensing is approved—roughly 24 weeks in—most of the system is already built.

With over 90% of the work done in the factory, the on-site labor is minimal. Modules arrive pre-fabricated and ready to slot into place. From there, there’s ten weeks of local assembly; followed by three months of testing, fuel loading, and final validation; and then… Power on.

The Path Forward

To go from this vision to electrons, Aalo has countless milestones to hit. But here are the most important:

In late 2025, they will break ground on the Aalo-X demonstration plant at Idaho National Lab and work toward gaining regulatory approval via the Department of Energy. This phase also provides data to fast-track an NRC application (i.e. Nuclear Regulatory Commission).

In 2026, they’ll achieve zero power criticality on Aalo-X. This means the reactor has achieved a self-sustaining nuclear chain reaction at an extremely low power output—essentially, near zero thermal power; a crucial step before the reactor begins producing significant heat or electricity.

In 2027, they’ll achieve full power at Aalo-X, which will validate the entire model and propel it towards commercial deployment in 2028, including NRC approval and construction of the first commercial Aalo Pod for a data center customer.

By 2029, the first commercial reactor will be powering a data center at full clip, setting the stage for a major financing event or IPO and a rapid, nationwide scale-up throughout the 2030s.

All of these objectives are geared towards a singular long-term goal: reducing the cost of electricity down to $0.03 per kilowatt hour.

In the US, the average cost is around $0.16 - $0.17, and rising fast, constraining countless plans to build, innovate, and grow. But if Aalo succeeds, those sci-fi dreams from the plains of Idaho become reality.

Imagine… the year is 2055. Electricity flows like water, powering not just homes and industries, but all new frontiers of human ambition.

Deserts bloom with desalinated water, city-based vertical farms feed millions, automated factories hum continuously, remote villages enter modernity, and reactors light up bases on the Moon and Mars. While back home… The air is cleaner, the climate is steadier, and prosperity is more accessible for all.

Yet abundant energy isn't just about raising the floor, it's about shattering the ceiling.

This starts with unleashing the potential of our most powerful new tool — artificial intelligence — which is woefully tethered to a short electrical leash.

While abundant energy won’t turn AI into a magic wand, it will allow the machines to run the trillions of simulations required to find a new drug, to model a planet’s climate with startling accuracy, or to design a material that was previously impossible.

It also lets us ask bigger questions; questions about our purpose, our origins, our destiny; because we can finally afford the computational cost of finding the answers.

This future will not be perfect. We will still be wrestling with our all-too-human limitations. But by removing the endless, exhausting struggle for energy, we are freed up to focus on everything else.

And while nuclear won’t solve all our problems. It will make solving the grandest of them profoundly more plausible; expanding the realm of the possible, and reminding us that the only true, lasting constraint on our future is what we dare to imagine next.

Which is why I’ve concluded — the world does remain split, but in an entirely different way: those who love nuclear energy, and those who just haven't looked closely enough.

Full transparency: I’m an investor in Aalo. But this essay came first. As I was writing and interviewing the team, my conviction grew — and when an allocation opened up in their Series B, I jumped all over it. Maybe it was luck, maybe it was timing, or maybe stories really do shape the future…

Spotify Podcast Link

Today’s guest is a legend within spatial computing. We're speaking with Philip Rosedale, one of our spaces most renowned pioneers and the visionary founder of Second Life.

Long before the 'metaverse' and 'web3' became buzz words, Second Life was the first, commercially successful, virtual world, with a self-sustaining, digital society where millions could create, trade, and become whoever they wanted to be.

Philip does a lot of podcasts, so I tried to make this one unique.

Call me biased, but mission accomplished. This is easily my favorite conversation of the year.

We explore the enduring magic of Second Life, what it taught him about economics, identity, and why adults—not just kids—still crave virtual connection.

We dive into the deep stuff—how avatars act as mirrors of the self, why virtual identity can actually reshape real-world behavior, and how platforms like Second Life may hold the key to understanding consciousness itself. We talk about the illusion of self, the Proteus Effect, and what it means to co-create identity in a shared digital world.

And we ask: what happens when AI agents start growing up in these spaces? Could virtual worlds become safe playgrounds for emerging machine minds—and what might they teach us in return?

We then get philosophical, in all the best ways. We explore:

• Avatars as mirrors of the self

• How virtual identity can reshape real-world behavior

• What happens when AI agents start growing up in these spaces? Could virtual worlds become safe playgrounds for emerging machine minds—and what might they teach us in return?

So, definitely strap in for this one and enjoy, its wonderfully wide-ranging, it's mind bending, and it's most importantly, full of soul...

With that, I bring you Philip Rosedale.

If you're just joining us — welcome! Excited to have you on this mind erasing bending journey.

That said, if you want this essay to truly land, highly suggest starting at the beginning with Part I. Links to each essay below:

• Part I

• Part II

• Part III

We've traveled far.

From the origins of this illusion, to wide-eyed babies, to how/why you 'have no head', to empirical evidence exposing the glitch in the Matrix.

The key takeaways?

You are not the thinker of your thoughts. You are not the pilot behind your eyes. You are not a permanent self riding around in a biological vehicle.

You're an ever adapting story, told in real time by a nervous system trying to keep you alive.

This might land intellectually. But if you're like me... it doesn't matter. Experientially, the illusion remains. And the more I poke at it, the feeling of self doesn't just remain... It rebels.

As if under attack, a self-defense mechanism is awakened. The whispers turn into screams,

"I'm real damnit! And I'm in control... All those desires? Success, money, power, sex.. we will and shall have it all. Onward!"

*sigh...

So what the hell is one to do? It all seems so... futile.

My first thought is to give up and just keep living my life and being... my self.

But then I see the progress of AI. I see how it's merging with augmented/virtual reality. I see how our sense of self — bombarded by virtual influence and mimetic desire — is soon to be swallowed whole, chewed up, and spit out.

I then turn back to my newborn daughter, Chloe. Her eyes are still too young to focus. But sometimes, her eyes lock onto mine. Or at least, they appear to... and in that moment of connection, I forget myself.

Not in the poetic sense. I mean literally: the “me” doing the watching fades. No inner narrator. No ego analyzing the moment for meaning or metrics. Just sensation. Soft skin. Warm breath. An open field of aliveness.

It lasts a few seconds, at most. Then my mind snaps back into its default groove— remembering all the things it wants to be. The roles. The performance. The problems to solve.

But that glimpse is enough. Enough to know that something deeper is possible. And not just possible—but necessary, especially in the face of exponential tech + change.

So what is that thing? And how do we hold onto it in spite of our inner tyrant?

We need a new perspective. A new mental model for reality; for what experience really is (and isn't).

And to achieve this, we need new tools. Or as I like to say, we need to use tech to combat tech. A form of jiu jitsu, if you will... using tech's momentum against itself.

Towards that end, we won't negate the self head-on. We'll first zoom out and move up stream of subjective experience.

We have to look at our environment and our relationship with it. Only then can we answer the ultimate question: if we're not a 'self', what are we? Because a mere 'story' feels... unsatisfying.

The answer is multi-layered. This essay will peel back each and expose the kernel of what makes the human experience so unique.

What I've found, is that the deeply human is paradoxical. It suggests we are beyond being just merely human...

Persons vs. Selves

So what are we?

The most direct answer I’ve found comes from modern day philosopher Jay Garfield.

In his his book, Loosing Ourselves, Garfield argues that we’re not fixed “selves” floating above the world—we’re “persons,” embedded in it.

What’s the difference?

A self is imagined as independent, essential, and constant. It's somehow... pre-existing, or prior to.

A person is none of those things. A person is a process—a pattern—a psycho-physical dance that only exists because of its entanglements with everything else; our biology, our social interactions, and our ongoing experiences.

The common metaphor in Buddhism is a wave on the ocean. A 'self' is the mistaken belief that you are a separate, independent drop of water, riding the wave.

A 'person' is the wave itself—a temporary form that is inseparable from the ocean it emerges from, shaped by the winds of culture and the tides of relationships.

As a pianist, I prefer the analogy of music: a person is more like a song—it only exists while it's being played. It is the interplay of notes and the instrument, not some hidden essence behind them.

In other words, we don't pre-exist and then decide to interact with the world; we emerge from it. We are a messy, dynamic, and beautiful tangle of our genetics, the language we speak, the culture we inhabit, and the people we love.

This isn't just a semantic game. It's a fundamental shift in our operating system. And this new OS matters, because the isolated 'self' is a fragile target for this next wave of tech/AI; easily hijacked by algorithms.

But a person?

A person is (as Bruce Lee would say) like water. It's fluid, adaptive, and finds stability not in separation, but in connection, with nature and with other people... Real people.

Spotlight vs Floodlight Consciousness

To further ground this idea, I turn to Alan Watts—a thinker who, in my view, revealed some of the deepest truths about modern affliction.

If Garfield gives us the academic language of “persons,” Watts gives us the poetic, spiritual framework for where we fit in. He tackles the same illusion but zooms out—from the mind to the cosmos.

In this lecture The Myth of Myself, Watts argues that our core delusion is the idea of the “skin-encapsulated ego”—that little ‘me’ we imagine trapped inside our heads, peering out at a foreign world. This hallucination splits “self” from “everything else,” casting life as a battle between a lonely subject and a cold, indifferent universe.

We feel like strangers here—confronting reality, facing facts, trying to conquer nature. Poet A.E. Housman said it best: “I, a stranger and afraid, in a world I never made.”

Watts traced this alienation to what he calls "two myths of the West".

First, is the myth of the universe as an artifact, i.e. a pot crafted by a cosmic potter, aka 'god'.

This is the religious view. And it makes us feel perpetually judged, like we’re on divine probation, perpetually proving ourselves worthy of existence (in this world or the next).

Then came the scientific view. This produced the myth of the universe as a cosmic fluke, i.e. an aimless probability machine, powered by blind energy. This view makes us feel like a meaningless accident.

One makes us sinners, the other specks. Both leave us feeling unfulfilled and like we don’t belong.

To rectify, Watts offers a third story. He argues that the universe isn’t a top-down creation, nor is it a bottom-up accident. Rather, it’s an intelligent, self-expressing organism.

And its form expression?

You, me, us.

We aren’t a fluke of the universe; we are a feature of it. Watts famously says, just as an apple tree “apples,” the cosmos “peoples.” We didn’t arrive in this world—we emerged from it. Like a leaf from a tree, we are functions of the universe, not anomalies within it.

So where Garfield’s “person” emerges from a social and biological context; Watts’ “you” emerges from the entire cosmos. It’s the same idea, just at a different resolution.

We are what the whole cosmos is doing at the place we call here and now. Each one of us is a unique "waving" of the entire universe.

But if this is true, why do we feel like a lonely pilot in the skull? Why the profound sense of separation? Of subject and object?

Watt's creates a bridge between cosmic truth and psychological illusion. He explains that our feeling of being a separate ego is a direct result of specializing in one particular form of consciousness, while ignoring another.

Those two forms are "the spotlight and the floodlight".

The spotlight is conscious attention. It’s the narrow, focused beam we’re trained from childhood to value above all else. When a teacher says, “Pay attention!” we fix our minds on one thing at a time. Our culture teaches us this capacity matters most. And so we believe, “I am my spotlight consciousness. That is my ego. That is me.”

But floodlight consciousness is always on. It is the vast, peripheral, holistic awareness that drives the car for miles while your spotlight is busy chatting with a friend. The floodlight is what beats your heart, grows your hair, and registers the entire field of sound and space around you without you “thinking” about it. Every nerve ending is its instrument.

The illusion of being a self is simply the habit of mistaking 'spotlight consciousness' for the whole show.

We’ve forgotten that we are also the floodlight. The floodlight is the waving cosmos, the entire field of being. The spotlight is just where the floodlight—the universe itself—has decided to focus its attention through you, for a mere moment.

Becoming aware of this is the mystical experience you hear about through meditation, psychedelics, and even states of flow. In these moments, you realize your real, deep self is not the fleeting beam of attention, but the entire field of awareness it moves through.

As Watt's states in his talk, you are not the headlight, unaware of the battery it’s wired to. You are the whole circuit.

Realizing this dissolves the ultimate fear... The fear of death. An isolated spotlight can be switched off. But the floodlight, the great field of being, simply transforms.

The wave returns to the ocean it was never separate from.

As Watts would say, you are the eternal thing that goes and comes. One day, you wake up, look out at the stars and realize, “Why, that’s me.”

Subscribe now

Neuro-Tech Buddhism

This might all sound a bit too hippy dippy woo woo for you.

It might also sound too abstract, if not entirely out of reach.

I felt the same way. These ideas tickled something deep in my being, but they remained slippery. And I struggled to see how to make them land for average Jane/Joe.

And then I came across a quote from one of my favorite authors, Ursula Le Guinn.

It echoes the same problem called out by Watts: the divide between humans and nature, and the resulting alienation.

Le Guin says:

"One way to stop seeing trees, or rivers, or hills, only as ‘natural resources,’ is to class them as fellow beings—kinfolk. Subjectify them, instead of objectifying them. I guess I am trying to subjectify the universe, because look at where objectifying it has got us. To subjectify is not necessarily to co-opt, colonize, exploit. It may involve a great reach outward of the mind and imagination."

A great reach outward... of the mind and imagination.

Well, if there's one tool capable 'subjectifying, and extending the mind... it's our favorite topic de jour: spatial computing, aka virtual reality, augmented reality, mixed reality, etc.

This brings us full circle, back to modern day philosopher, Thomas Metzinger.

If Le Guin calls for a new kind of perception—one rooted in kinship, not conquest. Metzinger suggests a way to build it, with spatial computing + AI.

I've explored similar use cases in my previous essays.

The first was the Ultimate Promise of the Metaverse back in 2021:

I think the key to mental resilience and joy are regular reminders of everything that exceeds us. Things that disintegrate the ego and remind us of the smallness of our everyday lives. Things that remind us we are mere dust in the cosmos, made of the stars themselves. That we are, as Carl Sagan said, “a way for the cosmos to know itself”. We need daily doses of something similar to the famous ‘overview effect’ that astronauts express when viewing the earth from space. What if the Metaverse could be littered with ‘overview effects’? On-demand experiences that would allow us to feel a connection to the grandness of the universe and the miracle of our own consciousness?”

I continued this thread in ‘The Ultimate Promise of Spatial Computing’, exploring the science of awe and its capacity to rewire us from the inside out:

“The ultimate conclusion? Awe shifts our focus from the self to the vastness of the external world, creating a sense of connection and well-being, while promoting greater generosity and cooperation. Towards this end, the researchers explore simple daily activities, such as taking an "awe walk" and intentionally noticing beauty in the world. Nature-based 'awe walks' will and should always be our first option. But 'awe on demand' via 'spatial/salient computing' could become the next best thing, with arguably more magic and tastefully curated forms of novelty, curiosity, and surprise. While virtually induced, I don't view these transcendent mind-states as disconnected from reality. They just might be the most real experiences we can have…”

Metzinger holds similar beliefs, viewing immersive tech as a way to hack our inner tech and reveal truths about the self; a kind of flight simulator for consciousness.

If you recall from Part I, Metzinger’s big idea is the "Phenomenal Self-Model"—the internal avatar your brain stitches together from a constant flood of signals: sights, sounds, memories, and bodily sensations.

Because that avatar is a construction, AR/VR can hijack and remix those signals, revealing the seams. A well-designed simulation, he argues, could grant ordinary people the kind of ego-softening insight that monks and mystics have pursued for millennia, but with a crucial twist: under controlled, reversible, lab-grade conditions.

He imagines a variety of tools for playing with the source code of identity:

A Dial for the Self: A virtual slider that lets you increase your sense of agency, until the entire scene feels like "you," then fade it out until you are nothing but a pure witness to the world.

Embodiment Hacks: The ability to body-swap into an avatar with elongated limbs, an animal’s form, or a different gender, showing just how effortlessly the feeling of "me" migrates to a new home.

AI Hijacking: You control a virtual body, but occasionally, an AI agent takes over your limbs. An action happens, but you didn't do it. The clean line between your intention and the world's action begins to blur.

Reality Faders: A simulation where the world’s textures slowly morph from photorealistic to overtly synthetic, allowing you to feel the precise moment your brain’s faith in an “outer” reality begins to crack.

The goal here isn't just a joyride. In pairing these experiences with brain-imaging technology, we can turn VR suites into engineering labs for the mind.

We could finally map how agency, body-ownership, and perspective click together—and potentially learn how to therapeutically loosen their grip in conditions like PTSD, depression, or chronic pain.

But like all frontier tech, this can cut both ways.

These tools would be like a digital psychedelic: potent, reality-bending, and profoundly double-edged. Without strict, open-sourced ethical guidelines, it could easily become a black-box for persuasion and addiction.

We would need to figure out the right measures of success/impact. Some way of separating a 'trippy novelty' from what ancient Greek's called 'eudaimonia', aka our capacity for well-being, flourishing, and realizing one's full potential.

If so, and if designed with wisdom, tomorrow’s headset may one day sit beside the meditation cushion as a core tool for waking up, proving that the wildest frontier isn’t outer space, but the inner cosmos of the mind itself.

Epilogue

Many people will resist these use cases and this future.

But the truth is… it’s inevitable. The train's far down the tracks. These simulations will exist, whether we like it or not.

The key will be to never forget: they’re still simulations; a means, not an end. A compass, not a destination. Perhaps technology can help here as well.

Regardless, the most powerful reminders of our true nature won’t come from code. They’ll come from something softer. Warmer. Human.

For me, that's now a baby, blinking up from her crib.

As I write these final words, Chloe sleeps peacefully, blissfully unaware that she's already living the truth I've spent thousands of words trying to uncover.

She doesn't yet know she's supposed to feel separate. She hasn't learned to hallucinate a solid self. She simply is.

Which makes me wonder...

Maybe the real revolution isn't in the technologies we'll build, but in remembering what we were before we forgot what we are.

Maybe the future of consciousness isn't about transcending our humanity, but about relaxing into it more completely.

Because as we've learned, the deeply human is yet another paradox. It shows that we are beyond being merely human.

As such, the illusion of self isn't a problem to be solved, but a game to be played more consciously—with more humor, more lightness, more love, and certainly... way less anxiety.

After all, if the self is just a feeling, and there's no one here to take credit for this essay, then there's also no one here to be worried about getting it wrong.

And that, perhaps, is the most liberating feeling of all.

THE END

If you made it this far… THANK YOU for making the time + effort.

I hope you enjoyed the journey and feel more prepared for this looming era of extreme change. If so, please consider sharing this essay with a friend or two. The more self-less 'persons’ out there the better ;)

In Part I, we defined the 'Self' and explored its origins.

In Part II, we explained the illusion and how to grasp it (by studying babies and realizing you 'have no head').

Here in Part III, we're taking a more empirical approach. It will make you question… well, just about everything.

If you haven't read Part I & II, highly recommend doing so, else this essay might not land 🙂

• Part I link here

• Part II link here

But I get it. Time is money and money is time, so here's an all-too-short recap: the self you think you are, that pilot behind your eyes, isn't what you think it is.

It's not a persistent, stable entity (e.g. a soul, an independent 'you' that exists across time & space). It's just a feeling, produced by a sophisticated model—weaving sensory data, memories, and social feedback into what feels like a coherent, persistent self.

Said another way... while the self is real as an experience, it is illusory as an actual object (in both our internal/external worlds). It's just evolution's biological user interface—helping a body navigate complexity without being crushed by it.

My favorite explanation comes from cognitive scientist Donald Hoffman, and his book The Case Against Reality.

Hoffman argues that evolution has not shaped us to see the truth, but to see what helps us stay alive and pass on genes.

His analogy is striking (if not a bit disturbing): all of reality is like a desktop interface. The icons on your screen don’t reveal what’s actually happening inside the computer—they give you a simplified, usable illusion.

The self is just one of those icons. It’s not real in the way it appears—but it helps you navigate the game.

This can be hard to believe. It implies our (internal & external) perceptions are some sort of lie. But as this essay reveals, at a fundamental level, that's exactly what they are. Lies. Or at the very least, white lies…

The Empirical Truth

Indeed, your brain hides the truth. It does not deliver raw data about the world. What feels like a live broadcast is really just a post-production, editorial suite (and sorry... but you aren't the editor).

I’ll explain with the most basic of games: a race.

Races don't start with a flashing light, they start with the crack of a gun.

But why? Light travels much faster than sound. Isn't it a better signal?

Nope.

Light travels faster but it takes longer to process through the neural cortex. In fact, all of our senses are on their own schedule; they hit our brains at different times, through different channels. Yet somehow, they seem in sync…

Neuroscientist, David Eagelman, has done multiple experiments on this front, revealing how our brain sorts and times our senses.

The first experiment is simple. Participants press a button and then see a light flash on a screen.

They continue to do so, but what they don't realize is that Eagleman is introducing a slight delay each time. Just 10 milliseconds at first. Then 20ms. Then 30ms, all the way up to a 200ms delay in the real world.

But as this happens, their brains don't perceive the delay. They re-adjust. Their brain tells itself, "I've done something and I should get something visual back, so let’s just adapt to this delay".

After a while, their brain gets used to it. But then suddenly, Eagleman removes the delay.

This time, when they hit the button they’re jostled with surprise.

The system glitched. They see the light flash before they fully press the button. As if the effect had caused the cause.

But that's not what happened. It didn't glitch, they did. In the physical world, the light went off as it should: after they pressed the button.

So what's behind this sensory flip?

Something called postdiction, i.e. your brain's tendency to finalize reality a moment after an event.

It quickly gathers recent sensory data, then retroactively constructs the most coherent story. Expecting a delay between a button press and a flash, its removal caused the brain's internal model to misfire, making the flash seem to precede the press.

This is not a bug in the brain—it’s a feature. The brain had rewritten its expectations and adapted to the altered world. When that world snapped back, it didn’t re-adapt fast enough.

Instead, it created an illusory reversal of cause and effect.

This illusion doesn't just happen with vision. It also happens with sound, touch, and even intentionality.

Indeed, we've found a rip in the Matrix; a window into how our brains construct reality—not as a live stream, but as a best-guess reconstruction, edited after the fact for narrative coherence.

More simply stated: your brain isn’t showing you the world as it is. It’s showing you a version of the world it thinks makes sense. And that feeling of "you" as a coherent self—acting with free will, in real time—is part of the constructed story.

The crazy part is that this is all by design. Evolution didn’t care about showing you the truth—it cared about keeping you alive. A coherent sense of self, like a coherent perception of a flash, is a useful shortcut. It helps you navigate the world without drowning in sensory chaos. But it’s not reality. It’s fiction.

And this shows how far your brain will go to keep the story consistent— even if it means bending the truth.

This brings us to perhaps the most dramatic evidence of your brain’s narrative prowess: studies of people whose brains have been split in half…

Subscribe now

One Skull, Two Selves

These editing tricks don’t stop with vision. The same narrative machinery shapes your sense of self, agency, and free will.

No experiment reveals this more clearly than the famous split-brain studies.

In the 1960s, neuroscientists studied patients who had undergone a rare surgery to treat severe epilepsy: cutting the corpus callosum, the thick neural bridge connecting the brain’s two hemispheres. As a result, two brain halves are now unable to communicate.

You’d expect total dysfunction. But strangely, these patients seemed fine. They could walk, talk, and go about life as if nothing had changed.

Until researchers looked closer.

They devised an experiment to send visual input to just one hemisphere at a time. The right hemisphere (which can’t speak but controls the left hand) was shown a snowy scene. The left hemisphere (which controls speech and the right hand) was shown a chicken claw.

Then came a test: the patient was asked to point to related images. The left hand picked a shovel (to clear snow). The right hand picked a chicken (for the claw).

When asked to explain both choices, only the left hemisphere could respond—and it had no idea the right hemisphere had seen snow.

You'd expect the patient to say something like, "Well, I picked the chicken because I saw a chicken claw, but I have no idea why my left hand chose a shovel."

Instead, the patient’s left hemisphere confidently said: “The claw goes with the chicken, and the shovel is to clean out the chicken shed.”

A complete fiction—but told with total confidence.

This is the brain’s default move. When it lacks information, it doesn’t stall or admit confusion—it fills in the blanks. Not to deceive, but to preserve the feeling that everything is part of one smooth, intentional, (you guessed it) story.

Neuroscientist Michael Gazzaniga, who led many of these studies, called this built-in narrator “the interpreter.” It’s like your brain’s PR agent—always spinning a narrative, even when it has no direct knowledge of the events. It values coherence over truth. Narrative over facts.

And this need for unity runs deep. In other split-brain experiments, the two hands—controlled by different hemispheres—actually fought. One hand would reach for a shirt, the other would swat it away. One hand would start a task, the other would try to stop it.

Two minds in one body, in complete disagreement. Talk about internal conflict…

All of this underpins our initial thesis: there’s no single self in charge. The “you” you feel is the interpreter’s creation—a tidy story woven from messy, distributed processes (or multiple ‘selves’).

This brings our journey full circle. Just as newborns reveal the self as a learned construct, and the Headless Way shows no solid "you" at the center of experience, split-brain experiments deliver our final blow to the myth of a single, indivisible self.

Your brain can operate like a divided government with multiple selves running in parallel, and the "I" you think governs it all is just one part of the system—often the last to know what's going on.

Which raises an even deeper question: if the self is a story, who—or what—is the author?

We'll explore this question and more in Part IV... stay tuned. Or, if you want to jump ahead, by all means. Hit this link here → Part IV | Future-Proofing Your Self.

In Part I, we defined the 'self' and explored its origins through the lens of science, philosophy and religion.

In Part II, we'll explore ways to see the illusion.

But first, let's align on the core argument here and why this effort matters.

When we say the self is an illusion, we are not denying that we exist.

Rather, we are attacking a very specific instinct — the feeling that there is a separate, unified "me" who owns a body, who owns mental states, and who acts independently (rather than causally) upon the world.

Our ultimate goal is to replace the mistaken idea of 'a self' with a more grounded understanding of what we really are: 'persons'. This idea comes from modern day philosopher and academic, Jay Garfield, via his book, Losing Ourselves.

Garfield defines persons not as pre-existing, independent selves, but as... "dynamic patterns constructed through ongoing relationships and social interactions."

Unlike the illusory “self” that seems to float above our lives, directing things from the outside, a person is more like a song — something that only exists while it's being played.

We emerge from the interplay of biology, culture, language, and experience, not apart from it.

Said another way… Illusions are not about complete non-existence; they involve something existing in one way but appearing in another. A mirage, for instance, is a real pattern of light that might appear to be water.

Likewise, a person exists as a real, socially and biologically embedded being. But it appears to be a separate, autonomous self.

The alternative is recognizing that we exist not as independent selves but as persons: physical beings embedded in networks of causes and conditions, social beings shaped through language and relationships.

We do not pre-exist and then interact with the world; we emerge from it, constructed through our biological bodies, our cultural context, and our interactions with others.

Our inability to grasp this is so natural, so automatic, that entire civilizations have been built upon it. As mentioned in Part I, The Greeks spoke of the "psyche," Christians of the "soul," Indians of the "atman" — all different ways of trying to name the mysterious "me" inside the machine.

But trying to define the self is like trying to measure how deep the water is in a mirage — the thing you’re seeing/feeling isn’t actually there.

The truth is, we aren't isolated pilots. We are more like waves on the ocean — shaped by the winds, tides, and other waves around us.

Like a five-dollar bill, we’re real not because of some intrinsic essence, but because of the roles, meanings, and agreements that give us form. To understand yourself, you don’t need to look within for some hidden core — you need to look around at the web of causes and conditions that make you who you are.

This reminds me of one of my favorite quotes.

"I am not who I think I am. I am not who you think I am. I am, who I think, that you think I am..."

Read that again. Chew on it. It’s this quote that sparked my curiosity on this topic…

And yet… even if this makes sense intellectually, the moment you stop thinking about it, you snap right back into the old operating system. The sense of “I” returns like a reflex.

So how do we stay rooted in this insight? How do we not just understand it — but see it, feel it, live it?

I don’t have the full answer, but I have uncovered a set of ideas, metaphors, and mental models that can help.

And it starts at the beginning, with how we all began... as babies :-)

Subscribe now

A newborn truth

I've spent the last couple years wrestling with all the big ideas here, spanning ancient philosophy, cutting-edge neuroscience, zen koans, etc.

And then last month, the clearest clue arrived: our newborn baby, Chloe.

Before language, before memories, before the world teaches us who we are, her experience reveals part of the answer: the self isn’t something we’re born with — it’s something we build.

According to developmental neuroscience, a newborn’s brain hasn’t yet developed the systems needed to separate “me” from “not-me.” Core abilities like proprioception (sensing where your body is), interoception (feeling your internal states), and multi-sensory integration (connecting sight, sound, and touch) are still maturing.

As a result, newborns live in what some scientists call an “undifferentiated soup of sensations.” Their vision is blurry. Their body map is incomplete. They don’t yet feel like a defined someone having experiences — they just are. That’s why one baby’s cry can trigger another’s: the line between self and other hasn’t solidified yet.

Even the brain’s Default Mode Network — the system responsible for self-referential thought and inner narration in adults — is mostly inactive in early infancy. It only begins to come online over months and years, reinforcing the idea that the self is a developed feature, not a built-in one.

Psychologists have long described this early stage as “primary narcissism” or the “oceanic feeling” — where the infant feels no clear separation from the caregiver. When a baby is held or fed, it’s not yet experienced as something done by someone else. It just happens. It’s only through repeated interactions — “when I kick, I see movement,” or “when I cry, someone comes” — that the brain begins to form patterns.

These patterns eventually give rise to a sense of agency and of self.

Milestones like recognizing themselves in a mirror around 18 months mark an early step in this construction, but even that is just the beginning.

This developmental arc aligns perfectly with the understanding that the adult self, despite its convincing feeling of solidity and permanence, is an emergent phenomenon; a "useful fiction" or "controlled hallucination," as Anil Seth suggests.

Newborns are the ultimate proof: they don’t "lose" a self or struggle with its illusions; it simply doesn't exist, yet.

The “I” we hold so dear, the inner center of experience, is gradually assembled through physical interactions, social mirroring, language, and memory.’

In other words, we aren’t born with a self; we learn to hallucinate one.

This might sound startling, yet it's validated by science. The boundaries we perceive between "me" and "not me" are not absolute givens. They're learned constructs, essential for navigating the world but not necessarily reflective of an inherent separation from it.

Why does this recognition matter?

Because it challenges our most fundamental assumptions about identity and liberate us from the confines of a rigidly defined ego.

If the self is a construct, it implies a degree of malleability and a deeper, perhaps forgotten, connection to the world.

Babies are the ultimate reminder that identity isn’t fixed. That we’re more fluid, more connected, more open to change than we think.

Beneath all the stories we tell ourselves, there’s still a trace of that original openness — that quiet, 'newborn truth' we came into the world with. And maybe, just maybe, we can find our way back to it.

Now, if you're like me, this might sound like a bunch of mystical mumbo jumbo — 'connection to the world', 'openness'. It's just all so hand wavey.

What does it really mean?

I used to write this stuff off. Until I stumbled upon something I couldn't deny…

That I have no head.

The Headless Way

If I told you that you have no head, you’d probably look at me like I have two.

I know I know, it sounds ridiculous. But there's a (disturbing) line of inquiry that makes it hard to ignore, if not plausible. It's also one of the more profound ways to think about the illusion of self, and to level set on what we really are…

The idea comes from Douglas Harding, a British architect-turned-philosopher who coined the “Headless Way.” His big insight wasn’t born from some ancient manuscript or psychedelic trip — it came during a walk in the Himalayas.

During his trek, he paused, looked down at his body, and then out at the mountains. In that moment something clicked: from his own direct perspective, there was no head. No face. No eyes. Just a wide-open field of vision, containing the world — his legs, the path, the sky, the peaks — but no “thing” doing the seeing.

The “me” he’d assumed lived behind his eyes was experientially absent.

His famous quip was, "I lost a head and gained the world."

What Harding realized is something most of us never question: we assume we’re looking out of a head, but we never actually see that head from our first-person perspective. It’s always inferred — never observed.

In other words... you can see your hands, your legs, your shoes, and the furniture around you. But if you try to see your own face—without using your phone’s camera or a mirror—you can’t. From your own point of view, your face is invisible. It’s just not there.

Rather, you’re just aware of everything in front of you. It’s like you’re a space or an opening through which the world appears. Other people see your face, but you never do directly.

This is difficult to intuit, so Harding created a simple experiment to help people (at least begin) to understand…

First, point at things around you—your shoes, your friends, the wall. You see them clearly.

Next, point back at your own face. What do you see?

Nothing—just the world in front of you.

The idea is that, at the center of your experience, you’re not a person with a head looking out—you’re more like a camera lens or a window: open, clear, and empty, with everything appearing in that space.

It's worth pausing here and letting that sink in…

When I first heard this, it hurt my brain. So here are a few metaphors that helped me understand.

Think about the nature of a window. The window doesn’t see itself; it’s just the opening through which you see everything else.

Similarly, think about a camera. The camera never photographs itself; it just shows what’s in front of it.

Now, consider the sky. It's open and empty, and all the clouds and birds just pass right on through.

Useful, I hope... But Harding wasn’t just being metaphorical.

He insisted this was a literal description of experience. What we see in the mirror is a construct — a few feet away, from another perspective. What others see is a socially agreed-upon “you.” But what you find at (what he called) 'zero distance', from your own center, is not a head looking at the world — it’s an empty space filled with the world.

So, what does this mean?

It means your true “self” isn’t a thing you can see or point to—it’s just an open awareness where all your experiences happen.

It’s a way of seeing yourself that’s very different from how you usually think about being a 'self/me' with a face and a head. Instead, you’re the space where the world shows up.

This 'space' sounds a lot like Garfield's notion of us as 'persons', not selves…

Indeed, Garfield and Harding are saying the same thing, but with different tacks.

Garfield says we're not selves separate from the world, but persons mediated by it and within it.

Harding would say... you are not a thing among other things. You are “No-thing,” and precisely because of that, you are the space in which everything appears.

This might just sound like a fun philosophical trick. But I assure you, as I spend my nights tending to a crying infant, changing diapers in a complete stupor, this has real world implications.

Seeing your “headlessness” disrupts the automatic belief that you are a separate, located ego trapped behind a face. It interrupts the habit of thinking of yourself as an object in the world, and reveals a more immediate, more intimate reality: that your true nature is open, borderless, and deeply connected.

This shift can bring psychological relief, (especially when mid diaper change, you suddenly yourself being both pooped and pee'd on).

You’re no longer a vulnerable little self perched behind the eyes, trying to manage a chaotic world. You are the space in which the chaos appears — and that space is untouched, undisturbed. It’s free.

Harding also noticed how this shift changes relationships.

When you see that you have “no face” here, you become space for the faces of others. Interactions become less confrontational — not face-to-face, but face-to-no-face. You’re more present, more available. Less busy protecting a self-image, more attuned to the world around you.

And the deeper you go, the more freeing it becomes. If the “self” you thought you were isn’t actually here, then what is there to defend? What is there to lose?

Well, for starters: the fear of death, the pressure to perform, the need to constantly assert your identity — all of it starts to loosen its grip.

This is what Harding meant when he said the Headless Way leads to “the peace that passeth all understanding.”

Not because you believe in some abstract truth — but because you looked, and noticed what’s always been right here; no head. No center. No thing. Just this wide, open space — quietly holding the world.

Making this applicable

If this topic interests you, check out Harding's book, On Having No Head. It's an esoteric gem and Harding wielded it for sixty years, trying to convince people to test this claim for themselves.

If only he was alive today... he would have known, this was his theory's moment.

Because that empty center he speaks of... it's the one dimension Silicon Valley can’t hijack.

Sure, algorithms can bend your preferences, curate your desires, and even clone your voice. But they can’t colonize a vacancy.

When you recognize the center as spacious, the latest mimetic frenzy on your feed loses sticking power. There’s no inner Velcro for it to latch onto.

Think of headlessness as a “psychotechnology", i.e. a perceptual trick you can deploy in the checkout line when your toddler erupts, or right before presenting quarterly numbers to a board of caffeinated VCs.

Pop the hood, notice nobody’s driving, and the nervous system resets to factory settings of ease and attentiveness.

To make this more visceral, here are are some personal 'day in the headless life' moments:

It's early morning. Too early... I stumble to the sink, baby monitor crackling on the counter.

Toothbrush in hand, I catch my reflection. A familiar father-face stares back, but here on this side of the glass? Vacancy. The mirror becomes a frame around the scene rather than evidence of a cramped identity.

Chloe wails; sound erupts inside the same openness. I respond, but the response arises unforced, like weather.

Now, the dreaded afternoon Zoom call. Ten rectangles float on the screen, each a curated persona. I toggle gallery view off and drop into headless seeing: colleagues speak; words appear; my own voice joins the mix; nowhere is there a solid epicenter absorbing credit or blame. Ironically, performance improves when there’s no one to impress.

The day's winding down. I'm on an evening stroller walk. Sunset flares across Austin’s sky. In place of “I’m under the sky,” the sentence flips: “Sky, trees, stroller, wife—all appear inside … this.” The boundary between inner and outer crumbles. I'm not zoning out; I'm tasting the scene without the usual packaging.

Objection handling

In closing, let's arm you with some objection handling. Because no doubt, they're going to come, be it from that little voice in your head, or a friend during happy hour.

Here are some of the most common…

“Isn’t this just escapism?” It would make for a strange escapism... given that headlessness drops you more fully into sensations than the normal selfie-centric mode. You’re not transcending life—you’re removing the tinted goggles.

“If there’s no self, who raises the baby or answers the email?” Tasks still get done. The spreadsheet edits itself the same way your heart beats itself. Agency functions, but minus the tight-fisted claimer of agency. It's responsibility without the migraine.

“Sounds cool, but doesn't it fade?” Yes. The default mind springs back like memory foam. Harding recommends dozens of micro-glimpses a day— a finger point, a mirror flip—to erode the reflex. He saw practice not as climbing a ladder toward enlightenment but remembering a joke you keep forgetting to laugh at.

But let's be honest, it's unlikely you'll find time to point at objects and then back at yourself, especially when it might be needed most, in public…

As such, we must go deeper into your programming, and hack it from the inside out.

Doing so starts with realizing the extent to which your mind is already an expert hacker in itself, keeping you from seeing what's actually happening, out there in the 'real' world (all for evolutionary benefit).

We’ll cover all this and more in Part III, by studying some scientific experiments that more empirically lift the veil. But if you want to jump ahead, go for it! Hit this link here → Part III | Future-Proofing Your Self

You've likely heard this idea before, i.e. the illusory self, nonself, nonduality, emptiness, etc. It's an ancient concept. One that has haunted and liberated the most introspective among us for thousands of years.

But like most ancient practices, it's woefully underutilized, under studied, and relegated to the realm of monks.

Fortunately, a resurgence of Buddhism, mindfulness, and meditation has put this illusion back into scope, at just the right time. A time when technology and AI is about to warp our sense of self, including all of its inputs: identity, agency, even free will.

As such, the illusion of self is no longer just a mystical inquiry. It’s a survival skill; a precondition for psychological sovereignty in the age of AI.

This distortion will bring many oddities and risks, some knowable, most not... But fortunately, they all rest on a single assumption: that beneath it all, there’s even a solid “self” to begin with; a persistent 'center' of being in the cross hairs.

This essay will help you challenge this assumption.

It's a collection of ideas, exercises, and experiments to help you not just understand the illusion intellectually, but to feel it experientially, and ultimately, live from the truth and freedom it reveals.

To be clear, I’m no guru. I’m far from an expert on this stuff. I’m just a guy trying to become a better father, friend, and leader. I’m also trying to become a more independent thinker, unmediated by the wants and needs of the outside world. Towards that end, I’ve tried all the things; all the self help books, the plant medicine, and the therapy. Each was useful in its own right. But this is the thing that has grabbed me most. It’s the furthest upstream, hitting the roots of all suffering, and of most all our modern day plights.

In the end, you’ll leave realizing this illusion isn’t just theoretical. It's structural, empirical, and becoming increasingly validated by science. And in an age of infinite avatars, curated identities, and AI companions, it's also becoming a dangerous trap.

Now, I'm far from free of it (hence the reason for this essay). But when I encounter those who are —people who radiate a certain stillness, clarity, and responsiveness—it’s clear that dissolving the illusion doesn’t detach them from life. It roots them more deeply in it. They move through the world with a grace we’re all quietly yearning for. A grace that will be critical in the exponential age.

Because one thing has become certain. This new era will indeed bring abundance. But largely within the material world. Our inner world faces a scarcity of meaning, purpose, and groundedness, largely due to a sense of self that will be maximally malleable, programmable, and monetizable.

The time to start preparing for this future is now. Doing so will involve a seemingly strange but profound set of perspectives and tools, from seeing that you have no head, to not wanting to ever trust your eyes again, to embracing your lack of free will.

So with that, let's begin, starting with a life update that sparked this whole pursuit.

A Wake Up Call

Life just changed in the most meaningful way.

We had a baby girl!

Her name is Chloe and she's a warm, wiggly, cooing bundle of joy, tucked into the pouch of my new 'kangaroo' shirt as I write this essay.

She's also a wake up call; to get out of my head and become more selfless.

No small task... I've always been more self-centered than I care to admit; lost in my inner world and tormented by a tyrannical ego; whipping me incessantly towards a better version; smarter, richer, funnier, more accomplished, more... admired. Psychologist call this 'the idealized self': a mask built not from who we are, but from who we thought we had to be to earn love and escape shame. This result is a collage of childhood pain, cultural ideals, and personal distortion.

Perhaps you know this feeling. It's a constant inner competition with yourself. But win or lose, the result is often the same: tension, guilt, exhaustion, and a burning frustration that these feelings even exist.

In our youth, this tension is useful. It fuels ambition as we carve out a place in the world. But as we get older, it starts to backfire, especially as other 'selves' begin to matter more than just our own; a life partner, a child, an employee. These people need us to show up in a different way.

Yet more often than not, we remain lost in the labyrinth of our own becoming.

Quite the paradox... how the act of becoming becomes the very thing keeping us from who we want to be.

This tension is a tale as old as time. But now, the labyrinth is evolving. Technology is warping it into a black hole. Tiny screens and algorithms fuel the 'act of becoming' on the grandest of scales, in the most profound of ways.

The net effect? We get what we want on demand, but rarely what we 'want to want'.

Philosopher René Girard called this mimetic desire — the idea that we don’t create our own wants/desires, we unconsciously borrow them, copying what others around us seem to want; their ambitions, their preferences, their ideals. More often than not, we don't intrinsically value these things. Our unconscious psyches just see someone else getting love and attention for having them, and decide to pursue the same.

Mimetic desire has always been part of the human condition, but it's inflamed by social media and influencer culture, and now being super charged by AI. Hell, the influencers are becoming AI's, leading us to crave the fullest forms of fabrication.

Needless to say, we're hurdling towards a strange new world, defined by virtual worlds and avatars, AI agents and AI companions, brain computer interfaces and gene editing.

When any person can be a curated projection, and every digital space invites you to perform a new persona, the question of “who am I?” isn't just philosophical— it's practical.

So... who are you? What exactly is this feeling of a self? Here’s my working definition and a target for our inquiry.

Subscribe now

Understanding the Self

For most of us, 'the self' is the main character of our story. It's the feeling of 'me' at the center of experience; a fixed entity who thinks, feels, decides, and acts. It's the pilot living in your head, with a personality, wants, dreams, and desires.

This feeling is so intuitive, so innate, that questioning it almost feels absurd. Of course you are you, and I am me, and well... that's just that. End of story.

But modern neuroscience offers a different story. According to cognitive scientist Dan Dennet, our sense of self is just a 'useful fiction', a mere narrative we tell ourselves, about ourselves. And to be fair, it's very useful, indeed. I need to think of myself as a father, a husband, a business partner— if only to make it through the week, much less excel at each.

As such, it's important to clarify: this is not the sense of self worth challenging. What we're targeting is more fundamental and subtle.

It's also the greatest source of suffering. It's the feeling of being a subject internal to our bodies; a mini-you behind your eyes. This version of self — an ego both produced by and identified with thought — is where the burden lies. Fortunately, it's also not a solid, concrete thing. It's a simulation. Or rather, to draw an eerie parallel to our new AI friends, it's a generative model.

Philosopher Thomas Metzinger calls it the Phenomenal Self-Model — a kind of internal avatar your brain assembles to unify experience and track itself in the world. And it works beautifully, stitching sounds, colors, smells, sensations into a coherent tapestry of “subject object"—instead of a random chaos of disconnected inputs. Neuroscientists call this 'sensory integration'.

But—and here’s the uncomfortable part—this model doesn’t reflect reality directly. It constructs it.

Take color: there’s no green “out there” in the world, only light waves bouncing around. Your brain interprets those waves and gives you the ‘feeling’ of greenness—what philosophers call qualia, aka the subjective, felt qualities of experience; the "what it's like" aspects of mental states.

As for the ‘self’, this feeling is what neuroscientist Anil Seth calls a 'controlled hallucination'. And it's entirely convincing because it’s transparent — we don’t see it as a 'model'. We just see it as me. One major reason for this is memory. It convinces us there's a stable 'me' who persists over time.

Now, by this point, some of you might be nodding along. “Sure,” you might say, “I get it. The self is fluid. I’m not the same person I was ten years ago. I’m a socially influenced, ever-changing mix of biology and psychology. I’m fine with that.”

And yes, that’s a perfectly valid view.

But even if you intellectually accept that the self is a construct, notice what still lingers beneath: a quiet but persistent sense of being someone—a distinct “me” behind the experience. A self who has a body and mind, but is somehow separate from both.

This is embedded in our language. In the West, we don’t say “I am a body.” We say, “I have a body.” That subtle phrasing reveals everything.

As a quick thought experiment, let's play with our innate capacity for imagination and desire. Close your eyes and imagine being your favorite athlete or super star, say Michael Jordan or Taylor Swift.

How'd that go? Could you imagine having MJ's jump shot and spraying champagne after a championship win? Did you envision a glitzy outfit and blonde curls bouncing in your face while 'shaking it off' in front of 100,000+ screaming fans?

But notice something subtle: you didn't imagine and desire being MJ or T Swift. That wouldn't make much sense. They've already lived those experiences. What you imagined/desired was being you, in their body with their talents.

This tiny twist... this possibility to form such a desire, it shows how we don't identify as a body or mind, but rather, as something beyond... something that has a body/mind, and in principle, could have another one.

This is just one example of many showing how deeply the illusion runs. Its purpose isn’t to argue about the power of imagination, or to prove or disprove the existence of a self. It’s to just highlight how our ability to even form such desires—and to imagine a “me” swapping out bodies like outfits—suggests that, on some deep, automatic level, we believe ourselves to be 'a self'.

And this implicit belief doesn’t stop with imagination. It finds its most enduring home in religion.

Because if we can imagine inhabiting different bodies, it’s not a huge leap to imagine the self surviving the loss of this one altogether.

Across cultures and traditions, the belief in a soul—a self that outlives the body—has been the default setting for most of human history. Whether it travels to heaven, hell, or into another life entirely, the soul is often treated as the real “you,” somehow untouchable by death.

Ancient traditions like Hinduism call in an Ātman — an eternal, unchanging self identical to the divine. This Atman survives bodily death and is central to the cycle of rebirth.

Islam calls it Ruh and Nas. Ruh being the "divine spirit breathed into humans by god". Nas being the sef or ego, portrayed as "the sea of base desires that must be purified".

Christianity evolved this into the immortal soul — the true “you,” distinct from the body, judged by God after death.

So whether through cultural inheritance or cognitive instinct, we tend to experience ourselves as selves. We are wired to believe in a self the same way we are wired to fall for the Müller-Lyer illusion:

Even when we know the lines are equal, they still look unequal.

So, if we hope to see the lines as they are... it helps to ask:

Did we arrive at the idea of the self or soul through a first-principles investigation of reality?

Or did we invent it—because it was socially and emotionally useful? A mere patch to manage fear, morality, mortality?

Also, if this ‘feeling’ of self is just an evolutionary survival strategy, what role does it play in our future? How can this newfound awareness support our next evolutionary leap?

We'll answer all these questions and more in Part II. But if you want to jump ahead, by all means! See this link here → Part II | Future-Proofing Your Self

Spotify Podcast Link

It’s time for some straight talk.

No hype, no press release optimism, no corporate spin. Just two guys with some scar tissue—and a lot of history—talking about what’s really happening in the world of spatial computing/XR.

In this episode, I sit down with Kharis O’Connell—former design leader at AR headset startup: Meta (the original Meta), Google, and Amazon.

Kharis has been in the XR arena since before the game began. From early experiments with Nokia's Point & Find in 2007 to helping define zero-learning-curve design principles at (OG) Meta, he’s seen the promise, the collapse, and the recycled hype cycles firsthand.

But this isn’t just a walk down memory lane.

It’s a reckoning.

Kharis recently left the XR industry. He’s no longer building AR glasses or pitching immersive futures to execs in Silicon Valley boardrooms. And that’s what makes this conversation special. He’s honest. Candid. Unfiltered. It’s the kind of clarity you only get once you’re no longer trying to raise funding or ship a prototype.

Reasons to listen:

• If you’re in the spatial computing space and having doubts

• If you’ve ever wondered why AR hasn’t had its iPhone moment

• If you’re building in this space and unsure whether to keep going

We cover:

• The philosophical vision behind the original Meta and why it failed

• Why “zero learning curve” design is mostly a myth

• What really happened at Google’s stealth AR program (and why it shut down)

• The failure of imagination across today’s use cases (yes, cooking demos again)

• Whether Apple Vision Pro was our “iPhone moment” (spoiler: it wasn’t)

• How the obsession with technology becomes an echo chamber—and how Kharis broke out

• The one big question every builder should ask themselves: Who is this for—and at what cost?

This episode won’t give you all the neat answers. But it will definitely give you better questions.

And if you’re like me—still carrying a flicker of belief that this tech can nourish the soul, not just the shareholder—then maybe that’s enough to keep going.

See the Substack audio above and Apple + Spotify links below :)

Apple Podcast Link

Spotify Podcast Link

If compelled by this new direction, be sure to subscribe to Dream Machines on your favorite podcast player.

And if you’re feeling generous, would beyond appreciate some stars/rating on Apple/Spotify! Goes a long way to helping us grow & spread the good word….

Apple Podcast Link

Spotify Podcast Link

Change is in the air! On multiple fronts...

The first? A warm, wiggling, cooing bundle of joy, currently in the pouch of my new kangaroo shirt.

Her name is Chloe :)

Yes, fatherhood recently arrived (hence the lack of content... and sleep).

It's been both beautiful and jarring, all at the same time. Needless to say, my world will never be the same, in all the best ways.

Our collective world is also changing and in a similar fashion; with equal intensity and equal pace.

Hence the reason for this announcement/post: I'm re-branding this newsletter and podcast.

Medium Energy is now Dream Machines, narrowing the focus on the most powerful force in technology today: the convergence of spatial computing and AI.

These tools are two sides of the same coin. Problem is... most people don't realize how and why. Nor do they realize what it means for their business, their careers, their children, their relationships, the list goes on.

This newsletter/pod will change that.

Now, after a decade on the spatial computing front lines, I’ll be the first to admit... spatial computing has been either 'too early' or just a 'nice to have'. Us industry vets have the scar tissue to show for it. Yet here we remain, patiently waiting for its breakout moment while staring longingly towards that trillion dollar TAM stuck on a perpetual horizon.

Amidst the hype cycles, spatial has also been misunderstood.

It's far more than just AR/VR headsets and the 'metaverse'.

Spatial computing is about any machine/device that can see, interact, and localize itself in the real world: this includes AR glasses, robots, drones, and self driving cars.

It also includes data, of an incredibly important kind... spatial data, i.e. data with an X-Y-Z coordinate, or what I like to call... 'world data' representing people, places, and things.

For decades, much of this data has lied dormant. Despite being immensely valuable, it's been woefully hard to access and to use. This ranges from 3D data (CAD, point clouds), to sensor data (IoT), to visual data of all types (photos, videos, scans)

But now, the terrain is shifting, and this data has an all new raison d'être...

AI is now supercharging spatial computing, and spatial is returning the favor — giving AI a body (robots), a perspective (3D computer vision), and a way to hear, navigate, and interact with the world (AR glasses).

Of most importance, it will give AI real world agency, and ultimately, the feedback loop needed to unlock the holy grail: AGI (artificial general intelligence).

At which point, the machines will dream, while helping humanity realize many of its dreams along the way.

As such, Dream Machines isn’t just a nod to the power of spatial + AI. It’s your window into all the ways the boundary between imagination and reality is about to collapse.

This convergence will unlock human potential in ways we can barely imagine. But it's also going to challenge everything we know about the human experience—about the self, identity, presence, and connection.

Which is why Dream Machines isn’t just about where technology is going—it's about where it’s taking us, fully devoted to a singular, existential question: how do we take full advantage of these tools, while staying grounded in the real world?

If that resonates, you’re in the right place, at the right time.

This convergence is happening now— from Meta's Project Orion and Llama, to Google's AndroidXR and Gemini, to NVIDIA's Omniverse and robotics, to 'large world models' like World Labs and Niantic's Geospatial World Model. Not to mention, a whole new bevy of startups in between.

The result? I’ll pull from this essay I wrote on the Apple Vision Pro…

AR’s arrival is going to shift people’s relationship with technology while reducing the afflictions born of abstracted, ‘flat computing’. Or the afflictions born of boxes tethered to a wall and TV (aka Xbox or PS5). These current form factors are what keep kids/people stuck inside, isolated, and socially inept.

In contrast… AR, in its ultimate form, will free kids from the confines of a screen and a living room with an outlet, thrusting them back into nature, back into face-to-face contact, and back into a world longed for by prior generations. A world of scratched knees from a treasure hunt in the park, of youthful pride from a fort forged in the woods, or of confidence from winning an argument while playing make-believe in the backyard.

Except for this time, the treasure becomes real, the forts become labyrinths, and the figments of make belief become not so make belief…

If this vision compels you, subscribe to both the newsletter and podcast below! We'd love to have you as part of our growing community of thinkers, builders, and doers.

Subscribe now

Apple Podcast Link

Spotify Podcast Link

And if you know anyone else who might also find this newsletter of use, please do send it their way :)

The new content plan is as follows:

• Weekly insights into our industries progress

• Monthly deep dives on companies and big ideas

• Monthly podcasts with the greatest minds in Spatial + AI

• Some other random smatterings along the way (I will get to more sci-fi stuff, one of these days, I swear!)

If you want a preview, here are my most popular essays:

• The Fate of the Apple Vision Pro: Part II

• The Ultimate Promise of Spatial Computing

• How to Defend the Metaverse

• Finding Solace in the Age of AI

• 2042: A day in the metaverse life

• The Humanoid Race

• Why Everyone Should Go To Japan

Lastly, if you’ve been a part of this journey to date, thank you… I beyond appreciate your interest and attention to date (as there are few things less scarce!)

See y’all on the front lines.

This feels like history in the making. And if you're a spatial computing enthusiast (AR/VR/XR) , it also feels like hope…

In case you haven't heard the news: OpenAI just acquired Jony Ives AI hardware startup, ‘io’ for $6.5B+.

The internet is now a buzz with speculation. Everyone is asking the same question — what exactly are they going to build?

My bet?

OpenAI is going to (eventually) build category defining AR glasses.

And in pairing with Jony Ive, they're not just trying to compete… they're trying to change the world.

This could yield the "iPhone moment" for AR that many of us have been waiting for, but with the AI-first twist this form factor can't live without.

The thesis is clear: AI needs a better way to understand and interact with our world, to break free from the confines of the 2D screen. It needs eyes, it needs context, and it needs to be with us, not just on a device.

And while that exact nature of that device remains shrouded in mystery, one thing is certain: when you fuse the world’s most potent AI brain with the planet’s most revered design sensibility, you’re not aiming to create a slightly more elegant chatbot interface or a me-too hardware product (e.g. speakers, earbuds, phone, etc).

You’re aiming for a revolution.

What could be more worthy of such a leap?

AR glasses will be a full circle evolution in human computer interaction, surfacing information in the way our brains have evolved to interpret the world. They’ll also bring our eyes back up to the world and to each other.

Imagine… a ‘generative UI’ that doesn’t just overlay information, it injects your perceptual system with maximum agency; an AI that moves beyond mere information retrieval to become an intuitive guide, an amplifier of your own curiosity and creativity.

The ultimate result? Machines that conform to us, rather than us conforming to them (hunched over, distracted, restless, addicted).

This is the vision that’s kept me in the industry for all these years, despite all the ups and (mostly) downs.

The implications for how we learn, work, create, and connect are profound and hard to imagine, especially if you’ve never tried really, really good augmented reality. I’d reckon less than a thousand people have. On the entire planet.

As for our wider industry, this should be a wake up call. Apple, Meta, Google, the whole AR/VR/XR/Spatial Computing alphabet soup – it’s time to kick things into gear and up our game.

And if you’re just a passive observer…. get your popcorn out. Because the AR + AI race just got a whole lot more interesting.

Okay… Easy there cowboy

I know what you're thinking. Pump the brakes, Evan, you're way over your skis.

There's zero guarantee OpenAI will get into AR. In fact, Sam Altman just told the WSJ they aren't making glasses.

What we do know is that Sam has a prototype, and he’s told OpenA staff it’s “the coolest piece of technology that the world will have ever seen.”

So far, the rumor mill points to some sort of little puck-like speaker thing that you can put on your desk, in your pocket, around your neck, etc.

Something like this…

Call me crazy, but this (or anything remotely close) would be far from “the coolest technology we’ve ever seen”.

I’m not buying it. I stand by my thesis, and here's why.

(1) AR Patents: First, OpenAI has filed numerous patents for AR.

Sam’s also one sly cat, and one of the last people I’d want to face at the poker table… Do you really think he’s going to show all his cards?

I could end my case here but let's keep going.

(2) A family of connected devices: Both Sam and Ive said they're going to make a family of devices.

Jony Ive has also suggested he's on a mission to rid the world of screens. Most notably, during this interview with Patrick Collison at the Stripe conference. It's an incredible conversation and will give you an inside glimpse into his motives. You can hint feelings of remorse for what the screen/iPhone has done to the world, reducing out attention span, creating all new forms of addiction, inflaming our insecurities, the list goes on...

He then said in the announcement video,"people have an appetite for something new, which is a reflection on a sort of unease with where we currently are..."

Now, I completely agree with getting rid of screens. But are we really going to get rid of the visual elements of computing?

(3) Seeing is believing: We are visual, experiential creatures. We need to see it to believe it. We need to see the inside our AirBNB, to fully inspect our new car, and to watch (360) videos from our loved ones.

I just don't see a world in which Ive doesn’t embrace our visual nature, and doesn’t embrace the world as our new desktop background.

What are we gonna do insteead… carry a puck/little box with a projector, point it at a bevy of surfaces, and hopes it pans out?

Me thinks not...

4/ AR = Ultimate Fun & Delight: In that same video, Ive talks about his design thesis and the ethos that drives him.

He believes product design should be all about eliciting delight, and creating more pure joy and fun in the world....

Again, most people haven't seen what truly good AR looks. I have, and I cried.

It was from trying an actual demo of this experience below, which we built with Nike back in my days at (OG) Meta:

This experience is what drives me to this day, and I’m telling you… it was the pinnacle of delight (within the realms of creativity & productivity).

The Strategy

I could keep going with my speculations, but I'll stop there for now.

Equally as important to what they build, is the how & why... aka their strategy. After all, what they build won't matter if they can't compete.

So, given who they're up against (Apple, Google, Meta), you have to ask the question... is this even the right strategy? Or is it a distraction?

Because in many ways... they're putting themselves into a big tech quantum state of superposition, trying to be a bit like Apple, a bit like Google, and everything else in between.

On one hand, you could argue OpenAI needs to focus on supplanting Google, and take a similar tack: building the OS/software/killer app (e.g. search) and have their apps on every device, in every ecosystem, in order to maximize scale.

And when Sam founded the company, it indeed was in direct opposition to Google. his catalyzing thesis was that Google was poised to run away with the AI race, and someone needed to compete.

But on the other hand... competing head on with Google is terrifying, especially now (did you see the last Google I/O? Mind blowing). Red ocean blood baths abound.

Now Apple? A different story. They're limping out of the gates in the AI race. Heck, you could argue they're still at the starting line. Their effort to date has been nothing short of disastrous, mired in politics and bureaucracy (as this recent Bloomberg article confirms).

In tandem, the iPhone hasn't meaningfully changed in generations, the Apple Vision Pro is a public R&D project (which I fully endorse, for the record), and developers continue to sour on Apple's App Store philosophy and mechanics.

In other words… Apple has cracks in its armor. Google is getting its mojo back. So going after Apple actually makes sense. And since we're obviously living in a simulation, OpenAI now has who Steve Jobs called his 'spiritual partner', making the onslaught now all the more fun and dramatic.

Regardless, things just got much, much more interesting.

Call it speculation, call it wishful thinking, call it hopium: AR glasses are coming at some point.

It might not be the first device, in fact, considering the technical hurdles that remain for AR, it almost certainly won't. But I'm betting you my bottom dollar: its on the mind, its on the roadmap, and its coming.

As such, I'm updating our AR + AI Racecard with OpenAI and Google.

You'll find below my 'pillar-based' analysis below.

For a similar analysis on Meta & Apple, check out my last article: The AR + AI Race Non-finale

With that, get your popcorn ready. The outcomes are uncertain, but the drama is sure to be first rate.

Oh, and if you're so compelled, tell me your thoughts in the comments below!

Who do you think is going to win? Where are the gaps in my thinking?

The AR + AI Racecard

The Contenders

1. Apple: The incumbent king of premium hardware, software integration, and a fiercely (but increasingly less) loyal ecosystem.

2. Meta: The relentless metaverse visionary, betting big on social VR/AR and open ecosystem development

3. OpenAI-IO: The AI-native, design-led disruptor, poised to redefine AI-human interaction.

4. Google: The re-awakened giant of AI and information, with deep pockets, vast data, and significant (if sometimes fragmented) AR/AI initiatives (think Project Astra, ARCore, Lens, etc.).

We analyzed Apple & Meta in our last essay (pasted further below). So let’s now breakdown our new entrants across the same critical battlefronts:

Software: AI Agents Will Redefine the OS

OpenAI-IO

• Forget app grids and traditional UI/UX. If my hunch is right, OpenAI-IO isn’t just building AR glasses; they're crafting the native interface for AGI.

• This means an OS designed from scratch around AI agents and a truly generative UI. Think less about swiping and tapping, and more about natural language interaction, with the visual interface materializing contextually, precisely when and how you need it.

• Why browse for an app when an AI agent can anticipate your need or summon the exact information/tool in a visual whisper? This is the ultimate "clean slate" advantage – no iOS or Android legacy to protect, just a relentless pursuit of the most intuitive AI-human interface. The "OS" becomes the AI itself, orchestrating experiences

Google

• Google has been telegraphing its AI agent ambitions for years (think Assistant, Lens, and the recent Project Astra demo).

• Their XR software play will undoubtedly leverage this deep expertise. Expect an Android-derived (or perhaps a new, lightweight, AI-centric) OS where AI agents are first-class citizens, deeply integrated with Google’s vast knowledge graph and services (Maps, Search, Workspace).

• The challenge for Google? Making it feel like a cohesive, new paradigm, not just Android shoehorned into glasses, and navigating the innovator's dilemma of how this new OS interacts with their existing Android dominance. But an AI that can "see what you see" and proactively offer assistance through AR? That’s Google’s sweet spot.

Innovator’s Dilemma: Who’s Free to Disrupt?

OpenAI-IO:

• This is where OpenAI-IO shines. They have zero existing hardware cash cows to protect. No iPhone sales to cannibalize, no dominant mobile OS to tiptoe around.

• They can be ruthless in their pursuit of the optimal AI-first AR experience, even if it upends current computing paradigms.

• This freedom is a massive competitive advantage. Their only "legacy" is cutting-edge AI, and these AR glasses are being built to serve that master.

Google

• Google faces a more complex path. While not as hardware-centric as Apple, their Android ecosystem is vast, and their primary revenue driver is advertising, fueled by data from services like Search.

• How does a deeply personal, always-on AR device fit into this? Will they create a truly open spatial OS that might compete with Android partners, or a more controlled experience to safeguard their service ecosystem?

• Their history shows a willingness to experiment (and sometimes abandon), but a full-scale AR push will require navigating these internal tensions. How do you truly innovate when you are the current paradigm in so many ways?

AI: The Engine of Spatial Intelligence

OpenAI-IO:

• OpenAI isn't just using AI; they are defining the frontiers of AGI. Their AR glasses won't just have AI features; they will be an AI, an embodied intelligence.

• Expect foundational models running (perhaps with clever on-device/cloud hybrid approaches) that offer unparalleled reasoning, personalization, and generative capabilities directly within the user's perception. The world itself becomes the prompt.

Google

• Google’s strategy? Unmatched data, powerful models, pervasive AI

• Google's AI prowess is legendary, fueled by unimaginable datasets from Search, YouTube, Maps, etc., and world-class research from Google AI and DeepMind.

• Their strength lies in contextual understanding, information retrieval, and services like real-time translation or visual search (Lens). For Google XR, AI will be about delivering Google's information and services in a more ambient, spatially aware manner. The key will be harnessing this power into a cohesive, delightful user experience in AR without feeling intrusive.

Hardware: Design Ethos and Go-to-Market Approach

OpenAI-IO:

• With Jony Ive at the design helm, expect nothing less than exquisitely crafted hardware that prioritizes user experience, elegance, and perhaps a minimalist aesthetic.

• The focus will likely be on a premium, aspirational device where the technology feels almost invisible, seamlessly blending with the user.

• The challenge? Manufacturing at scale and hitting a price point that, while likely premium, doesn't relegate it to a Veblen good. But Ive's involvement practically guarantees the product will be desirable.

Google

• Google’s hardware play will be more open and partner-centric

• Google's hardware history with XR is... varied (RIP Daydream, ahem, Glass). They might not aim to be the sole, vertically integrated hardware provider like Apple or potentially OpenAI-IO.

• Instead, we might see them focus on creating a core AI-XR software platform and reference designs, empowering a broader ecosystem of hardware partners (think their Pixel strategy, but for AR, or their renewed partnership with Samsung).

• This approach accelerates reach but risks fragmentation and inconsistent user experiences. Their "secret sauce" will be the AI and services powering these diverse endpoints.

Founder Vision & Leadership:

OpenAI-IO:

• Sam Altman is driven by a singular, audacious goal: achieving AGI.

• Jony Ive is driven by an obsessive pursuit of perfect design and user experience.

• This combination is electric. Altman isn't playing for incremental gains; he’s playing for a paradigm shift. Ive ensures that shift is human-centric. This is conviction-driven leadership at its finest, focused on building something genuinely new.

Google

• Sundar Pichai has clearly stated that AI is the core of Google's future, infusing every product and service. Google XR will be a key expression of this.

• While perhaps not with the same singular, public-facing "bet-the-company" XR fervor as Zuck, the leadership commitment to AI as the enabler of new experiences is undeniable.

• The question is whether XR hardware itself becomes a top-tier strategic priority or remains primarily a vehicle for their AI services.

Philosophy: Open vs. Closed Systems

OpenAI-IO:

• This is an unknown. Will OpenAI opt for a more controlled experience, or will they finally live up to their name with a more ‘Open’ platform?

• Given the tight integration of AI and the Jony Ive design ethos, the initial hardware/software experience will likely be relatively controlled to ensure quality and coherence.

• However, OpenAI's broader strategy involves making its AI models available through APIs. We might see a future where the glasses are a premium conduit to an increasingly open AI platform, allowing developers to build "skills" or "agents" for the AR experience.

Google

• Google will likely continue the Android legacy – Openness with guardrails

• Google’s DNA is rooted in Android’s openness, which spurred massive ecosystem growth. They’ll likely aim for a similar model in XR to encourage developer adoption and hardware diversity.

• However, access to Google's core AI services and data will undoubtedly come with Google's terms, creating a "semi-open" environment.

• The challenge is fostering true innovation while maintaining the quality and integrity of the user experience tied to their services.

Developer Trust & Traction:

OpenAI-IO:

• If OpenAI-IO delivers on the promise of AR glasses as a direct interface to their cutting-edge AI models, developers will flock.

• The opportunity to create applications powered by AGI-level intelligence in a spatial context is a greenfield of immense proportions.

• Trust will be built on the power and accessibility of their AI tools and a clear value proposition for developers.

Google

• Google will be aiming to leveraging the existing Android Army, while principally rebuilding the XR Faith

• The existing Android developer base is massive. The key is providing them with compelling tools, consistent platform support (their XR efforts have been a bit start-stop), and clear monetization paths to translate their mobile success to spatial. Project Astra and advances in their AI services are powerful carrots.

Fashion & Culture

OpenAI-IO

• Jony Ive isn't just a designer; he's a cultural icon. His involvement alone lends AR a level of design credibility and "cool factor" it has struggled to achieve.

• If anyone can make face computers fashionable and desirable, it's Ive. This could significantly shift public perception.

Google

• Google certainly has tougher path here due to the cultural missteps of Google Glass (aka the Glass Hole legacy).

• They’ll need to lean heavily on sleek partner designs (if they go that route) or develop an incredibly compelling aesthetic internally, perhaps focusing on minimalist, almost invisible tech, to win hearts and minds from a fashion perspective. Their focus will likely be more on utility first, fashion second.

Product Iteration & Consumer Comfort

OpenAI-IO

• Ive is known for striving for a perfectly polished V1.

• OpenAI, however, comes from a research background where rapid iteration is key. Finding the balance will be crucial.

• The AI will evolve constantly, but the hardware will need to feel "right" from day one to gain consumer trust.

Google

• Google is comfortable launching products in beta and iterating publicly based on user feedback.

• This can lead to faster learning but also a perception of products being perpetually unfinished.

• For AR, where comfort and reliability are paramount, they'll need to ensure a high-quality baseline experience.

Distribution

• OpenAI-io

• Given the likely premium nature of an Ive-designed product and OpenAI's current standing, a direct-to-consumer model, perhaps with select high-end retail partnerships, seems probable for initial launch.

• Scaled distribution will be a new muscle for OpenAI to build.

Google

• If Google focuses on a platform/software approach with hardware partners, their distribution is instantly global through the existing Android device ecosystem and retail channels.

• If they launch their own "Pixel-for-AR" device, distribution will be more controlled but can leverage their existing hardware channels.

Product Portfolio: The Ecosystem Driving AR Use Cases

OpenAI-IO:

• OpenAI's "portfolio" is its suite of world-leading AI models.

• The AR glasses become the prime delivery mechanism for these models to provide value across countless applications – from hyper-intelligent assistants to creative tools to powerful analytical interfaces.

• The use cases will be defined by what their AI can do spatially.

Google

• Google's immense portfolio – Search, Maps, YouTube, Workspace, Assistant, Translate – provides an almost endless list of powerful services that can be integrated into an AR experience.

• The AR glasses become a new, more contextual way to access and interact with the Google ecosystem you already use.

Initial Readout (Not a final verdict, just a summary)

OpenAI-IO

• Poised to deliver massive disruption by redefining the experience of AR through an AI-first, design-led approach.

• Their path is to create something so compelling and intelligent it carves out a new premium category.

• Their biggest challenge: scaling hardware production and distribution from a (relative) standing start.

Google

• Has all the ingredients for a dominant AR play – AI, data, an existing OS footprint, and vast services.

• Their path is likely to make AR an ambient extension of the Google services ecosystem.

• Their biggest challenge: a historically fragmented XR hardware strategy and the innovator's dilemma of how AR fits with their existing Android dominance.

This sets the stage for a truly fascinating four-way race.

Each player brings unique strengths and philosophies to the table. The coming years won't just be about who builds the best device, but who crafts the most compelling, useful, and ultimately indispensable experience at the intersection of AI and our perceived reality.

If you’d like to dive into a similar analysis on Meta and Apple, check out my last article here: The AR + AI Race Non-finale.

Spotify Podcast Link

At a time when most conversations about technology skew dystopian—AI anxiety, screen addiction, digital isolation—Dr. Walter Greenleaf offers a refreshingly grounded and hopeful counter-narrative.

Walter is the Yoda of therapeutic virtual reality. He’s spent nearly four decades at the intersection of neuroscience, immersive technology, and medicine. In this conversation, he outlined a vision of the future in which virtual environments help us heal some of our greatest and most common afflictions.

His thesis: AR/VR is not just a source of novelty and entertainment. It’s a medium for personal transformation.

From Simulation to Self-Regulation

My favorite part of this conversation is Greenleaf’s reframing of neuroplasticity—not as an abstract biological process, but as learning. Learning how to respond to fear. How to stay calm in moments of stress. How to rehearse difficult conversations. And perhaps most importantly, how to change one’s behavior through experience, not just advice.

“If we can evoke a cognitive state,” Greenleaf said, “then we can teach someone the skills to manage that state.”

VR makes this possible. Whether it’s public speaking anxiety, PTSD triggers, or social challenges for people on the autism spectrum, immersive environments can simulate these situations with precision—and allow patients to practice new responses in real time. It’s a safe space to fail, adjust, and try again. Over time, this repetition becomes wiring. The brain reconfigures itself to meet the challenge.

This is a radical upgrade from traditional coaching or therapy, which often relies on memory, imagination, and verbal reporting. In VR, patients experience the thing. And then they change.

Emotional Data, Finally Quantified

For decades, mental health has operated in a kind of emotional darkness. Clinicians ask patients how they’re feeling or how a medication affected them a week ago—and they rely on subjective, often unreliable answers.

Greenleaf believes that’s about to change.

“We finally have tools to objectively measure emotional state,” he said. Through voice tone, eye gaze, body language, and passive biometrics, immersive systems can now detect emotional shifts in real time—and adapt accordingly. These systems act not just as display screens, but as biosensors.

This breakthrough opens the door to a revolution in psychiatry, psychology, coaching, and personal wellness. Emotional state can become visible. Interventions can become dynamic. The invisible becomes trackable, and therefore, improvable.

It’s like going from sailing without a compass to having a full navigational dashboard. “We’ve kind of been flying blind,” I said during the conversation. “But this is our radar. Our telescope. Our map.”

The Empathy Machine, Upgraded

Beyond therapy, Greenleaf sees VR as a way to deepen connection—especially through empathy and perspective-taking. One of the most compelling use cases he shared came from the Stanford Virtual Human Interaction Lab, where users interact with an avatar of their future self. That avatar can give you feedback—encourage a better choice, or show you the long-term consequences of today’s behaviors. It’s a powerful tool for behavior change, but also for self-compassion.

There are also use cases for interpersonal empathy: understanding what it’s like to be homeless, or aging, or recovering from trauma. But Greenleaf is clear: these experiences must be designed with care. Without agency or context, VR empathy can backfire—leaving users overwhelmed, numb, or disengaged.

The right design includes agency, positive feedback, and a clear path to action.

Digital Wellness by Design

Of course, immersive tech still raises concerns. What about screen addiction? What about the fear that we’re escaping life instead of engaging with it?

Greenleaf doesn’t dismiss these concerns. In fact, he shares them. But he believes the solution isn’t to reject the technology—it’s to design it better.

The problem isn’t screens, he argues. It’s poorly designed digital experiences that flatten emotion, reduce connection, and addict rather than inspire. He envisions a future where digital experiences are emotionally intelligent, adaptive, and pro-social—helping people learn, connect, and thrive.

In this vision, AR glasses won’t isolate us. They’ll serve as context-aware companions—warning us when we’re agitated, encouraging us to pause, offering guidance, or simply helping us understand ourselves better.

As Greenleaf put it, “It’s not about escaping the real world. It’s about deepening our experience of it.”

The Road Ahead

So where are we now?

Greenleaf sees early traction in places like the Veterans Health Administration, where VR is being used to treat PTSD at scale for over 30 clinical indications. But the broader healthcare system still faces hurdles—hardware friction, a lack of interoperability, and outdated business models that don’t yet know how to reimburse for immersive care.

Still, the technology is ready. The science is sound. What’s left is better design, broader adoption, and thoughtful regulation.

For those of us building the future—whether in AR, AI, or health tech—Walter Greenleaf offers both a challenge and an invitation. The challenge: design wisely. The invitation: build tools that make us more human, not less (aka the Medium Energy ethos).

If we get it right, AR/VR won’t be a form of distraction or escape. It’ll be a place where we rehearse who we want to become—and then return to the real world better equipped to live.

With that, I bring you Dr. Walter Greenleaf.