In 6th grade, all we wanted was a calculator. In my German school, you got to use them starting in 7th grade, and we were very jealous of the older kids. Why were we still doing math by hand when next year we’d have a machine for it?

When we complained, my math teacher responded: “Yes, there are calculators. Yes, they will make your lives easier. But you have to earn them first.”

He was right, of course. When you do something by hand, you build intuition that the shortcut can’t give you. The calculator is better at computing than you’ll ever be. But it can’t tell you what to compute, or what the result means. You only develop that judgment by doing the work yourself.

Once we’d earned it, of course we used the calculator. That’s the whole point. It shifts the work away from mechanical computation to the interesting part—what the math actually means.

Now I have LLMs. They’re the most powerful tool I’ve ever had, especially for programming. I use them heavily and I don’t want to go back.

But I still remember my math teacher. When I’ve written this type of code a million times, or it’s boilerplate, or the code is just a means to an end—I let the LLM do it. I’ve already earned that.

When I’m learning something new, or doing research, I write the code myself. Not because I’m old-fashioned. Because when I skip the struggle, I don’t build the intuition. I can’t see the shape of the problem. I don’t know what to try next. I don’t notice the questions and choices that naturally come up while writing the code. The understanding is the work, and there’s no shortcut to it.

The trouble is that we’re all in school again, with an amazingly powerful calculator but without a teacher to hold us accountable. Nobody’s going to make you earn it. That’s on you.

I recently came across a GitHub repo that I found fascinating. It’s called ruvnet/RuView; it has ~29k stars and is the #1 trending repo for this month as of this writing. It claims to turn commodity WiFi signals into real-time human pose estimation and vital sign monitoring (the idea is real and based on actual research). It made the rounds on social media. People on Reddit are even asking how to protect themselves against it.

This particular repo, however, does nothing useful. The project’s “pose estimation” is a hardcoded skeleton template wiggled by sine waves, and with a hard-coded walking animation. Before the author rewrote (obscured?) everything in Rust, this was even clearer: The code that was supposed to return data from the WiFi sensor returned random numbers.

It’s essentially a deepfake, but for code. There’s plausible-looking Rust and Python code, a convincing README, all the right buzzwords. It looks real enough to fool thousands of developers into starring it.

This isn’t an isolated incident; the author has over a hundred similar repos on their GitHub, all of them seemingly AI-generated.¹

These repos are symptoms of something more fundamental. The internet always made it cheap to distribute noise—spam and SEO clickbait have been around for decades. But AI has made it cheap to produce convincing noise too. A single person can now mass-produce plausible-looking repos, articles, and images at near-zero marginal cost. And every receiver pays the verification cost—cross-referencing, fact-checking, checking if code actually works—independently.

But AI doesn’t just increase the noise. It simultaneously makes it possible to extract signal at scale. The same technology creates the noise and powers the filter—but only for those who can afford it.

Consider Meta. They’ve invested billions in AI infrastructure for ad targeting. Apple’s App Tracking Transparency (ATT) made it significantly harder to track users across apps, degrading the signal that advertisers rely on. In response Meta built models to infer user intent and behavior from noisier data—and it worked: by Q4 2023 they were reporting record revenue of $40B, up 25% year-over-year, and completely recovered from a brutal 2022.

On the Q3 2025 earnings call, Zuckerberg said:

But any compute that we don’t need for [AI research], we feel pretty good that we’re going to be able to absorb a very large amount of that to just convert into more intelligence and better recommendations in our family of apps and ads in a profitable way.

Meta has essentially built industrial-grade signal extraction for their advertising channel—and they think it will keep getting better the more compute they can throw at it.

ATT, it turns out, was a gift in disguise for Meta. If ad targeting becomes harder due to noise, the companies that are best at signal extraction gain a more defensible moat. The increasing noise actually helps incumbents by raising the bar beyond what smaller players can afford. This is happening for Meta and I expect the same dynamic to play out in financial markets, intelligence, and anywhere else that extracting signal from noisy data confers an advantage.

This asymmetry is structural—it follows directly from the economics of AI deployment. Building and running sophisticated models for signal extraction is expensive while the cost of noise generation is approaching zero. So we end up with a world where only well-resourced actors—large tech companies, governments, sophisticated financial firms—can afford verification at scale. For the average user, the utility of the internet degrades. For smaller companies, it becomes increasingly difficult to compete. This should worry anybody who wants an open and egalitarian internet.

I love robots. I’ve worked on AI and robotics for ~5 years. So naturally I would very much love to have a humanoid robot in my home that does my dishes and washes my clothes and perhaps even does my grocery shopping. Who wouldn’t?

But I really don’t think that we’re close to having general purpose humanoid robots at home. Somehow, in 2026, it seems like that’s a controversial opinion; so I figured I’ll write it down.

Here’s my claim: General purpose humanoid robots are like self-driving cars, but actually much harder. Before I explain why, let’s talk about the good news first.

There are some very impressive improvements in humanoid robotics that have happened over the last 5 years. We do now have:

• Much better, cheaper hardware. Unitree and Boston Dynamics are genuinely doing impressive stuff.¹ For Unitree in particular, prices have also come way down: The Unitree G1 starts at $13.5k and the H2 starts at $30k. 1X lets you pre-order Neo for $20k. That’s a fraction of the cost of the robotics hand that I used to work with, and it gets you an entire robot.
• Multi-modal foundation models and world models. All the frontier labs are now training their models on images and text; Google is also including videos in Gemini’s pre-training. We have models like OpenAI’s Sora 2 and Google’s Genie 3. This is great news for robotics because you can definitely use these pre-trained models as your foundation for embodied AI and probably use the world models to train them on diverse tasks without collecting lots of real-world data.
• Massive capital inflow. People are real excited about humanoids and investors are pouring a lot of money into the field. You have companies like Figure (raised $1B at a $39B valuation), 1X (trying to raise $1B at a $10B valuation), Neura (also trying to raise $1B at an $8-10B valuation) and Unitree (seeking a $7B IPO). That’s a lot of capital. Tesla is also investing heavily into their Optimus humanoid robot, including re-purposing their Fremont assembly line to manufacture humanoids.²

But building general purpose humanoid robots remains extremely difficult. The closest thing we have is self-driving cars—those are basically autonomous robots on wheels. And since we’ve worked on self-driving for a while now, we can look at how that went.

Well, self-driving turned out to be really hard, and it’s still not fully solved. Waymo is arguably mostly there after 17 years and billions of dollars, but only in a handful of cities, and they still need to do things like hiring DoorDash drivers to close car doors. There’s also a long list of companies that tried and failed: Cruise, Uber, and Argo AI. The tail of problems is long, you know.

And yet, self-driving is actually the easier problem. Compared to humanoid robots, self-driving cars have significant structural advantages:

• Mature hardware. Humans have built cars for more than 100 years. Modern cars are incredibly reliable and safe.
• A constrained environment. Cars drive on public roads. Public roads are heavily regulated environments. Of course there’s some level of chaos and a wide range of possible events, but the environment overall is very structured.
• A single, well-defined task. Self-driving cars need to go from A to B. Driving is complicated because you need to handle complex and diverse situations, not because there’s diversity in the task domain.
• Self-driving is a feature, not a requirement. A regular car that you drive yourself has a lot of utility. You don’t need it to drive itself for it to be useful.
• Scalable data collection. Because of the aforementioned property, there’s a very scalable path to data collection: Record what the human drivers do day-to-day. There’s no need to hire or pay someone for this; this data can be naturally collected as a by-product of the intended use.

Now, contrast that with general purpose humanoid robots:

• Immature hardware. Yes, there was progress here but we have not built and deployed humanoid robots for 100 years at scale. They will be unreliable. They will break a lot. Somewhat tellingly, 1X’s Neo is already broken in the demo they did for Joanna Stern at the Wall Street Journal. The hardware problems here are far from solved. Also, there is no infrastructure for maintenance and repairs yet, which is ubiquitous for cars.
• An extremely open-ended environment. You’re deploying these things in people’s homes. Homes are extremely different: Some of them are very small, some are very large, some are neat, some are messy, some are colorful, some have multiple stories. There are also very few rules and regulations around how people set up their homes. Also if you want your robot to go to the grocery store, you have to deal with public roads and the store as well. That’s a lot!
• A very large variety of tasks. Today I want you to do my laundry. Tomorrow, can you cook me a dish? Also can you vacuum the place, pick up some groceries, fetch my mail from the mailbox and maybe walk the dog. The breadth of tasks a general purpose robot is expected to do is really, really large!
• Full autonomy is table stakes, not just a feature. In contrast to a car, the robot itself is utterly useless if it’s not automated. Imagine buying a robot for $20k and then doing the dishes yourself by teleoperating it. Yeah, nobody wants to do that. The whole point of a robot is that it does something that I do not want to do myself. So you have to either a) have it work autonomously via AI or b) you pay a person to teleoperate the robot in your home on your behalf. There’s also a social aspect here: The robot operates in an environment with humans and therefore has to adapt to social conventions and behaviors, which raises the bar even more.
• Data is very scarce. This is the big one. Tesla collects driving data from every car on the road—millions of miles, for free, as a by-product of people just driving. That data advantage is real; it’s a big part of why FSD has gotten as far as it has. Humanoid robots don’t have this. The robot is useless without AI, so there are no human users generating data. You either pay someone to teleoperate the robot (doesn’t scale), you pay someone to collect human demonstrations,³ collect data from the robot’s own autonomous operations (only works if the AI is already decent), or rely on simulation. World models might be a big unlock here, but we don’t know yet. For homes in particular, there are also obvious issues around privacy.⁴

To be fair, humanoid robots also have some advantages: There are fewer laws and regulations around their deployment, the cost of individual failures is lower (a dropped plate vs. a car crash), a home robot can pause and ask the human for help in a way that a car on a highway cannot, and perhaps most importantly, perception and foundation models are dramatically better today than when Waymo started in 2009.

That said, I think these advantages are more than offset by the challenges above, especially the data problem. So, all told, the odds still seem really stacked against general purpose humanoid robots happening anytime soon.

But that doesn’t mean nothing is happening in home robotics. Look at how much robot vacuums have improved over the years; Roborock is now even attaching an arm to their flagship model. Robot lawn mowers from companies like Husqvarna and Mammotion are now a thriving market. These work precisely because they sidestep some of the problems above: simpler and more mature hardware, well-defined tasks, and existing deployment fleets that generate data. Less exciting than a humanoid butler, sure. But they’re real and they’re shipping today. I think these companies have a real shot to scale to much more powerful home robots over time. They also ask a key question: Do we really need the humanoid form factor?

There are also companies like Physical Intelligence and Generalist AI who are building foundation models for industrial robots. This might turn out to be another viable path: Start with proven industrial hardware and operate in a more constrained environment. Then gradually scale your capabilities.

If you’re working on general purpose humanoid robots: godspeed. The problems I outlined above are real and they’re hard. The key challenge is to figure out how to overcome these structural deployment and data flywheel issues. But someone has to try, and I hope you succeed!

In the meantime, let’s stay level-headed. The vision is exciting. I want it. But wanting something doesn’t make it close. And a cool demo does not yet make for a real product.

I have a confession: I’m a bit lazy. Not dysfunctionally lazy. But definitely a bit.

Early in my career, I thought this was a problem. There’s so much more I could do if I weren’t lazy! But, over time, I realized it’s the opposite: My laziness is an asset. It’s a regularizer.

When you write code, you have an infinite blank canvas. You could build anything. This is pretty distinct from lots of other domains: Code is inherently less constrained than, say, the work of a mechanical engineer because you don’t have to fight the constraints of physics every step of the way (of course a computer is still physical, so there are some real constraints; but you get the point). This is fantastic. But it’s also daunting.

Enter AI agents. These things are great and I use them all the time. The amazing thing is you can build anything now. The trouble is also that you can build anything now. They never get tired, they never get frustrated, they are infinitely patient and motivated. In other words, they reduce the execution cost of building something towards zero.

So a natural question arises: What should you build? This was always a question, of course, but AI agents make it a lot more urgent.

When it comes to deciding on what to build and how, taste is an important concept. A lot has been written about this. For example, there’s Paul Graham’s famous “Taste for Makers” essay and John Schulman’s excellent guide to ML research, which identifies taste as a critical skill as well.

But what they don’t talk about is laziness. Taste is about quality and direction and simplicity and beauty. Laziness is about restraint and friction and effort. Laziness regularizes taste—taste without laziness gives you beautifully overengineered systems nobody needed. Also, tastelessness without friction is by far the worst: that’s slop.

Remember when you had to write all code and do all research yourself and you went “ugh”? Because you really didn’t feel like doing something? That’s healthy! That “ugh” carries information. It’s telling you: “This is going to be work.” And it’s asking: “Is this work really worth doing or should I just… not?” The answer might be yes or no, but it really matters that this question gets raised in the first place.

AI agents remove this natural friction. They are so fast. They never stop. But they have a price: You have to remember your very human laziness. You have to remember to ask yourself: “Should I really build this?”—or should you just be lazy today.