Robot Maids in Every Home Soon?

Chapters: 0:00 The robot housework test · 1:03 Testing humanoid intelligence · 3:23 The role of the Index app · 5:17 Scaling laws in robotics · 8:03 Reality check and limitations · 9:37 Investment and market analysis · 11:30 The future of general robotics

0:00 This robot just walked into a house it never seen before. Nobody showed it the house. Nobody trained it there. It cleaned the living room. It folded the towels and it made the bed. Then it did it again in 30 different homes. Everyone is sharing this video, but almost everyone is missing the real story. The real story is not the robot. The real story is a phone app. an app that pays normal people to film themselves doing housework. And that app is just helped a $ 39 billion company to do something robots could never do before. In this video, I will show you what really happened, what they didn't tell you, and the number that has investors so excited. At the end, I will answer the big question.

0:53 Is General Robotics about to be sold? My answer might surprise you. I'm Anie Bergstrom and this is AI for finance. Let's go. The company is called Figure. They build humanoid robots with a body like ours. Two arms, two legs, two hands. In September, they show the new AI brain for their robot. It is called Helix 2.5. Helix is the part that looks at the world and decides how to move. Here is the test they did. figure rented 30 real homes in California. Real sofas, real beds, real mess. The robot got three jobs. Clean up the toys in the living room, fold the towels, make the bed. And there were three rules. Rule one, no training data from these homes. The robot had never been inside them. Rule two, the toys, towels, and the bed sheets were all new to the robot. Rule three, the same robot brain in every home. no changes after it arrived. Think about hiring a cleaner. You don't train them for 6 months in your apartment.

1:59 They walk in, they look around, they start working. Humans do that. Robots could not until now. The test was also very strict. In the living room, there were 13 to 15 toys on the floor. The robot had to put every single toy in the basket. Miss one toy, that is a fail. No half points. And one thing really surprised me. When the robot made a mistake, it fixed it. It stepped back. It moved its body. Sometimes it walked around the whole bed and it tried again. For a robot in a new room, that is huge. No. You might think making a bed that's easy. For you, yes. For a robot, no. In 1997, a computer beat the world champion at chess. Almost 30 years later, we still get excited when a robot folds a towel. Why? Because for computers, the easy things are hard.

2:59 Chess has clear rules. A towel does not. Every time you drop a towel, it lands in a different shape. Every bed is different. Every home is different. You cannot write a rule book for that. And the humanoid has another problem. It has to walk to see things, then move its body to reach them, then use both hands at the same time, all at once. Now, here is the business part. Until now, robots had to learn every new place from zero. New warehouse collect new data. New customer collect new data again. So, every new place costs more money. If you have worked in it, you know this problem. It does not scale not to millions of homes. So the real question was not can a robot make a bed. The real question was can a robot learn once and then work anywhere. And the answer came from a surprising place your phone.

3:58 Chad GPT learned by reading the internet. Billions of pages already written almost free. But robots have a problem. There is no internet of movement. Nobody uploaded millions of videos of how to fold a sheet filmed from their own eyes. Figure tried to buy this data. It was not good enough. So they built their own. In August they launched an app. It is called index. Here is how it works. You download the app. You film yourself doing real tasks. Cooking, cleaning, laundry, and figure pays you. When they launched, figures said the app already had 264,000 downloads, people in 108 countries, more than 16 million videos, and $15 million already paid to users. Every second, 30 minutes of new video came in. Every single day, almost 5 years of human work. A few weeks later, it was 35 minutes of video every second. Let that sink in. The top humanoid robot company in the west is [music] also running a geek economy up. People get paid to film the housework that robots are learning to do. Remember that. I will come back to it. Now the numbers from the start. This is the most important test in the whole release.

5:22 Figure trained two robot brains. Same design, same training on the task, same test. Only one difference. The first brain started from zero. The second brain first watched all the human videos from the app. Then they tested both in the 30 homes. The brain that started from zero succeeded 9% of the time. The brain that watched humans succeeded 56% of the time, more than six times better. And the whole change was watching humans. It didn't just memorize how to make a bed. It learned something bigger. How a body move through a home. That is the real breakthrough, not the robot, the recipe.

6:09 And the reason investors got so excited is what happened next. To understand this, we need to talk about Chachi PT again. Why could AI companies spend billions of dollars on training? Because of something called a scaling law. It means this, more data plus more computers give you a better AI and you can predict how much better before you spend the money. For a finance person, that changes everything. Research is a gamble. A predictable return is an investment. Now, Figer says they found the same thing for robots. They trained four robots brain. Each one got more human videos up to eight times more. And each time the robot got better in a small, clear line, so clear that they predicted the result of the biggest brain before they even trained it. Now look at the timing.

7:04 August 25, they launched the app. That is the data. September 3, they signed a deal for up to 100,000 Nvidia chips. $3.5 billion, maybe more than 6 billion later. That is computing power. September 17, they show the robot that is the proof. Data, computers, proof in 3 weeks. That is not just science news. That is a story for investors told step by step. And there is one more number. The new brain needed half the training data of the old one. And it worked in 30 homes, not just one. Two times cheaper to teach. 30 times more places to work. Remember the IT problem? Every new place costs more. If this is true, that problem starts to go away. And that is when a robot company starts to look like a software company. if this is true because now I need to put on my finance hat and read the small print. Here are five things the video does not tell you.

8:06 Number one, 56% success also mean 44% failure. In total, the robot finished about 237 out of 420 tries. The best task was masking the bed about 67%. Would you hire a cleaner who fails almost every second home? A demo needs to work once. A product needs to work almost every time. Number two, homes were new, the objects were new, but the three tasks were not new. The robot was trained on three tasks in other places first. So, it is new on place, not new on task. It didn't walked in and decided to do the dishes. Number three, figure tested figure. Their test rules are stricter than most robot demos, but nobody outside the company has checked the results yet. Number four, watch the editing. People online notice that the video cuts away before the task are fully done. That does not mean the numbers are false. It means trust the test, not the trailer.

9:09 Number five, this was three jobs in dirty homes, all in California. Not a small Stockholm apartment, not a messy kitchen, not a factory. And to be fair, figure says this too. They write it clearly. General robotics is not sold. We have seen this before. Self-driving cars looked almost ready more than 10 years ago. The last few% took many years and many billions. Homes are full of surprises. Now, let's talk about money. Figures started in 2022. In 2023, it was worth about half a billion. In early 2024, $2.6 billion. In September 2025, $39 billion. 15 times more in about a year and a half. The investors include Nvidia, Microsoft, Intel, and Jeff Bases. And the revenue very small. So investors are not paying for today. They are paying for the idea that robots will one day do a big part of the human work. That is why this breakthrough matters for investors. A scaling low turns trust us into here is the plan and figure is not alone.

10:21 China sells the most humanoid robots by far around 90% of them. Tesla is building its own robot Optimus, but you still can't buy it. 1X, a company roots in Norway, sells a home robot called Neo, $20,000 for $499 a month. And here is a Nordic surprise. When BMW chooses a humanoid robot for its first test in Europe, it did not choose figure. It choose a robot from Hexagon, a company listed right here in Stockholm. So, how do I look at this as an investor, not as a fund? I watch five things. One, does someone outside figure get the same results? Two, does 56% go up to 90% and more? Three, can one brain learn many tasks, not just many homes? Four, what does one robot hour costs compared to a human hour? In Sweden, human work is expensive, so this math may work here earlier. Five, the price of the next founding round or a stock market listing. There is no listing plan yet. Quick note, this is not an investment advice. This is just how I read the signals. And remember the app, people paid to film the housework.

11:33 Robots will learn. That is not a side story. That is the business model. Human work today creates the data. So robots can do some of that work tomorrow. Is that exciting or scary? Maybe. It depends on which side of the camera you are on. So is general robotics about to be solved? No. Not this year. Probably not next year. Can the robot work in homes it has never seen? Yes, a little. 30 homes, three jobs. Can it learn new jobs on its own? Not yet. Is it good enough to sell? No. 56% is not a product. Does more data make it better? Bigger says, yes, we need proof from others. But here is what I think really happened. For many years, robotics was a research problem. Every new task needed a new idea. Now, it may be turning into a scaling problem. More human videos, more computers, a better robot. Research problems are hard to predict. Scaling problems are expensive, but you can put them in a spreadsheet.

12:40 And when you can put something in a spreadsheet, the money comes in. We saw this with CHP. So this is not the chip moment for robots. It may be the moment before it. The moment the line on the chart appeared, before the product did. Now, I want to hear from you. Would you let a robot with 56% success make your bet or would you wait for 99%. And would you pay $499 a month for it? Write it in the comments. I read everyone. If you want more videos like this where I explain AI through the eyes of someone who builds data systems in finance, subscribe. I'm Annie. See you in the next one. Bye.