Meta's Humanoid Robot Max and the Race to Put AI Bodies in Homes and Workplaces

The video opens with reports that Meta has built a humanoid robot called Max. According to the report, Meta sent a team to Shanghai, to the factory of a Chinese firm that makes robots like this. The robot took just three months to build, and a second version with better hands and sensors is already in development.

The presenter adds that one day Max could become the body for Meta's all-seeing AI assistant, Muse.

The presenter notes that Big Tech is spending hundreds of billions of dollars on AI infrastructure. To make that pay, many believe AI must leave the screen for the real world, with examples such as a robot in the home or a robot in the workplace.

The segment poses the question of whether this is the future or Big Tech's biggest gamble yet.

Ed Zitron, who hosts the podcast Better Offline, says Meta is not building a robot. In his view, the company is doing more ancillary work for its large ad-targeting operation and is not becoming a robot company. He describes the story as one about fooling the media and investors into believing Meta is anything other than an ad-targeting company connected to a social network.

Zitron says Meta sent an engineer to China for two months to work on a robot, and he dismisses that as unimportant. He points to a graveyard of dead projects and acquisitions, citing Instagram and Oculus as examples, and says Meta's glasses and Meta Muse lose money. He calls the robot effort a complete waste of money and time.

Zitron argues that large language models are inherently probabilistic and make mistakes by definition. He says OpenAI said in October last year that they are mathematically certain to hallucinate, which makes controlling a robot with them a problem.

He uses washing dishes as an example. Picking up a dish is not enough; a robot must measure pressure in nuanced ways so it does not crush the bowl, rip the sponge in half or squeeze out all the washing-up liquid. He says Meta will not solve these problems by connecting an LLM or through training data, calling them massive scientific efforts. He adds that impressive demos often show an articulating hand, because miniature motor movements are so difficult.

Zitron also says machine learning is trained for the 99 percent of cases but fails on the 1 percent edge cases, and real life is full of them. He cites his own coordination disability, dyspraxia, and says a slightly mistrained robot could crack a table or break things, which would be unacceptable for a product. He also says Meta has changed its AI team four or five times in the last year.

Stephanie Hare says she thinks Zitron is right in one sense: we have bad versions of everything, and robots are probably at the early adoption stage. She compares them to mobile phones that did not work very well and to bad early versions of computers.

John Hanke says he thinks humanoids are inevitable and that they are not here yet. He says he grew up with bad versions of technologies and enjoyed using them before they were ready for prime time, citing early smartphones before the iPhone and painful early dial-up internet. In his view, these were glimpses of what was coming, and seeing early humanoids is a window into the future.

Asked why robot companies are interested in the human form when there are myriad ways to build a robot, Hanke says many robots today do boring things that do not look like humans, such as arms in factories. The vision in people's minds, however, is of C-3PO or robots seen in movies.

Hanke says there are good reasons to build something human-shaped because the world itself is built for humans. A machine that can access things, turn knobs, reach objects and move them as a human would gains many advantages from similar reach and mobility. He acknowledges that this is energy-inefficient in many ways, but says compatibility will be important for some jobs.

The presenter describes Hanke's career, from Pokémon Go and Google Earth, which were about how humans see the world, to a question of how machines see the world and what maps AI-driven robots will need to navigate.

Hanke says maps for humans are encoded to match how people perceive and navigate space, a skill learned at an early age. A computer does not need that encoding. Mapping natively for AI can capture a much richer depiction of the world, including the semantic contents of an area, meaning the objects and materials present. For grasping, a robot needs to know the texture, shape, resistance and weight of an item. So a 3D map should capture not only the geometry of the environment but also its materials, so a robot can predict how the world will react to its motions and actions.

The presenter explains that a home robot would need cameras for eyes and microphones for ears, and would map every room to supply information to its processor. To be truly useful, it would have to learn everything about people's homes and habits. The presenter wonders whether that will be the industry's problem and asks whether it is really ready for that.

Zitron says the conversation should not start there, because it suggests the industry is near doing this. He describes his own Matic cleaning robot, which is covered in cameras and mostly cleans and wipes floors. It cost $1,800, and he says he received it as a review unit. Even that narrow task took years to get done, though it was assumed to be a solved problem.

Zitron argues the difficulty is not only what the robot sees in one home but how many homes it has seen. He says millions and millions of hours of training data are needed to do basic things and handle edge cases such as different floors, stairs, and objects like cats, kittens and Munchkin cats. He also notes that homes differ: English homes are much shorter than American ones, and some American McMansions have 12, 15 or 20-foot ceilings. He says this is a training-data problem plus the mechanics of turning that data into something useful.

Zitron adds that robotics conversations keep starting in the middle, and he does not think early days is a fair description. He says headlines about shipping thousands of humanoid robots often describe a robot doing the world's worst dancing and then falling over, which people call progress.

Zitron says the tech industry is jingling keys because it knows it is out of hyper-growth ideas. He says the AI industry, including OpenAI and Anthropic, is trying to bring up robotics so people who dislike AI will also dislike robots, even though large language models are very different. He concludes that he considers all of this marketing because he does not think it is real right now and believes the technology is far away.

The presenter describes a video of six humanoid robots from one of China's other big robot makers, Agibot, working in a convenience store. The store was live-streamed for 24 hours. The robots stock shelves, make coffee, greet customers and take out the rubbish.

The presenter says what is not known is how much human involvement there is in what the robots are doing, and asks whether this could be a forerunner to a humanoid in the home, since a convenience store is a cluttered environment.

Hanke says a convenience store is a pretty good early training ground for something that might work in the home in the future. He notes that there are lots of objects and lots of clutter, which are difficult for computer vision to map accurately in real time, and for a robot to understand and interact with.

The presenter notes Zitron's point that robots need a lot of training data. Hanke agrees and says one of the tricks is to do much of the training in simulation, if the simulation is good enough. He describes the sim-to-real gap, meaning simulation environments have been lacking in many ways, which has limited what can be done.

Hanke says his team has worked to make simulation environments much richer, both in the visual depiction of the scene, so computer vision can train on a simulated 3D environment, and by including solid surfaces so movement can be accurately trained. Adding semantics and other material understanding makes the simulation even better.

The presenter asks about the difference between a Gaussian splat and a mesh on screen. Hanke says Gaussians give an accurate and very accurate visual representation, similar to what a camera or human eyes would see. This gives the robot's camera an accurate model for training its behaviours.

The mesh shows traversable and hard surfaces, meaning where the robot can safely go, where it can stand and where it can place an object. Hanke says both representations capture different elements of the scene, and both are needed to do useful work in an environment.

The presenter reports that in July the US banned imports of new foreign-made humanoid robots, a market China is sort of cornering. The presenter notes a fear that such robots could spy on Americans or work remotely for Chinese intelligence agencies, and asks whether a robot should be treated as a security issue.

Hare says this is a good question. On one hand, the move could be a knee-jerk reaction by the administration, which, she notes, has a certain view on China on both sides of the aisle in the United States. From a strictly cybersecurity perspective, she says any machine that can be hacked can potentially be weaponized, offensively or defensively. She cites stories in Europe of hacked refrigerators that led to millions of dollars of food being thrown out, and notes the same worry exists for self-driving cars, where hacking could weaponize a vehicle or take the data it gathers. She concludes that it is not entirely crazy to want to regulate and control robots, but that the idea a Chinese robot is more dangerous than an American one is probably more of a political question.