AI drives service costs to near zero

Author | Wang Lei

Editor | Qin Zhangyong

"The number of robots in the future will far exceed that of humans!"

"I have to intentionally ignore the potential threats posed by AI to sleep well."

"The pace of AI technological advancement is much faster than any other technology I've seen."

Musk's latest interview is filled with jokes, self-expression, and even crude language. It also touches on sensitive topics like freedom of speech and regulation.

On September 10, the annual All-in Summit was held as scheduled, and Musk, a regular guest at the summit in recent years, once again became the center of attention.

Musk discussed the profound impact that humanoid robots and autonomous driving could have on the economy and social life in the future, under the umbrella of AI. He believes that the bright future of AI will usher in an era of extreme prosperity, where material scarcity will become a thing of the past.

However, risks coexist. Musk warns of a 20% chance of destruction but overall paints a picture of a future filled with technological wonders and unprecedented prosperity for those present.

Enjoy the following:

01 "Great Harmony in the World"

"I believe that useful humanoid robots represent the greatest opportunity mankind has ever had."

Musk expresses unexpected optimism about the future of the struggling Optimus project, believing that robots will eventually replace humans in various tasks.

In his view, humans are essentially "meat" versions of general-purpose robots. Through designing the Optimus, he has gained a deeper understanding of why humans are shaped the way they are.

For example, why we have five fingers, why the pinky is shorter than the index finger, and how the primary muscles of the hand are actually located in the forearm, with fingers operated mainly by tendons.

The current Optimus robot hand has only 11 degrees of freedom, while the human hand has approximately 25. The next-generation Optimus hand prototype has already been developed, with the muscle system moved to the forearm and operated by cables similar to human hands, giving it 22 degrees of freedom.

"We believe this is sufficient to accomplish almost everything humans can do."

Regarding costs, Musk firmly believes that mass production significantly reduces costs, approaching the material cost of the product. This means that even humanoid robots like Optimus will eventually have costs that approach their raw material value. It may take a decade, but the cost of Optimus will eventually be lower than that of a small car.

After three significant iterations and reaching a production scale of one million units annually, Musk estimates the price could be around $20,000. Two years of development have already been completed, with subsequent iterations expected to take less than two years each. This suggests annual production could reach one million units in five to six years.

At that point, according to Musk, every person on Earth will be able to afford a robot companion.

Musk also envisions the next 30 years, predicting that the number of robots will far exceed that of humans, with a potential ratio of 2:1 or even 3:1 humanoid robots to humans. This scenario may exceed the imagination of our generation but will be the norm for future generations.

Imagine a future where multiple robot assistants are available, "They can help you walk your dog, mow the lawn, babysit your kids, and even teach them. There will be many invisible robots producing goods and providing services."

Robots could also be sent to distant Mars, ushering in a new era of interstellar exploration and colonization. In Musk's view, Mars is already a "robot planet," filled with machines like rovers and helicopters, where only robots reside.

"Therefore, I believe that useful humanoid robots represent the greatest opportunity Earth has ever had."

02 20% Chance of Destruction

Regarding the future of AI, Musk believes it will usher in a "golden age" for humanity, an era of abundance where goods and services will no longer be scarce.

Everyone will be able to afford what they desire, turning once-unattainable dreams and personalized services into everyday realities. However, Musk acknowledges that this future is not without risks, estimating an 80% chance of extreme prosperity but also a 20% chance of destruction. "I have to intentionally ignore the potential threats posed by AI to sleep well at night," he admits.

It's worth noting that Musk made similar comments at last year's All-in Summit: "AI may be the most important technology humans have ever created, but it could also be more dangerous than nuclear weapons."

Even when intentionally ignoring AI's potential threats, Musk finds it difficult to sleep at times. He believes a more pressing question is "finding meaning in a world where AI can do things better than us," posing a greater challenge.

However, this presupposes that AI's development must overcome crucial factors like humanoid robots and autonomous driving technology.

"Once we have these humanoid robots and autonomous driving technology, we can theoretically produce anything. Economically, the size of the economy is essentially the productivity of each individual multiplied by the population. If humanoid robots can work intelligently and their numbers are unlimited, there is no true upper limit to the size of the economy."

In other words, despite the 20% chance of destruction, the immense benefits of AI's rapid development may outweigh the risks, making it worthwhile to embrace the technology first.

03 Becoming NVIDIA's Competitor

As mentioned earlier, Musk believes that AI's breakthrough depends on two crucial technologies: humanoid robots and autonomous driving.

However, the core of these technologies lies in the development of chips and computer clusters.

At the All-in Summit, Musk proudly introduced xAI's Colossus computer cluster, calling it the most powerful AI training supercomputer in the world, currently housing 100,000 NVIDIA chips and set to expand to 200,000 by year's end.

This supercomputer provides a solid foundation for the autonomous driving system FSD and humanoid robots.

Many believe that the economic value of AI largely flows to NVIDIA through chip sales, and Tesla still uses NVIDIA chips. However, Tesla plans to soon introduce its alternative, the Dojo supercomputer chip, specifically designed for image and large-scale video processing.

Musk intends not only to use Dojo but also to rent it out, directly competing with NVIDIA.

"At Tesla, we have two projects: the Dojo training computer and the inference computer in every vehicle. We currently have Dojo 1, and we expect Dojo 2 to enter mass production by the end of next year. We believe Dojo 2 will be comparable to existing high-end training systems like NVIDIA's B200. In the future, we may offer it as a service."

Musk believes any system requires three significant technological iterations to achieve optimal results. This year marks the first iteration, with the second (Dojo 2) expected next year and the third (Dojo 3) potentially in 2026. Only then will the full potential of the Dojo series be realized.

Despite his confidence, Musk admits there's no guarantee of success. Nevertheless, the development of the Colossus supercomputer significantly enhances Dojo's chances.

However, Tesla's Dojo faces different challenges in autonomous driving scenarios compared to large language models. Tesla's context is extremely long, with billions of bytes of data.

For example, Tesla vehicles have seven cameras, and one minute of HD video from a single camera generates several gigabytes of data. Compressing this massive context into manageable pixels is a significant challenge for Musk.

Tesla's FSD requires continuous data collection and processing, identifying and classifying objects around the vehicle, all while making decisions at human-like speeds. These steps must be executed seamlessly, akin to human cognition.

The difficulty lies in selecting key frames from the vast video data captured by multiple cameras and compressing them into a range manageable by the vehicle's onboard inference chip, which consumes only a few hundred watts.

"You must compress pixels in both the spatial and temporal dimensions, then make inferences on a small computer that consumes only a few hundred watts."

Musk takes pride in Tesla's chip design team, believing their inference computer surpasses any product bought from suppliers.