Li Xiang: AI can change the physical world|36Kr interview

Text | Li Qin Li Anqi 

Edit | Li Qin

On March 2, Tesla released the third chapter of the "secret grand plan" on the investor event day. On the same day, Li Auto also held an unprecedented communication meeting. Musk trumpeted "we are investors in the planet" to outline a grand final chapter of a sustainable energy economy for the smart electric vehicle industry.

Li Xiang, CEO of Li Xiang, was more pragmatic and talked for more than 2 hours, sharing why he founded Li Auto, the starting point for extended range, how to layout pure electric and supercharging networks, how to do business management, build a supply chain and enter the era of intelligent electric vehicles.

In fact, Musk has used three secret grand plans in the past 17 years to open up three major business scenarios for the smart electric vehicle industry: car, mobility and energy. At present, domestic car companies are still struggling to build and sell good cars. The ideal of relatively robust physique has begun to initially look forward to the future of smart travel.

Regarding technology, Li Xiang bluntly said that the company's self-driving technology path is no different from Tesla's FSD. When it comes to energy, Li doesn't think much about it, mentioning more about "energy security" and the company's practical thinking in this context.

Source: Ideal Car

Li Xiang is still sharp and outspoken. "Consumers do not have any obligations, when I buy an electric car, the entire experience, the radius of exercise, the efficiency is lower than that of fuel vehicles." Talking about how to promote fuel vehicles to new energy vehicles, Li Xiang made this decision clear.

Based on this, Li Auto will still adhere to the two sets of technical solutions of extended range and pure electric by 2023. Around the pure electric layout, Li Xiang also has a clear plan: 800V high-voltage + 480V high-power charging pile to achieve the energy replenishment effect of fuel vehicles.

This set of ideal cars has long been publicized, but this time Li Xiang also gave a detailed decision-making model. First of all, the 800V high-voltage platform and silicon carbide power chip, combined with better wind resistance coefficient and overall efficiency optimization, can greatly improve the endurance, thereby saving costs.

"Compared with today's traditional 400V mainstream electric vehicles, under the same size and the same drive form, it can probably reduce the cost of parts by 30,000 to 40,000 yuan for one vehicle." Li Xiang said, "Everyone thinks that 800V silicon carbide is a more expensive price, but it is not. ” 

Li Xiang's positioning of the charging pile is that this is a product integrated with electric vehicles and must be built. He calculated that a charging station of more than 640 kVA requires about 1 million yuan of investment, and the construction of 1,000 stations is 10 billion yuan, amortized for 5 years, 2 billion per year, which does not pose pressure on a company with an annual revenue of 100 billion yuan.

And high-power charging, will also shorten the charging time to 10-20 minutes, which will make the charging pile from the past 2-3 orders per day, to 5-6 orders, to achieve breakeven, "if you can achieve 10 orders a day, the profit will exceed the gas station." ”

Sustained investment requires a healthy financial model. "20 points is the minimum gross profit margin that a company can develop healthily for a long time." Li Xiang said, "The higher the gross profit margin, the more difficult it is to operate, the lower the gross profit margin, the easier it is, and the negative gross profit margin is equivalent to sending money, which is the easiest." ”

In addition to the electric layout, the ideal car is also deepening the investment and application of intelligent technology, including not relying on high-precision maps, and landing in the city NOA by the end of the year. Li Xiang's judgment is that by 2024, whether it is equipped with high-speed NoA will begin to affect the consumption decisions of mid-to-high-end vehicles, which will also open the curtain of the era of smart electric vehicles.

The industry has changed drastically, and competition has increased significantly. Since the beginning of this year, Tesla has launched four consecutive rounds of price cuts around the world, with a reduction of up to 40,000 yuan, and domestic old car companies such as BYD have also followed up with 10,000 yuan price reductions. The price war between leading companies has made the industry smoke.

Li Xiang admitted that this is a manifestation of the huge pressure in the industry, "in one industry, very few two leading companies can fight a price war".

Although Tesla cut the price to less than 300,000 yuan, making some room for the ideal car, Li Xiang was very clear, "This is our lucky place, but we will not always be lucky, and we have to be prepared." ”

The following is the sharing and exchange content of Li Xiang, CEO of Ideal Auto, which has been excerpted:

The better the car sells, the more expensive the battery is likely to be

Because smart electric vehicles are one of the few industries that involves both energy and cutting-edge information technology, this is an important opportunity for us to see.

From the user's point of view, the demand for energy is actually three levels. The first level is the convenience of energy access, we also often see that in the forum, in Weibo, in Douyin different electric vehicle users will give a completely polar evaluation of their electric vehicles, that is, those with charging piles will say that the electric car is really fragrant, and they can never go back after use, and many users without charging piles, including users without electric vehicles, often run long distances, expressing regret why they bought an electric car.

Source: Ideal Car

The second point is the cost of energy acquisition for users, but the cost of energy acquisition actually has two aspects, one is whether the product I bought is more expensive, so here is the composition of the battery price.

Another is actually the cost of use, the cost of electric vehicles is very low, I think this is also the past so many years, especially in the current economic background, there will be more and more consumers choose electric vehicles, because the money you can save every year can buy one more iPhone, you can buy a new iPad, this is actually saving for the entire economic growth is very critical after slowing down, so this is the second user demand.

What are the needs of the third user? That is, the user's demand is comfortable, environmentally friendly and zero pollution. What do users think is environmentally friendly? In fact, for example, we drive a car to pick up our children from school in the summer, if I turn on the air conditioner, there is smell everywhere, this is pollution, and I drive an electric car with the air conditioner on.

On the other hand, in the government, there are three needs:

The first aspect is the autonomy and controllability of core technology; The second point is the security of energy supply, in fact, whether it is our oil production, lithium carbonate production, is different from our own actual shipments and usage, we produce more than 60% of the world's fuel cells, but our lithium carbonate energy storage is less than 10%, we use more than 20% of the world's oil, but our oil reserves are about 2% less than 3%; Third, carbon peak and carbon neutrality, has also created China's entire electric vehicle industry, the entire industry chain, including motors, electronic control, and then the third generation of semiconductors, including the following batteries, new battery technology, and the whole industry chain like photovoltaics.

Then back to a car company, then we feel the same as in 2016 when doing the ideal ONE project, we face all the challenges on two, the first challenge is the problem of charging difficulty, I mean the charging difficulty is to replace the standard of fuel vehicles to carry out charging, because our consumers do not have any obligations, when I buy an electric car, the user's entire experience, exercise radius, efficiency is lower than that of fuel vehicles.

Then I think the second challenge is the high cost of batteries, because batteries are composed of commodities with core precious metals, so what kind of phenomenon will it present? That is, the better the electric car is sold, for example, the stronger explosive growth began in the second half of last year, so it will become more and more expensive; In January and February this year, almost every company set a full-year target of only 4%-5% in the past two months. At this time, the price of lithium carbonate dropped significantly, from 550,000 tons to 390,000 tons.

We judge that the overall cost will definitely not fall to the original 4-50,000 yuan a ton, but 200,000 to 300,000 a ton is a long-term stable price. In this way, the cost of batteries still cannot be greatly reduced, and even today's battery costs are much more expensive than in 2018.

Charging pile business: 10 orders per day, more profitable than gas stations

From the very beginning, we analyzed that there are three paths to solve this problem, the first path is also the power exchange I did when I was a director of NIO. In fact, power exchange is a very good way, because the experience of the first power change can be infinitely close to the experience of refueling. Second, power exchange allows consumers to buy an electric car at a lower cost than a fuel vehicle.

Of course, power exchange will also have to prepare more batteries for the construction of power stations, including the risk of large fluctuations in battery price increases and price reductions. We didn't have much money at the time, so we were wondering if we could put it directly on the product side instead of relying on the server side, so we made an extended-range electric vehicle.

In addition, the third path is the charging pile. The vast majority of today's charging piles are not profitable, because the business of charging piles is easy to calculate, that is, about 6 to 7 single is 6 to 7 times a day, basically break even, if you do 10 basically a charging pile The profitability of a charging pile will be better than the profitability of the gas station.

However, nowadays, the common 2 to 3 orders are basically losing money, then the common way to do it is that this charging pile can only be used for 5 years, but I allocate it to the cost of 10 years, and reduce the loss in this way.

Users need a lot of waiting time costs to charge, so it is common to have 2 to 3 orders a day. If the charging can be shortened within 20 minutes, the user will not leave the car, and if it is shortened to 10 minutes, the user experience is basically the same as that of the fuel vehicle.

Electric vehicles are not as expensive as you think, and there will be extended range in 2030

In order to achieve the charging experience and fuel vehicles, in fact, there are these two cores, the first core is that the high-voltage platform must be used to achieve such a charging speed, the second advantage is that when we use silicon carbide, with the high-voltage platform, and then with a better wind resistance coefficient, the battery cost can be greatly reduced, that is, the efficiency can be significantly improved.

For example, the best selling medium and large SUV at present is about 100 degrees of electricity, to achieve 600 kilometers, but in fact, we use an 800-volt high-voltage platform, with better drag coefficient and silicon carbide and overall efficiency optimization, we can probably use 80 degrees of electricity to achieve the same mileage, and the vehicle will be lighter, because the car can reduce the use of aluminum after the lighter, and the cost is about 30,000 to 40,000 yuan. Behind this is also related to our own silicon carbide mold and three-in-one motor.

When these things are done well, compared with today's traditional 400V mainstream electric vehicles, in the same size, the same drive form, can probably reduce the cost of parts by 30,000 to 40,000 yuan for a car, so you can see that we launched 800V high-voltage platform electric vehicles, because 4C still has a certain degree of selectivity, probably can achieve the same price as the extended range, which may be different from what everyone imagines, because today everyone thinks that 800V silicon carbide is a more expensive price, but it is not. It can save a lot of costs.

So I think these are our own two core routes, whether we do extended range electric, or do high-voltage pure electricity, in fact, the core purpose is the first to solve the problem of charging, when we use 4C batteries can do 10 minutes to charge 400 kilometers, when we use 2C batteries can do 20 minutes to charge 400 kilometers, this is the core inside.

We think that if we want to launch an electric car, the charging pile is a product for us, not a service, and the charging pile is the product itself. If we provide 4C, but there is no 4C perfect charging network, it is actually equivalent to us buying a 4G mobile phone, but there is only a 2G network, so we think it belongs to the product itself, and we don't have to worry about the speed and determination of us to lay charging piles.

Supercharging stations are not as expensive as everyone thinks, and they are much cheaper than building a factory. I can calculate an account, if a high-speed supercharging station, can pull about 640 kVA to 800 kVA electricity, because the general will do 3 + 1 station, will be 1 480 kW, with 3 250 kW. Because if we put 480 kilowatts, it is useless, and the entire supercharging station can only reach the peak level of 800 kVA. If by 2025, we build 3,000 supercharging stations, the total cost will be 10 billion. According to 5 years of amortization, the allocation to 2 billion yuan per year, for enterprises with annual revenue of hundreds of billions of yuan, the cost is not as high as everyone thinks.

Looking forward to 2030, we will still adhere to the two routes of range extension and pure electric to go forward at the same time, and then make the range extension efficiency higher. On the other hand, it does better in the entire high-voltage pure electricity, providing two needs according to the user's own choice.

Cars are not sold as cheaply as possible

One of the most stressful things in this industry is that the top two companies can fight price wars. This is simply not seen in mobile phones and various industries. And I think it's easy to overlook the problem, that is, we diagnose the problems of various companies, and we find that consumers still have some very basic common sense cognition. Let me give you an example, as a Chinese brand, if we sell a mid-size B-class SUV for more than 300,000, there will be no volume of 100%. That's some very important common sense.

If we look at Tesla and BYD after adjusting the price, they have always put their prices in a reasonable space, not too cheap, but in a reasonable range that meets user cognition. Many people say, why doesn't the ideal reduce the price to less than 300,000? The same problem, consumers do not accept a medium and large SUV of a not bad brand, sold for more than 200,000.

When we were in Autohome, we observed the state of the user's psychological cognition, if he buys a medium and large SUV, it is a good brand and cannot be sold for more than 200,000. Because everyone must think that your car is problematic, you may probably cut corners, which is a user's cognition. There will also be another perception, that is, the perception between users and users, that is, you spend so much money and then buy such a big car, you must be poor. You must have something wrong financially. These are very common sense cognitions, so many times the car is not sold as cheaply as possible, and the car is not sold as expensive as possible. You can draw a curve very clearly in each level.

20% gross margin is the bottom line

Why is gross margin so important? It's still related to the automotive industry, because if you want to become a car company that stays on the card table, you basically have hundreds of billions of revenues, and if you want to become the world's leading company, you must basically have trillions of revenues. How to operate such a scale of revenue? This becomes very critical, can not rely on financing forever, when small billions of income, tens of billions of income can rely on financing, no matter how big how to rely on financing?

And we can see clearly from the history of the development of the automotive industry over the years, when the car is the most difficult to sell, such as the economic crisis, the car is the most difficult to sell, and the financing cannot be financed at all, there will be even large companies like GM and Chrysler, in the 2008 economic crisis, car sales fell by about 40%, and both companies went bankrupt.

In addition to revenue, there is too much leverage, including the separation of companies like Delphi later, and they bear all the leverage, so the 40% decline in sales is almost entirely borne by themselves.

What is gross margin? Gross profit margin is the sales price - cost of sales, cost of sales includes the BOM cost of the car, including the manufacturing cost of the car, the corresponding sharing of the car, as well as the transportation cost of the car, the software cost of the car, and taxes, are the cost of sales of the car. The expenditure of personnel in the store is called sales expenses, which is the part other than gross profit margin, and the part of our gross profit margin is to cope with sales expenses, so the gross profit margin is equal to sales revenue - cost of sales. So after the sales revenue minus the cost of sales, the rest is the money that the company can invest, which is the part that improves its own capabilities.

Back to the perspective of the car, as a smart electric vehicle company, we believe that a healthy threshold is 20 points, our R & D investment is basically more than 10 points, sales management expenses are very good 7 to 8 points, but also to bear certain risks, as well as capital investment, such as building factories. We think 20 points is relatively healthy, Tesla is more than 20 points at present, we are also stable more than 20 points, BYD is also more than 20 points. BYD's sales network is not directly operated, if the dealer expenses are included together, the gross profit margin of BYD cars is also more than 20 points, we think this is relatively healthy, otherwise how to invest in research and development.

It cannot be said that the capital market is good when it invests in R&D, the capital market is not good to shrink R&D, cars are not like this, cars should be more than five years of cycle to do any plan, so gross margin is the key point here. Of course, the higher the gross profit margin, the more difficult it is to operate, the lower the gross profit margin, the easier it is, and the negative gross profit margin is equivalent to sending money, which is the easiest.

Source: Ideal Car

What we see is that 20 points is the lowest gross profit margin that a company can develop healthily for a long time, and you can see that Tesla has been losing money for so many years, but since Model S, Model X, and even from the Roadster, the gross margin has been stable at more than 20%. I can spend a lot of money, but always stick to the gross margin above 20%. Tesla only has one year to drop to eighteen, which should be 2018, the year Model 3 is difficult to produce. Although Tesla has cut prices on a large scale this year, he said at the annual report meeting that I still maintain a gross profit margin of more than 20 points this year, because only in this way can the company develop healthily, and when the company makes any long-term investment, it can not be intimidated, and can it firmly invest down.

At present, ideal R&D includes product R&D, platform R&D, and system R&D. How does the cost of the platform continue to fall, and is it to use someone else's system or develop your own operating system? There are also chips, if you make your own inference chips, you can do the same cost as Tesla, because the algorithm, training platform, and chip are in your own hands.

In February, all of our supporting store staff had achieved a sales capacity of close to 7 vehicles per capita. If we remove the management, then there are more than 10 cars per capita, which is about three times more than the peers of the new forces, so we have the opportunity to let our employees do the work of three people and earn two people's money, so our people are now the hardest to dig. Including the charging network just mentioned, we look at the necessity of strategy in the entire structure, not where to go, but what must be built, and to what extent.

Supply chain: 24 months forward

The most critical thing to limit our climbing speed is the three electrics, we divided all the parts of the car into four parts.

The first is the traditional rearview mirror, bumper, the automotive industry originally has a production capacity of more than 20 million passenger cars a year, there is no difference between electric vehicles and fuel vehicles, which do not require us to do it, and it can be given to suppliers;

The second category is new electronics, chips, such as our domain controllers, China used to have a very good foundry system, and China purchases about 40% of the world's chips, the manufacturing efficiency of this system is very high, give us a car to produce 200,000 sets, 300,000 sets of domain controllers without any problems, companies like Foxconn can do it for you, and the solution is not complicated;

Everyone is talking about the lack of chips, but there has never been a lack of Qualcomm chips and Qualcomm computing platforms, no lack of NVIDIA chips and NVIDIA computing platforms, nor lack of Horizon chips and Horizon computing platforms, lack of traditional small function chips, because China is still very mature in this regard, China is the world's strongest electronic foundry.

The third type of problem is more troublesome, which is related to the three electricity. For example, when we develop the L series, a car uses three motors, two drive motors, and a generator. If they sell 300,000 vehicles a year and place an order for 900,000 units from any motor factory, they can't take it, because 900,000 units means that he has to build 10 factories, and if we don't want it, then they will bear all these things.

There is a very interesting phenomenon, when these suppliers face new forces, the order quantity we report to them is different from the order quantity they calculate, and they will have an analysis system to analyze how much we can actually sell, so that they are willing to produce so much. When car companies used to sell only 10,000 units, but told suppliers that they needed to produce 30,000 units later, they wouldn't believe it. At the end of last year, all the suppliers thought we could do about 150,000 vehicles this year, and it wasn't until we proved that volume that we were transferred to 250,000. And behind this situation, like the three electricity, including the range extender, we all produce it ourselves, otherwise our suppliers will not be able to keep up at all. So we built our own range extender factory in Mianyang, and the first five-in-one motor of our L series is produced by ourselves next to Changzhou.

We will produce our own silicon carbide modules for 800 volts, including our own silicon carbide motors. It's all our supply capacity.

In addition to our factory in Changzhou, we will also establish a factory in Beijing to produce pure electric vehicles.

Supply capacity needs to be 24 months forward, otherwise it will be too late. Including taking land, building factories, recruiting people, and trial production, it basically takes 24 months to come. It's the same with suppliers, if you find a European manufacturer and only make us 30,000 to 40,000 sets of air suspension a year, we talk to him for a year, and we may be able to supply up to 60,000 sets. But we need hundreds of thousands of sets a year, and then we will discuss with Chinese suppliers how to build production lines and factories to ensure that second- and third-level components are in place. The supply chain needs to be done very well, and we have not done well in the past year, but some aspects of our advance layout have brought us a lot of help.

The homemade rate does not exceed 30%, otherwise it carries high leverage

Another problem with the supply chain is that the self-control rate cannot be too high. All companies in the world have experienced it, such as Toyota to spin Aisin and Toyota Boshoku out and share with others, and GM to take Delphi out of Delphi.

We know that many parts manufacturers are now dismantled. Because if you do all your own homemade, it means that when sales fall by 40%, you have to bear all the leverage and all the costs in 40%. At this time, I am particularly afraid of encountering such a big problem as an economic crisis

Therefore, we internally believe that 30% self-made rate is relatively healthy, and we can cooperate with suppliers, even if we develop ourselves, we can also hand over to suppliers to produce, rather than produce ourselves.

AI changes the physical world

At Ideal Auto, we have four artificial intelligence algorithm teams that are solving problems in four areas, including:1. The intelligent cockpit team established last year does multimodal algorithms; 2. Intelligent driving; 3. Smart Factory; 4. Retail intelligence.

Why should "smart factories" do artificial intelligence, the reason is that we have hundreds of inspectors in a production process every day on the production line, which can be replaced by artificial intelligence algorithms; At the same time, it can also help us use algorithms to solve quality problems, from a quantitative point of view, each car in the allocation of quality costs, about 50% of the same price Mercedes-Benz, BMW, Audi.

Source: Ideal Car

Therefore, if you measure the quality status in terms of money, the same price of the product, our quality is significant BBA these luxury car brands.

In addition, this algorithm can also help us decompose various accidents, if you ask friends around you who drive ideal ONE, you will get an important feedback - our insurance is the lowest price in all peers, even lower than fuel vehicles, which is different from ordinary electric vehicles, because the insurance price of general electric vehicles is more expensive than fuel vehicles.

As a car of more than 300,000 yuan, the part of commercial insurance in the second year of renewal is 3,300 yuan, and if you add strong insurance, the whole is 4,300. The standard price given by all insurance companies is so low, and this price is even similar to the insurance of many new energy vehicles of about 100,000 yuan, which is thanks to our entire algorithm system.

Going back to why we made such a system, it is actually related to my previous Autohome. When I was doing Autohome before, basically all products were successful, but the only one that failed, and the mess that failed, was the car e-commerce.

Moreover, not only did we fail in automotive e-commerce, including Bitcar, Taobao and JD.com, as well as many automotive e-commerce startups and second-hand car e-commerce companies at that time. This is not only a domestic brand, but also Carvana's second-hand e-commerce that has achieved a market value of 560 billion US dollars in the United States in the past few years, and today it has fallen to more than one billion or two billion US dollars, and we have not succeeded in learning True CAR in the United States.

When we reflect on why such a problem occurs, there is a very important consideration, 90% of the entire business chain of automobile e-commerce, or even 95% of it occurs offline, in the entire offline warehousing, inventory, logistics, delivery experience link, we did not do any transformation at that time, the whole cost is even higher than the traditional car dealership, we just spent 5,000 to 15,000 yuan more CPS costs online, and the rest of the transformation is nothing. That is to say, in the entire chain of measuring efficiency and business with money, we have nothing left to create value except for spending a part of the extra money.

The same problems are the same in new retail, which everyone thought was very popular in the past, such as community group buying and taxi software. Whether it is Uber in the United States or Didi in China, today's market value is far less than the market value of its financing at that time, and its core problem is that it was software 1.0 in the past, and there is no way to solve the problem of the offline physical world.

What is Software 1.0 as we define it? Whether it is our use of the Internet, the APP we use, or other terminals, it is the same principle, the main core is to run in a digital world, but everything that happens in the physical world we can't change.

We can take Uber very quickly in the United States to get a car, but everything that happens between this car from point A to point B, the cost of what we call TCO has not decreased, even higher than leasing, because the air rate is higher, so these platform companies will have a problem, when it wants to get a high-speed growth and share it will lose, if it wants to make money, the only way is to exploit drivers, such as the United States collects 30%, China receives 25%, and the previous rental company only charges 20%, So they now charge drivers more than rental companies.

Retail e-commerce is different, e-commerce when this expensive shopping mall to sell goods to a warehouse can sell goods, the overall cost drops by 35% to 40%, but 35% to 40% of the entire industry is divided by three, the first is that consumers take a part, so that the retail price is cheaper; The second is that the platform side took away part of it, creating trillion-dollar companies like Ali and Amazon, and the third is these brands - brands that use new retail to do e-commerce.

So, we were thinking about the question – what can change the physical world? Is there any way to change the physical world? We found that there is only AI.

AI can change the physical world, because the core of AI is actually learning, not programming, not logic, but relying on its imitation and learning of humans.

At present, Ideal has four artificial intelligence algorithm teams, which are challenged in four fields: intelligent cockpit, intelligent driving, intelligent factory, and retail intelligence.

The ideal divides AI into two categories, one is behavioral learning, that is, the cerebellum is the main and the brain is supplemented. Use the brain to think and train yourself, but after successful training, the cerebellum can process tasks, such as driving or production line detection, etc., and many times do not need to think completely by the brain, and respond quickly and in time.

Another category is cognitive learning, with the brain as the main and the cerebellum as the supplement. ChatGPT has become popular recently, including "Ideal Classmate" and NIO's "NOMI". When a complex command is given, it has to go back to the cloud to process, and then after training, it has a relatively deterministic execution, and if it is placed on the car side, the result must be given through the GPU or BPU.

Autonomous driving has also gone through two stages: behavioral learning and cognitive learning. In the first stage, the sensor and computing power are not strong. For example, the earliest chip that uses Mobileye 2 TOPS computing power is ideal, and the camera with a forward view of 2 million pixels. These hardware is not enough to support three-dimensional perception, identify two-dimensional pictures one by one, and even "two-dimensional ranging", the wider a picture, the closer the distance, the narrower the picture, the farther away. To solve the problem of ranging, add a millimeter-wave radar.

That's how 99% of the cars on Chinese roads worked, even if they were on the NVIDIA Orin computing platform. Tesla's AP and FSD in China both work in this way. This is a 2.5D way of working, which is like a bat, using radar to measure the distance between me and various obstacles and target objects, which is the premise of the next step.

Therefore, the judgment and decision-making of these mass-produced and delivered cars have nothing to do with AI. Algorithms are all rule-based, algorithms written by humans with transparent rules. What do I do when I get stuck in the lane, what do I do when I leave the lane, what do I do when I get closer, and so on. These things are implemented by a large number of people debugging, relying on human programming.

Whether the intelligent driving experience is good or not also depends on who writes more rules than AI algorithms. A definite conclusion is that it can be solved within the rules, but it cannot be solved outside the rules, and the accident is the responsibility of the consumer himself, because it is only an assisted driving.

In terms of control, it is also completely rule-based. Therefore, when encountering a traffic jam, if you follow closely, you will get motion sickness, because the brakes are sharper and the start is faster. A car that is more comfortable, acceleration will be more stable, but it is easy to be plugged, and major car companies are constantly adjusting, to find a balance between the two.

The fourth part is feedback growth. Due to the collection of more video data, behind this is the human calibration team, such as Tesla's calibration team in India, the calibration team of Chinese brands are also outsourced, and a large number of calibration teams are in Guizhou.

At this stage, the video needs to be manually calibrated, for example, a person can identify that there is a pile, and there are reels and two reels can be identified as a car. When the calibration is completed, the car will not hit it directly. Without calibration, the system may think that there is no object ahead, and the strategy will be to continue driving and cause a direct collision.

At this stage, whether it is artificial intelligence for vision or artificial intelligence for voice in the industry, we joke internally - as many people there are, there is as much intelligence.

The cognitive learning changes in the autonomous driving industry began with Tesla's FSD. Tesla's former senior director of self-driving algorithms, was recruited by Musk from Open AI, and since his pass, Tesla's intelligent driving has changed essentially.

The FSD chip is very efficient, although it is only 144 TOPS, but the effective computing power is basically the same as the two Orin, and the effective computing power is about 3 times the level of the general-purpose GPU. In addition, Tesla has upgraded to a 360-degree camera, turned into three-dimensional vision, and began to do BEV perception in the way of "big model". The good thing is that cars are starting to see the world in a human-like way.

Although there is still a certain difference with people. At present, the camera only has 2 million to 8 million pixels, while the human eye basically has 150 million to 200 million pixels. There is no problem for people to see 400-500 meters away, and today the camera sees about 100 meters is basically the limit, especially at night.

With three-dimensional perception, judgment decisions are different. Judgment and decision-making are actually divided into two layers: one is a white box, programmed to obey traffic rules and obey the law. The other layer is from three-dimensional perception, to judgment and decision-making, to execution and control to form a whole, we call it "end-to-end". It can clearly see how humans do, what things humans see, what judgments they make, and what kind of execution they make, and this process can be recorded.

This is an essential change that we call "shadow learning." It is a learning process and requires the ability of the terminal GPU or dedicated BPU to have a closed loop.

From here, the "feedback growth" level is also different. Structural data is transmitted back to the supercomputing platform for training, and training is a black box process. Some algorithms are white boxes, which belong to the 1.0 or 1.5 stage, completely black box algorithms, that is 2.0, which is an essential change.

Cloud training can do automatic labeling, classification, and training, and no longer need so many people. At present, the number of Tesla and Open AI algorithm teams is very small, which is very different from the number of AI companies in China.

And because cloud training is a black box, training is not observing right and wrong, but observing the degree of improvement of training quality. Only if there are enough samples of this "end-to-end" training, the overall driving experience and safety will become better and better, just like how humans work in various scenarios to learn.

To establish this set of autonomous driving cognitive learning system, car companies need to compete in three things in intelligent driving: First, how to reduce the cost of the computing platform and sensors on the vehicle end, and achieve the standard configuration of each vehicle as much as possible. The cost of intelligent driving is including sensors (radar, camera) + computing platform, Tesla is 1500 US dollars, the ideal is 4000 US dollars;

The second is more "end-to-end" closed-loop data than who, only a part of the data is useless, complete end-to-end training can be, different countries, different scenarios of training are different, can not be simply moved over and reused. This requires companies to sell enough cars full of sensors and computing platforms to complete this part;

Third, complete the training of large models with higher costs and greater challenges. The United States began to restrict the development of Chinese training chips last year, as well as Nvidia's sales of high-bandwidth training chips to China. Today, whether it is using the GPU on the car side or using the GPU in the cloud, it is not the best way to perform for large models, it only has about 20% of the performance, and theoretically the other 80% of the cost is wasted.

Therefore, like NVIDIA as its own BPU and D1 training chip, after the whole system is built, it can probably replicate the entire computing cluster at the cost of NVIDIA A100 1/6, and the scalability becomes higher. This is also the part that companies determined to do artificial intelligence must face, solve and invest in research and development.

In addition, in the future, the real competition will appear at the lowest level - the operating system of AI. Because the operating system of AI requires hard real-time, the specified time and place must be executed, and the operation in the physical world cannot be delayed.

Many people say that we are more careful in artificial intelligence, but in fact, we have been doing in-depth perception in this area for a long time, and after determining that this route is feasible, we will do it all-in. After Tesla ran through the approximate logic model, we began to officially invest.

Many companies are developing well or investing in AI early, but it is not reflected in sales. The key is to find an accurate user experience value point, which will fully explode. Just like smartphones meet three conditions at the same time: 3G network, App Store launch, OTA technology, mobile phones really enter the era of smartphones from the era of touch screens.

In 2024, enter the era of intelligent electric vehicles

Back to smart electric vehicles, when are smart electric vehicles really smart? The terminal is not smart, the terminal just continues the function of the mobile phone, why continue the function of the mobile phone? Because Android has built a very good ecology, consumers are not to use Android, but to use the ecology behind Android, because there is a map ecology, an entertainment video ecology, and a large number of application software developers' ecology.

But I think the era of change that truly belongs to smart electric vehicles should start from the real 2.0 can run. I make my own prediction, and I think this point in time, especially for mid-to-high-end cars, will appear in 2024.

True urban NOA based on large models, based on BEV technology. Today, the ideal car has been considered to be relatively good, and the utilization rate of the entire assisted driving is more than 13%. Because high-speed scenarios are limited, although our assisted driving usage rate in high-speed scenarios has exceeded 50%, there is no way to use it in urban areas. I think that a technology that can only be used by users every day, and the mileage usage rate is stable by more than 60%, this technology will become inseparable for users to no longer do without.

What are the results of the city NOA? In fact, it does not need to become completely autopilot, but it can help drivers improve the solution experience when commuting to and from work and stuck in traffic every day. One is that more than 60% or even more than 80% of the cars in the city are handed over to this function to drive, and the safety is no problem, unless others violate the rules, even if there is an accident, the probability is a small scratch. At this time, there will be a huge change, just as consumers buy a 20-30 story building, there is a fundamental difference between an elevator and no elevator.

I think one of the phenomena this year, using the Orin computing platform, these companies will basically deliver the earliest NOA for testing in the fourth quarter of this year, based on large models. Because it requires rich training, it is not just putting the computing platform on it.

I personally believe that by the end of this year, most of the leading companies will be able to achieve the level of Tesla at the end of 2011, and by 2024, everyone will generally be able to achieve the level of Tesla in North America at the end of 2022 and the beginning of 2023. I think that after this, at least mid-to-high-end cars, if they cannot provide urban NOA, will affect consumers' purchase decisions.

At this point in time, starting from mid-to-high-end cars, we will completely enter the era of intelligent automatic vehicles, the era of intelligent electric vehicles based on software 2.0. Otherwise, you can only sell non-smart electric cars forever.