Tesla will die "stärkste KI der Welt" in seine Autos integrieren. Hierunter leiden Huawei und Li Auto am stärksten unter Druck.

Elon Musk's most pride - and - joy, the Tesla intelligent cockpit, has been as unnoticed as a stray on the roadside in the Chinese market. Presumably, he must be feeling quite down.

Just like Mark in A Better Tomorrow who vowed, "I'll take back what I've lost with my own hands," another "Mark" across the ocean seems to be gearing up to regain Tesla's throne in automotive intelligence after half a year of biding his time.

Last Thursday, a day after the official release of the Grok large - language model, Musk excitedly announced on X that Grok AI will be launched in Tesla cars "at the latest" this week (i.e., the current week).

If realized, it means Tesla will take an extremely important step in intelligence. This is not only the first time Musk has given an accurate schedule for "Grok in cars," but the Grok to be installed is also what Musk calls the "world's most powerful AI":

For example, Musk has long claimed that Grok 4 will "rewrite the human knowledge base." At the press conference, he once again emphasized that Grok 4 is currently the smartest AI in the world.

With such a powerful AI coming to cars, can the leaders in the domestic intelligent cockpit field, such as Huawei and Li Auto, hold their ground?

Empty boasts or real capabilities?

When it comes to Grok, what most people might first think of is its "reckless talk." Whether it's saying "the biggest firework in Japanese history was the atomic bomb explosion" or supporting Hitler's "anti - Semitic remarks," they are all quite memorable.

After the release of Grok 4, people's attention shifted to its "sky - high" annual subscription fee of up to $3000. So when Musk said Grok 4 is the "world's smartest AI," most people were skeptical.

Sure enough, Grok 4 "flipped over" during actual testing.

Blogger @karminski - dentist posted on Weibo that in his test of the classic question "20 balls bouncing inside a heptagon," Grok 4 had syntax errors in 2 out of 3 code generations. Even the only successful attempt showed a significant gap compared to the initial version of DeepSeek R1.

In another more difficult test, the "chimney demolition simulation," Grok 4's performance was also mediocre. The scattered particles were rough and blurry, and most ridiculously, the simulated chimney base wasn't even cylindrical... A comparison with the adjacent DeepSeek R1 made the difference obvious at a glance.

Anyway, Grok 4 is far from the top - tier level in code - writing, let alone living up to the over - hyped claim of being the "world's smartest AI."

However, although Musk hasn't confirmed what role Grok will play in Tesla's intelligent cockpit, foreign media analysis suggests it will most likely serve as a "voice assistant" in the car system first.

If it's just for the "voice assistant" role, it's a piece of cake for Grok.

Especially in terms of the key capabilities required for a voice assistant: multi - disciplinary reasoning, semantic understanding, context understanding, and Agent capabilities, Grok 4 is indeed one of the best among current large - language models.

In the multi - disciplinary reasoning aspect, in the HLE (Humanity’s Last Exam), a test often translated into Chinese as "Humanity's Ultimate Exam" or "Humanity's Last Exam," without tool assistance, Grok 4 scored 25.4%, significantly surpassing Gemini 2.5 Pro's 21.6% and OpenAI o3's 21%. With tool assistance, Grok 4's score can further rise to 38.6%, and Grok 4 Heavy can even reach 44%.

Moreover, in the ARC - AGI test, Grok 4 even "set a new record." This test aims to evaluate the general intelligence level of AI models, mainly through visual reasoning questions. Grok 4 scored 16.2%, leading by a wide margin on the score coordinate system.

In terms of context understanding, Grok 4 supports a context window of 256k tokens. Although it can't match Gemini 2.5 Pro's astonishing 1 million tokens, it still surpasses Claude 4 Sonnet, Claude 4 Opus, ChatGPT o3's 200k tokens, and DeepSeek R1 0528's 128k tokens, which is more than sufficient for a voice - assistant role.

Moreover, Musk also said that the seventh version of the Grok 4 base model will be completed this month, which will have better video - understanding and tool - invocation capabilities. In the next few months, xAI will also launch a code model, multi - model agents, and a video - generation model.

In short, even the single - agent version of Grok 4 is definitely sufficient for adjusting the air - conditioning, windshield wipers, rear - view mirrors, etc. in a car.

With the support of Grok, Tesla's intelligent cockpit can be compared to upgrading from a simple rifle to a powerful Barrett. Imagine that on top of Tesla's already impressive sales, having a voice assistant that can communicate smoothly in natural language and control a wide range of in - car functions directly addresses a long - standing pain point for consumers.

As long as Grok on Tesla can reach the average level of voice assistants in domestic new - energy vehicles, there will be one less point for Xiaomi, Li Auto, Huawei, XPeng, etc. to "mock" Tesla at their press conferences, putting more pressure on their publicity departments.

But is the "voice assistant" the full significance of Grok in cars? Obviously not.

Infusing soul into FSD

Besides filling the gap in human - machine interaction, let's think more boldly: Is it possible that such a powerful Grok 4 can even influence or control FSD?

This idea is not unfounded. In China, new - energy vehicle manufacturers have long included "natural - language - controlled assisted driving" as a must - have feature for the next - generation assisted - driving systems in their intelligent - assisted - driving function plans.

At the NVIDIA GTC2025, a Li Auto executive mentioned that in their next - generation autonomous - driving technology, MindVLA, they aim to turn the car's intelligent - driving system into a professional driver that can understand, see, and locate.

According to Li Auto's description, users can interact deeply with the car system. For example, during intelligent driving, if the car is going too fast, users can use the vague voice command "You're going too fast" to limit the speed. After the car has planned an intelligent - driving route, the owner can also change their mind and say to the car system, "Turn right at the next intersection," and the intelligent - driving system will modify the route accordingly.

This "science - fiction - like" function is also being planned by other car manufacturers such as Huawei, XPeng, Zeekr, and Xiaomi.

Interestingly, the capabilities demonstrated by Grok 4 seem to be tailor - made for "natural - language - controlled assisted driving."

In "natural - language - controlled assisted driving," the core lies in the AI's understanding of natural language and its accurate mapping of driving intentions. At this time, Grok 4's support for a 256K context window and multi - agent collaborative architecture (exclusive to Grok 4 Heavy) has the potential to handle the complex correlations among user voice commands, vehicle status, and environmental data simultaneously.

When the user says, "There's going to be a traffic jam ahead. Find a less - congested route," Grok can parse three levels of semantics: environmental perception (identifying the congestion status on the map), user preference ("less - congested" = low - traffic route), and action generation (triggering route re - planning), and then quickly select a new driving route.

In addition, Grok 4 Heavy's 4 - Agent parallel architecture technically provides a "human - like decision - making brain area" for FSD, which is somewhat similar to the "big - and - small - brain" architecture emphasized by domestic car manufacturers in their intelligent - driving systems.

For example, the four Agents of Grok 4 Heavy can be a perception agent (fusing data to build an occupancy network), a planning agent (route planning based on spatial - simulation capabilities), an interaction agent (mainly handling in - car voice interaction and commands), and a safety agent (monitoring real - time conflicts between Action Tokens and traffic rules).

When the FSD function is activated, these four agents can collaborate. In a simple scenario where the user gives a voice command to "overtake on the left," the interaction agent perceives the user's command, the perception agent identifies the lane conditions on the left, the safety agent determines if there is a safe overtaking window, and finally, the planning agent plans the specific overtaking route, and FSD executes the overtaking.

If this function can be implemented in cars, such voice commands can be extended to many car - using scenarios, including Li Auto's proposed "finding a parking space" and "flexible route planning."

Moreover, considering Musk's actions in autonomous - driving driverless cars, the integration of Grok in cars may be a crucial step for Tesla to build a new in - car AI ecological closed - loop and deploy autonomous driving.

Ultimately, the competition in intelligent driving boils down to the competition in autonomous driving. Starting from June 22, Tesla launched its first batch of Robotaxi services in Austin, Texas. Compared with FSD, Robotaxi may need a smarter "brain," or rather, a "soul."

With "semantic - level control of FSD," passengers can control the driving rhythm, style, and route of the Robotaxi under legal and safe conditions, and can also ask the car, just like they would ask a taxi driver, "Why did you brake suddenly just now?"

At this time, Grok can query the car's decision - making log and translate the reason for the sudden braking into understandable language, such as "The car in front suddenly decelerated."

However, compared with Li Auto's announcement that MindVLA will be installed in cars in the second half of 2025 and Huawei's similar - effect Pangu large - language model 5.0 empowering ADS starting to be rolled out gradually from August 2025, it's still uncertain whether Tesla can truly achieve "semantic - level control of FSD."

Meanwhile, "natural - language - controlled assisted driving" also faces a series of challenges such as regulation, liability, ethics, and computing - power deployment. The most pressing question for Chinese consumers is, Can Grok enter the Chinese market?

This issue affects not only the distant "natural - language - controlled assisted driving" but also the "Grok voice assistant" that is supposed to be launched this week. Most likely, North American users will get to experience it first, leaving Chinese users longing.

Although there are many uncertainties about the "world's most powerful AI in cars," it doesn't mean domestic new - energy manufacturers can relax. As long as there's a possibility, domestic car manufacturers will face great pressure.

Domestic cars under huge pressure

Based on the current technology and sales, Tesla has the confidence to be calm. Moreover, with the potential addition of Grok in the future, Tesla's intelligence level may reach a new height.

In the just - passed June, although the Xiaomi YU7, touted as the "ultimate Model Y killer," was launched and indeed achieved an unprecedented record of nearly 300,000 orders in one hour, Tesla Model Y still sold a staggering 44,848 units (according to Dongchedi data), and in some institutional statistics, it even reached 51,253 units (according to Gasgoo). It remains unfazed by the competition.

Even the long - range version of Model 3 dared to increase the price by 10,000 yuan.

Tesla's technology is also equally robust. Since the launch of FSD in the Chinese market in March, it has been mocked for its high subscription price and inability to recognize Chinese traffic lights and bus lanes, becoming an easy target for other car manufacturers' intelligent - driving systems.

But when domestic car manufacturers are touting "recognizing traffic - police gestures" as a major selling point, many netizens noticed months ago that FSD had quietly "learned" to recognize traffic - police gestures and drive accordingly.

Netizen @Huang Site, who dared to use FSD, posted