Pogbagram

Reports indicate that Apple's standalone Siri app will support automatic conversation deletion and significantly replace its underlying model with Gemini.

Last night, Bloomberg reporter Mark Gurman revealed more details about Apple's upcoming standalone Siri app in the latest Power On newsletter.

Privacy is one of the key selling points of this new app. Gurman revealed that the app will have a built-in feature to automatically delete conversations, and users can choose how long conversations are retained in the settings: 30 days, 1 year, or permanently. Apple is expected to continue emphasizing privacy protection at this year's WWDC and highlight that its AI experience is ad-free.

In terms of models, Apple has largely replaced the underlying AI model of the new Siri with Google's Gemini, and runs some services on Google's cloud infrastructure. Previously, there were reports that Apple's relationship with its former partner, OpenAI, had soured, potentially leading to a legal battle, because OpenAI believed the collaboration had not yielded the expected results.

It's worth noting that although the new AI Siri has been delayed for about two years, Gurman said that Apple's internal test version is still labeled "Beta" and may continue to be used until its official release.

Malta will provide all citizens with one year of free ChatGPT Plus.

OpenAI recently announced that it has signed a cooperation agreement with the Maltese government to provide Maltese citizens and residents with a free one-year subscription to ChatGPT Plus.

It is understood that residents must first complete the free online course "AI for All" developed by the University of Malta. The course covers the basic concepts of AI, its capabilities and boundaries, and responsible use. After passing the course, they can apply to the Malta Digital Innovation Authority for a free subscription.

Xiaomi REDMI Product Manager: If you don't plan to change your phone, you can wait a bit longer; more models will support battery upgrade services.

Xiaomi REDMI product manager "Suncun" recently posted on Weibo suggesting that users who don't plan to change their phones in the near future should consider waiting to replace their batteries. He also pointed out that the capacity of some models may be increased after the new battery is replaced, referring to the current policy of the Xiaomi 13 series.

Sun Cun further explained that Xiaomi is actively encouraging existing users to upgrade to high-capacity batteries using GSR Ventures technology, with the Xiaomi 13 series already having this service available. She also revealed that specific supported models will "have to wait."

Previously, the Xiaomi 13 series had launched a battery upgrade service, supporting the upgrade of the Xiaomi 13, Xiaomi 13 Pro and Xiaomi 13 Ultra models to larger capacity batteries. The cost is 189 yuan, including 149 yuan for the battery upgrade and 40 yuan for labor.

With a target valuation of $1.75 trillion, SpaceX could go public as early as June.

According to Reuters, SpaceX has accelerated its IPO timeline, planning to list on Nasdaq as early as June 12 under the ticker symbol "SPCX," about two weeks earlier than the previously scheduled late June.

According to three sources familiar with the matter, the company plans to release its prospectus as early as this Wednesday, with the roadshow scheduled to begin on June 4th and pricing expected to be finalized as early as June 11th. The main reason for this accelerated timeline is that the SEC's review of the prospectus is progressing faster than anticipated.

This IPO is expected to raise approximately $75 billion, targeting a valuation of about $1.75 trillion, potentially making it the largest stock offering in history. The underwriting lineup includes Morgan Stanley, Bank of America, Citigroup, JPMorgan Chase, and Goldman Sachs.

arXiv introduces new rules to ban AI spam, with violators facing a one-year ban and enhanced peer review as a double penalty.

Recently, the preprint academic paper platform arXiv announced that if an author's submitted paper contains irrefutable evidence of AI-generated content, the author's account will be banned for one year, and thereafter all submissions will be required to be accepted by a peer-reviewed journal before they can be uploaded.

Thomas Dietterich, chair of the Computer Science section on arXiv, stated in an article on X that if a paper contains obvious signs such as "illusionary citations" or large language model meta-annotations (e.g., "The following is a 200-word abstract, should it be modified?"), it can be concluded that the authors did not verify the AI-generated content.

arXiv, currently managed by Cornell Tech, will become an independent nonprofit organization this July. Cornell Tech Dean Greg Morrisett stated that independence will allow arXiv to access funding from a wider range of donors to address the operational challenges posed by "AI-generated junk content."

Lu Weibing: The Xuanjie chip will definitely be upgraded this year; the price of some domestic flagship candybar phones may exceed 10,000 yuan in the second half of the year.

Recently, Lu Weibing, partner and president of Xiaomi Group, said during the "Max Summer" outdoor camping live broadcast that the Xuanjie chip will definitely launch an iterative version this year, and that it will be "a very powerful chip" that will be installed in a "very excellent product".

He also cautioned that the various rumors circulating online are not highly credible, and specific specifications cannot be disclosed at this time. Lu Weibing further revealed that the application of the new generation of Xuanjie chips will no longer be limited to smartphones, and there are plans to widely integrate them into various smart terminal devices under Xiaomi.

Regarding mobile phone pricing, Lu Weibing predicted in a live broadcast that the prices of some domestic flagship candybar phones may exceed 10,000 yuan in the second half of this year, especially towards the end of the year. He also predicted that the pressure of rising memory prices will continue at least until the end of 2027, and possibly even into 2028.

Regarding the upcoming Xiaomi 17 Max, he stated that the pricing is still under discussion and will be officially announced at the launch event. However, Xiaomi's principle is to provide "the most cost-effective product" in its price range.

Li Bin of NIO: Firefly, like the iPhone, has transcended the function of a car as a tool and become a fashion brand.

According to Economy, at the 18th Xuanyuan Auto Blue Book Forum held recently, NIO founder, chairman and CEO Li Bin attended online and compared the firefly to the iPhone in the dialogue, saying that it has become "a very high-quality product".

He emphasized that Firefly has, in a sense, transcended the utilitarian nature of automobiles, evolving into "a lifestyle brand, a fashion brand," and has begun to form a unique brand culture.

He stated that Firefly breaks down the dual limitations of age and car purchase budget in terms of user groups. "Young people who have just graduated from university can buy it if they want to, and wealthy bosses can drive their children to school themselves without any problem."

Demand surged by over 700% in a year, making forward deployment engineers the most sought-after job in the AI ​​era.

According to Business Insider, demand for forward deployment engineers (FDEs) has surged by more than 700% in a year, making it one of the fastest-growing jobs in the tech industry amid the AI ​​wave.

Data from the recruitment platform Indeed shows that the number of job postings for this position reached 5,330 in April this year, an increase of approximately 729% compared to 643 in the same period last year. Anthropic, OpenAI, Palantir, Stripe, and Google Cloud are all actively recruiting for this position.

The core responsibility of an FDE (Financial Design Engineer) is to assist companies in integrating AI into their business processes and to develop customized AI tools based on client needs. In terms of compensation, Indeed data shows that the annual salary for this position ranges from approximately $170,000 to over $200,000 .

Starting at just 1 yuan for 400,000 tokens, the three major telecom operators launch token packages.

China Telecom launched a series of trial commercial Token packages yesterday, providing integrated "Token + Connectivity + Security" services for developers and SMEs, as well as individuals and families. Each package offers three tiers: 10 million Tokens starting at 9.9 yuan/month for individuals and families; and 15 million Tokens for developers and SMEs at 39.9 yuan/month.

Both types of users can opt to purchase a broadband uplink speed boost package (uplink speed increased by 50 Mbps, 100 yuan/month) and a smart security butler service (5 yuan/month). Individual and family users can also additionally purchase the 5G-A premium package (including 10 GB of data, up to 300 Mbps uplink speed, 20 yuan/month).

China Unicom and China Mobile, the two major telecom operators, have also launched their own token packages.

• China Mobile Shanghai launched its Token Universal Service at a press conference on World Telecommunication Day, May 17, priced at 1 yuan for 400,000 Tokens.
• China Unicom: On May 16, China Unicom Shanghai Branch released the fully domestically produced "domestic chip, domestic model, and domestic cloud" security platform and launched multiple token products and integrated packages.

Tesla Robotaxi crash report released

Recently, Tesla officially declassified and released the full, unanonymized reports of 17 "Robotaxi" autopilot collisions submitted to the National Highway Traffic Safety Administration (NHTSA).

According to Electrek, the 17 incidents occurred between July 2025 and March 2026, all involving 2026 Model Y vehicles undergoing "Robotaxi" testing in Austin, Texas, and all involved safety drivers present during the tests.

The casualty data shows that the overall accidents were relatively minor: 13 incidents resulted in only property damage, 2 incidents resulted in no injuries, 1 incident resulted in minor injuries that did not require hospitalization, and another incident resulted in minor injuries that required hospitalization (the vehicle was rear-ended by an SUV while crawling at 2 mph at an intersection).

Official reports indicate that most accidents were not due to fault of Tesla's Autopilot (ADS) system, but rather occurred when the vehicle was completely stationary at traffic lights, stop signs, or in traffic jams, and was rear-ended or sideswiped by a human driver.

However, the data also revealed flaws in the system and backend intervention:

Two incidents occurred during remote operator takeover, resulting in vehicles hitting metal fences and construction barriers at speeds of 8 mph and 9 mph respectively. Several other incidents revealed blind spots in ADS spatial awareness, with vehicles hitting metal chains, trailer hitches, wooden utility poles, and curbs while turning or reversing.

Zhihu CEO Zhou Yuan: In the AI ​​era, every real person possesses irreplaceable value.

According to Tingtong Tech, Zhou Yuan, founder and CEO of Zhihu, said at the 12th New Knowledge Youth Conference that the value of real creators in the AI ​​era is irreplaceable.

AI creates wonders every minute, but the AI ​​era still needs new knowledge. And new knowledge comes from real people. Every real person possesses irreplaceable value.

At the main forum of the conference, several guests discussed the relationship between AI and human creation. Liang Yong'an, a humanities scholar from Fudan University, proposed, "Technology has its uses, but how can we refrain from using it? This requires human judgment and humanistic qualities." Writer Zhang Chun stated, "Imperfection is our perfection."

At the end of the conference, Zhou Yuan concluded: "The more advanced the technology, the more valuable authentic works of art that are warm and profound."

Xiaomi YU7 GT Volcanic Gray Color Option Announced

Recently, Xiaomi Auto released another teaser image of the YU7 GT, revealing new color options and interior information.

Among them, the Xiaomi YU7 series will add a new color option called "Volcanic Ash". The official description is as follows:

Inspired by the volcanic landscape under the morning mist, the high-purity gray tone, layered with nano-sized mineral particles, presents a delicate and layered metallic texture.

Meanwhile, the interior of the new car also features a brand-new "Racing Red" color scheme: a striking red and black color combination with exclusive embroidery.

The official statement indicates that the Xiaomi YU7 will also offer a "volcanic ash" exterior color option, and the cherry red YU7 GT is already on display in some Xiaomi car stores.

ChatGPT can help you manage your money now, supporting connections to over 12,000 financial institutions.

Recently, OpenAI launched a personal finance management feature for ChatGPT Pro users in the United States, supporting connections to bank accounts and credit cards from over 12,000 financial institutions via Plaid.

After connecting the account, ChatGPT automatically reads the user's actual balance, spending history, and debt data to generate a financial dashboard covering investments, daily expenses, subscriptions, and recent bills. When the user asks questions, the model can directly access the billing history, identify which categories are prone to overspending, and provide specific cost-cutting solutions.

In addition, users can tell ChatGPT some off-balance information, such as "I'm saving money to buy a car" or "I still owe my parents money," and the system will save this information as financial memory for later conversations.

Musk: The 1.5T "game-changer" model Grok V9 will be released within three to four weeks.

Musk posted on X yesterday that SpaceXAI's 1.5T parameter Grok base model version 9 (V9) has completed training and is expected to be officially released within 3 to 4 weeks.

The current public version of Grok 4.3 is based on the V8 base model with 0.5T parameters and is currently iterated every few days. V9 will increase the number of parameters to 1.5T, which Musk calls a "major upgrade".

At the end of the post, Musk even used the word "banger" to describe his expectations for this version.

ByteDance Open Source Cola DLM: No More Left-to-Right Generation, Diffusion Model Can Also Write Text

ByteDance's Seed team recently open-sourced its diffusion text generation model, Cola DLM. Unlike mainstream large models that write text right-handed word by word and token by token, this model first "thinks out the general meaning" in a continuous semantic space and then restores it to specific text.

Technically, Cola DLM consists of a Text VAE and a block-causal DiT. The Text VAE compresses the text into a continuous latent space, and the DiT learns the semantic distribution within it through flow matching. Finally, the decoder reconstructs the text. Diffusion denoising occurs at the semantic layer, not the token layer.

The total number of parameters in this open-source version is approximately 2.3B. In eight benchmark tests, including LAMBADA, MMLU, and HellaSwag, the paper states that its average score under the unified generative protocol exceeds the baselines of AR and LLaDA of the same scale.

GitHub: github.com/ByteDance-Seed/Cola-DLM Hugging Face: huggingface.co/ByteDance-Seed/Cola-DLM

Tencent Vibe Coding's product "Toast" is coming soon.

According to Tech Planet, Vibe Coding's product "Toast," incubated within Tencent, is about to launch. The Android version has already been released on App Store and the official website, with the iOS version to follow later.

"Toast" is positioned as an "application generation and inspiration co-creation platform". Users do not need coding skills. They can describe their needs through natural language, and AI can complete the entire process from function decomposition and prototype generation to APK packaging.

In addition to application creation, version 1.0 also supports social sharing, an inspiration square (one-click replication or secondary creation of others' works), and AI application recommendations. It is free for a limited time during the public beta period, providing 5 creation slots.

Mixue Ice Cream launched the same Peach Four Seasons Spring Ice Cream as Lao Huang, and sales skyrocketed.

Recently, Mixue Ice Cream's official ordering mini-program launched a special category called "Celebrity Picks," featuring the "Peach Four Seasons Spring" drink at the top. Previously, Nvidia CEO Jensen Huang visited a Mixue Ice Cream store in Nanluoguxiang, Beijing, and ordered a "Peach Four Seasons Spring," which garnered widespread attention.

According to Dahe Daily, a staff member from Mixue Ice Cream responded that after the launch of the "Celebrity Picks" mini-program, the sales of their Peach Four Seasons Spring ice cream increased by nearly 140% compared to the previous day and by more than 90% compared to last Friday. At the Nanluoguxiang store visited by Huang Renxun, sales of Peach Four Seasons Spring ice cream increased by more than 90% compared to the previous day.

Bank of China has shut down its standalone credit card app and integrated its services into a main app.

According to CCTV News, Bank of China announced yesterday that its dedicated credit card mobile application, "Colorful Life," will officially cease service on July 1st. Previously, several other banks had already announced the closure of their respective standalone credit card apps.

Bank of China responded that the shutdown will not affect customers' normal user experience, as all existing services have been integrated into the Bank of China main app. Other banks that have also announced the shutdown of their credit card apps stated that this is part of mobile application integration and will not lower credit card service standards.

Industry experts generally believe that reducing the size of standalone apps helps banks lower operating costs; while integrating credit card services into a comprehensive app also allows customers to switch smoothly between various financial services, improving the overall user experience.

Filming for the live-action "How to Train Your Dragon 2" has wrapped.

The live-action "How to Train Your Dragon 2" has officially wrapped filming and is scheduled for theatrical release in North America on June 11, 2027.

Based on the animated version's plot, the sequel will continue the world view of humans and dragons coexisting. Hiccup may experience a major turning point in his life: he finds his long-lost mother, but Toothless is controlled by the villain, and his father dies as a result.

Sonic the Hedgehog 4 has officially wrapped filming.

Jeff Fowler, director of "Sonic the Hedgehog 4," announced yesterday that filming has officially wrapped, along with a photo of himself with the new character, Metal Sonic. The film is scheduled for release in North America on March 19th next year.

Metal Sonic's presence was foreshadowed as early as the end of the third film, and this time he makes his official debut as a main character, marking the first time this classic mechanical villain has been brought to the big screen in the series. Furthermore, Amy Rose will also appear in this film, voiced by Kristen Bell.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

Finally, in May 2026, we witnessed the completion of a full wave of flagship imaging products.

Xiaomi created the Leica Instant, vivo developed the V-Single, which combines photos and videos, and OPPO surprised us with its dual telephoto solution.

At the very end of this wave of flagship upgrades, Huawei has prematurely launched the new Pura 90 Pro Max.

To be honest, at first glance, I was disappointed with the hardware. The impressive "dual-lens" design on last year's Pura 80 Ultra is gone, replaced by the most reliable and common large-sensor telephoto solution in the industry.

However, after a thorough experience, I think the Pura 90 Pro Max has a new solution that is more interesting than simply focusing on hardware – XMAGE intelligent shooting.

Fully automatic, worry-free.

Over the past few years, I have systematically reviewed the imaging architecture of Huawei, Xiaomi, OPPO, and Vivo. Although each company has its own understanding of hardware selection, the general direction is surprisingly consistent: they are all working hard to maximize sensor area within extremely limited body space, competing to see who can get the largest aperture and the most light.

It's understandable to strive for excellence in hardware. But in photography, a solid hardware foundation is only the first step.

To take a good photo, composition and post-processing are equally important.

Composition ensures that a photo is visually balanced and that the subject is clearly defined; while post-processing determines the visual quality and stylistics of the photo.

However, both of these have always been skills with high barriers to entry.

Huawei intends to address both of these issues in one go with the Pura 90 Pro Max.

Those familiar with their product line may recall that Huawei experimented with AI-assisted composition on the Pura 80 Ultra released last year.

AI-assisted composition is relatively basic, mainly serving as a guide and reminder in actual shooting. With the Pura 90 Pro Max, this technology has undergone a transformation from "assistance" to "dominance" and is named "XMAGE Smart Shooting".

Peeling back the surface to explore the underlying logic of XMAGE intelligent shooting, it is a localized image pipeline driven entirely by edge computing power, which breaks down the shooting process into three modules that run simultaneously: automatic composition based on real-time semantic segmentation, intelligent zoom that ignores the limitations of physical lenses, and color reshaping for the image.

In other words, with this feature, the Pura 90 Pro Max can automatically recognize the subject, zoom in and compose the shot, and finally provide a series of XMAGE styles suitable for the current scene for you to choose from.

It's worth mentioning that XMAGE Smart Shooting doesn't require a network connection or to transmit massive amounts of image data back to the cloud. It operates entirely on edge AI, recognizing the outlines of the main subject, architectural lines, and light source distribution in the image in real time. Then, based on the underlying trained aesthetic model, it delivers a combination of optimal composition, zoom, and refined color scheme.

The complex professional parameters are made black-box here, and the composition and color adjustment work that used to require human judgment is now entirely handed over to the local algorithm.

In addition, the Pura 90 Pro Max also launched AI pose recommendation. This feature also uses semantic segmentation to understand objects, environment, poses and background in the picture, and then directly draws a proportional wireframe outline in the viewfinder.

Users with mobile phones can simply follow the image to guide the model to the pose shown. Unsatisfactory poses can be refreshed at any time; it even supports importing sample photos saved on social networks and extracting the essence of the pose with a single click.

It sounds impressive, but whether this pipeline, which relies entirely on edge computing, can handle the complex and ever-changing lighting conditions in the real world remains to be seen. We need to take it to the streets to find out.

AI has its strengths and weaknesses.

Upon arriving at a coffee shop, the multi-layered halo lights on the wall caught my eye.

Holding up my phone, the XMAGE smart camera automatically zoomed to 2x, using the halo on the wall as the main subject. This choice is fairly standard; the circular halo is positioned in the lower center of the frame, resulting in a stable and harmonious overall composition.

Meanwhile, the coffee machine with its metallic texture on the control panel is also a great subject. The XMAGE smart camera accurately recognized my shooting intention, zoomed to 192mm, and kept the coffee machine in the center.

Because I wanted to preserve the indoor lighting and colors more accurately, this photo did not use any of the color styles recommended by XMAGE Smart Shot.

After taking several sets of photos with it, I think XMAGE Smart Shooting is a feature that's very user-centric.

In traditional shooting habits, the focal length of a mobile phone is something that needs to be carefully calculated. The brain will instinctively try to match the native focal length of the physical lens, for fear that falling into the middle focal length will result in a loss of image quality.

But XMAGE Smart Shooting completely abandons this obsession. It doesn't care about the zoom level; instead, it focuses entirely on the image quality: how to crop the current scene to look best. Even if it inevitably loses some edge pixels, as long as the visual center stands out and the proportions are harmonious, it will act decisively.

This approach is very practical; everything is geared towards the final product.

Looking back at the café from outside, the greenery and orange exterior walls complement each other. XMAGE Smart Capture automatically captures the part with the strongest color impact while accurately identifying the main building of the café. The two large color blocks occupy similar proportions in the frame, creating a balanced and aesthetically pleasing effect.

Looking up again, the trees are lush in summer. Using XMAGE to slightly zoom in, the branches form a natural frame composition, with the Canton Tower suspended right in the center of the image.

During the shooting of this series of photos, I also noticed a very interesting technical detail. When you press the shutter, you'll find that the filter loads much faster than the composition changes.

This is because filters typically only need to recognize the image and apply the underlying color mapping, while composition requires edge computing power to recognize architectural lines, light source distribution, and human outlines in the image in real time. The computational demands are on completely different orders of magnitude.

What's even more interesting is that Smart Shot applies filters in a very precise way. In addition to selecting the XMAGE style, it also makes fine adjustments to the color palette to ensure that the style is suitable for the current environment.

Of course, this system also has its quirks. When actually walking through the streets and alleys, XMAGE Smart Shot will sometimes automatically pop up and take over the view, and sometimes it will remain completely silent, requiring you to manually click to wake it up.

Even when manually activated, XMAGE Smart Shooting may encounter situations where a style has been selected, but the current scene cannot find the optimal composition. This frequently occurs in extremely complex environments, where the elements are too chaotic and the amount of information is overwhelming, making it difficult for the algorithm to extract a suitable composition.

Of course, if we look at it from another angle, perhaps the 1× perspective at this moment is the best composition for the current environment.

Everything can be AI-enabled; now it's photography's turn.

185 years ago, Daguerre invented the daguerreotype process, making the preservation of time a privilege; in 2011, film giant Kodak announced its bankruptcy, and smartphones began to grow wildly, stuffing lenses into the pockets of ordinary people.

Looking back at the entire history of film, it is actually a history of breaking down privileges and bringing technology to the masses.

Today, the rapid advancement of mobile imaging has hit an invisible wall.

As mobile phone camera modules get bigger and bigger, a rather meaningful rumor has recently emerged in the industry: due to high costs and physical limitations, most of the "super-sized" versions of the next generation of flagship imaging cameras will likely be discontinued.

Why is this path blocked? Because behind the rapid advancements of imaging flagships lies a dead end that has been deliberately avoided.

Ultimately, the quality of a photo is determined by the mind behind the lens. Ordinary people may have an eye for beauty, but they often lack the understanding of aperture, shutter speed, and composition. Handing a phone packed with top-of-the-line hardware to someone with no prior knowledge will likely result in a mediocre, unremarkable snapshot.

If top-tier hardware only serves a small group of photography enthusiasts, then it becomes a false proposition.

At a time when its competitors were all hitting the brakes, Huawei launched the Pura 90 Pro Max ahead of time, and XMAGE intelligent shooting was the solution to break the deadlock.

Prior to this, there had been many AI applications implemented in the industry, such as vivo's ability to display the changing seasons and OPPO's elimination of reflections. However, most of these were post-production creative adjustments and retouching after the shutter was pressed.

XMAGE has taken a giant leap forward, making algorithms the driving force of the entire shooting process.

This approach isn't exactly an artistic subversion, but it definitely works. It lowers the barrier to entry, allowing even people with no photography background to relatively easily use a flagship phone effectively.

Good technology should never be an isolated island. Smartphones once gave people the freedom to "take pictures anytime," and now AI has further empowered the public with the ability to "take good pictures."

This may be a good remedy to break the current deadlock and a way to make the most of the extra-large film.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

In the past few years, the market for new energy SUVs priced around 200,000 yuan has been like a river that has suddenly widened.

On one hand, there are the rigid demands of family users for space, range, and comfort; on the other hand, intelligent and performance-oriented features are constantly being made available at lower price points. Many features that were previously only found in vehicles priced at 300,000 yuan or even higher are now permeating the market like rain. LiDAR, air suspension, large-scale intelligent driving systems, and zero-gravity seats—these terms, originally associated with "high-end" features, are becoming bargaining chips in the new round of price segment competition.

Shouldering the important task of GAC Group's high-end development, the Haobo brand has also restructured its product architecture in line with the trend of GAC Group's revitalization this year. The Haobo A800 flagship sedan was launched recently, using Qiankun Intelligent Driving and HarmonyOS Cockpit to enhance product competitiveness.

Following closely behind, the Haobo S600 SUV, targeting the 200,000 yuan price range, has also arrived, offering four versions: the pure electric 660 Max, the pure electric 660 Max+, the range-extended 230 Max, and the range-extended 215 four-wheel drive Ultra+. The official pre-sale price is between 209,900 and 229,900 yuan, with pre-sale privilege prices ranging from 188,900 to 208,900 yuan. Positioned as a "new luxury intelligent sports SUV," the new car offers both pure electric and range-extended powertrains, with dimensions of 5015×1933×1700mm and a wheelbase of 2936mm.

Compared to its ambitious early days, Haobo is now much more pragmatic. It is trying to make the Haobo S600 a model with a high configuration density: performance, intelligent driving, cabin, comfort, safety and benefits are almost all crammed into the keyword "fully equipped".

The Haobo S600 has hired two brand ambassadors, one representing "elegance" and the other representing "sportiness".

In the past, when people mentioned mid-to-large SUVs, their first reaction was "stability": the space should be large, the seating position should be high, the rear seats should be comfortable, and it would be best if it could also handle the trivial but real scenarios of family travel.

For example, on a Friday night, a car might be packed with a folding chair, a camping table, a child's scooter, and a box of snacks in the trunk; or on a long trip home, an elderly person in the back might want to recline the seat, a child might want to watch cartoons, and the passenger might want to take a nap on the highway. For these users, the car is not primarily a "driving machine," but rather a mobile living room, a temporary rest area, and a second container for family life.

The Haobo S600 hasn't abandoned these requirements; in fact, these features have become a basic standard in domestically produced mid-to-high-end SUVs. The Haobo S600 boasts a 90% usable floor area ratio, 750mm of second-row legroom, 5.316 square meters of total glass area, and 1292L of second-row extended space. In terms of comfort features, it also includes the only 18-point massage zero-gravity queen-like passenger seat in its class, 143° reclining rear power-adjustable seats, a 22-speaker ADiGO SOUND sound system, and five double-layered laminated soundproof glass pieces.

However, the Haobo S600 does not completely package itself as a "sofa car".

On the other hand, it emphasizes the sporty attributes of posture and handling, so the Haobo repeatedly uses the expression "King of the Mountain". Obviously, Haobo hopes to make a large SUV that is more than five meters long less bulky, so that it is not only suitable for urban commuting and family travel, but also retains a certain driving participation in scenarios such as mountain roads, curves, and highway overtaking.

In terms of styling, the Haobo S600 adopts a coupe-like fastback posture with a low front and high rear. The length of more than 5 meters and the wheelbase of nearly 3 meters provide visual volume. Design elements such as double-layer ducktail, wide rear, and 20-inch sports wheels attempt to reduce the heavy feeling common in mid-to-large SUVs. From the side, it does not look as bulky and boxy as a typical large family SUV, but rather seems to be designed specifically for young families.

Therefore, Haobo has also invited two spokespeople this time: English football legend Michael Owen as the "sports spokesperson" and well-known actress Chen Shu as the "elegance spokesperson". The two spokespeople represent a different characteristic of this car and are also trying to attract potential consumers who are family-oriented and personal sports-oriented at the same time.

They put four-wheel drive, advanced driver assistance, and dual-chamber air suspension into the 200,000 yuan price range.

On the top-of-the-line Ultra+ extended-range 215 four-wheel drive version, the Haobo S600 offers both dual-motor four-wheel drive and dual-chamber air suspension. Haobo states that this is the first time in the industry that dual-motor four-wheel drive and dual-chamber air suspension have been offered simultaneously at this price point. Furthermore, the Haobo S600 comes standard with a front double wishbone and rear five-link independent suspension, SDC variable damping shock absorbers, and Eagle Claw System 2.0, and offers three road assist modes: snow, slippery, and off-road.

When a large SUV changes lanes on an elevated city road, the body needs to follow the steering wheel's movements, not react sluggishly like a large ship. On a weekend mountain drive, consecutive curves test roll control and body roll recovery. On rainy days, driving over flooded roads requires tire grip and the vehicle stability system to correct the vehicle's posture before the driver can even react. In these scenarios, dual-chamber air suspension effectively enhances the driving experience and boosts driver confidence.

In terms of powertrain, the Haobo S600 offers both pure electric and range-extended options. The range-extended four-wheel-drive version boasts a combined power output of 370kW, achieving 0-100km/h acceleration in 4.3 seconds on a full charge and 5.6 seconds with a depleted battery. GAC Xingyuan's range-extending technology provides a continuous power generation capacity of 85kW. The official specifications also mention that the vehicle can achieve a continuous lap time of 165km/h for 2 hours with a depleted battery, and a top speed of 193km/h.

The most common criticism of range-extended vehicles in the past was their poor performance when fully charged, resembling an electric car when the battery is low, leading to issues such as reduced power, increased fuel consumption, and louder engine noise. The Haobo S600's emphasis on its 85kW continuous power generation and low-battery performance addresses this pain point. A range-extended SUV shouldn't just be fast in the launch event's figures; it should also maintain basic performance in long-distance driving, highway driving, hill climbing, and low-battery scenarios.

In terms of intelligent features, the Haobo S600's strategy is also quite straightforward: try to make high-end configurations standard across the entire series.

The new vehicle comes standard with LiDAR, Qualcomm SA8650 intelligent driving chip, 27 high-precision perception units, and is equipped with GAC Xingling Zhixing ADiGO GSD 3.0, which integrates the Momenta R6 reinforcement learning world model, supporting urban and highway scenarios. The information also mentions that the Haobo S600 will be among the first to be upgraded with the Momenta R7 reinforcement learning world model in the second half of the year.

In terms of the intelligent cockpit, the new car is equipped with a Qualcomm 8295P cockpit chip, a 17.3-inch 3K central control screen, a 27-inch W-HUD, and 11 built-in AI intelligent agents. Compared to the past when the in-vehicle system was only responsible for navigation, music, and voice control, the current cockpit is becoming more like a mobile terminal: it not only needs to respond to commands, but also understand the scenario, and even know in advance what the user might need in the car.

After getting off work, you sit in the car, the navigation system automatically provides the route home, the HUD projects lane-level information in front of you, the child sits in the back watching cartoons, the car's infotainment system adjusts the fresh air intake based on the car's temperature and air quality, and the seat massage starts working… These experiences that are possible in new energy vehicles are undoubtedly a leap forward for users who have switched from traditional gasoline cars in the 200,000 yuan price range.

GAC Group has a large customer base, and if these customers want to purchase or replace their vehicles with new energy vehicles, GAC naturally wants to meet their needs with its products. Just like AI says, "I'm here, straight ahead, without flinching or running away, firmly catching you."

To some extent, in the past few years, GAC Group's development of new energy products has focused too much on the B-end and has not been attractive enough to C-end users, failing to "steadily meet" user demand.

For consumers, the 200,000-yuan-level new energy SUV market is already quite crowded, with models like the Zhiji LS6, Jike 7X, and XPeng G7 all putting pressure on this segment. For the Haobo S600 to break through, the key issue isn't "whether it has a lot of features," but rather whether these features can be organized into a clear and perceptible driving experience: quiet and composed in the city, not cumbersome on mountain roads, stress-free on long journeys, relaxing for family members, and engaging for the driver.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

The "Father of Lobster" consumed 603 billion tokens in a single month, totaling a staggering nine million RMB.

China Mobile, China Unicom, and China Telecom, the three major telecom operators, are all promoting Token packages. For just 199 yuan, you get gigabit broadband and 100 million tokens. Interested?

From Silicon Valley to major domestic companies, Tokenmaxxing has become the mainstream practice. Whoever consumes the most tokens is considered a good employee in the AI ​​era.

A Gen Z alumnus donated 2 billion tokens to his alma mater, prompting netizens to jokingly say that based on DeepSeek's price of 5 yuan per 100 million tokens, it would only cost 100 yuan.

Within six months, tokens underwent a transformation: from technical terminology to KPIs, to phone plans, to donation currency. They became the "measuring stick" of the AI ​​era, the only problem being that no one can clearly define what they are actually measuring.

We buy our own tokens, use the company's tokens, and deploy a bunch of agents. Code, papers, and weekly reports are all generated by burning tokens.

On the other hand, employees of large companies, due to the token consumption ranking system, have started using company tokens to handle personal matters, play games, and develop dozens of useless sub-agents to improve their own rankings.

"Returns" are difficult to quantify, but "usage" can be quantified.

So everyone chose the one that was easy to quantify. This isn't a new problem in the AI ​​era; it's an old problem in management.

The company that uses AI to eliminate bullshit jobs is creating new bullshit jobs.

Amazon, the company that laid off so many employees that it turned its website into a 404 error, has recently been the subject of another "laughable" story.

Originally intended to eliminate "bullshit jobs," AI has instead become a source of new "bullshit jobs."

According to the Financial Times, in an effort to force employees to embrace AI, Amazon has come up with an extremely retro management method: a "Token Consumption Leaderboard" that tracks each employee's usage.

The company mandates that over 80% of developers must meet AI usage metrics each week, and even uses the amount of tokens consumed as a performance evaluation criterion.

The workers reacted directly: since the company uses this metric for performance evaluation, everyone decided to fight fire with fire and started the "Tokenmaxxing" tactic.

Just then, Amazon launched an AI agent called MeshClaw, which can initiate code deployments, organize emails, and manipulate Slack. An internal company memo described it as: "It integrates what it learns during the day by dreaming at night, monitors your deployments while you're in meetings, and categorizes your emails before you wake up."

MeshClaw thus became a tool for manipulating leaderboards. Developers started using it to plan trips, process private emails, and have AI analyze the silly things product managers say on Slack.

On Team Blind, an anonymous workplace community (a message board for verified employees of companies like Google and Apple), a post by an Amazon employee received a flood of upvotes.

I'm burning through tokens like crazy just to rant about my product manager. Whenever he spouts nonsense on Slack, I dump the chat logs on to AI, activating 10 sub-agents to do a comprehensive, in-depth analysis and critique of him. This is absolutely the perfect use of GPU computing power.

In its response to the Financial Times, Amazon stated that MeshClaw "helps thousands of employees automate repetitive tasks every day," and that the company is "committed to the responsible deployment of generative AI." The company also stated that token statistics will not be used for performance evaluation.

But the employees said, "Managers look at this data. When they track usage, it creates distorted incentives, and some people become very competitive about it."

The company says it doesn't count towards KPIs, but the manager is secretly watching. This is the same tactic used by big companies to say "year-end bonuses are unrelated to the 996 work schedule."

It's not just Amazon; Meta employees are doing the same thing.

Back in April, The Information reported that an employee of Meta used internal data to create a dashboard on the company intranet, allowing colleagues to compete to become the company's number one AI Token user.

This ranking compiles the AI ​​usage data of over 85,000 Meta employees and lists the top 250 super users, with Zuckerberg not making the top 250.

The ranking list was taken down two days later. In response to media inquiries, Meta issued a statement saying, "The employee made the decision to remove the dashboard on their own; Meta did not request this action."

After you laugh at the absurdity of this ranking, you'll realize that this actually reflects the reality for most companies. They haven't figured out how AI will be used, but they've already started laying off employees; they haven't figured out how to use tokens, but they've rushed to use them as a productivity measure.

What did the 600 billion tokens burned in a month produce?

Before we could even process the absurdity of the Token Consumption Ranking, something even more surreal happened.

Three alumni born in the 2000s donated 2 billion tokens to their alma mater, Zhengzhou Sias University. Netizens calculated the value based on DeepSeek's pricing and said it was worth 100 yuan.

Later, media outlets clarified that the 2 billion tokens included not only API call volume but also the right to use the generation tool and platform points. But the act of "donating tokens" itself is already bizarre enough.

Three alumni said they didn't have the financial means to donate a teaching building, so they donated tokens. The logic of charity in this era is also being refreshed: if you can't afford to donate a building, donate computing power.

The value of tokens is being refreshed, and the boundaries of token usage are also being refreshed.

Nat Friedman, former CEO of GitHub and current CEO of Meta Superintelligence Lab, told a story at a public event. One day, his OpenClaw program determined that he wasn't drinking enough water, so he casually gave the instruction: "Make sure I get enough water at all costs."

OpenClaw acted quickly. It instructed him to go to the kitchen and drink a bottle of water, casually mentioning that it was monitoring him through his home security cameras to ensure he actually did so. After he complied, OpenClaw sent him a screenshot of him drinking water with the comment, "Well done."

What was originally just a phone setting to remind you to drink water daily has now turned into a crazy burning of tokens, using the camera to "remind you to drink a glass of water".

When the consumption of tokens is no longer important, and we no longer need to consider the value and usage boundaries of tokens, we will use them for something else.

The most interesting thing about OpenClaw recently is Peter Steinberger's sharing on X on Saturday. He posted a screenshot of CodexBar with the caption "CodexBar's latest update makes API fees more user-friendly."

However, netizens soon discovered that this screenshot was remarkable, showing that 603 billion tokens were used in 30 days, with a total consumption of $1.3 million, equivalent to about 9.3 million RMB.

The comments section is full of questions: How much code have you delivered? What's the ratio between the tokens consumed and the usable code? What useful things have you created so far? If you hadn't joined OpenAI, would you be allowed to consume Codex tokens like this?

Dude, you'd better come up with something that even million-dollar engineers can't make, or this could be the beginning of the bursting bubble in cutting-edge labs. And this is with a subsidy price, my god. If it were the actual cost, the price would definitely be higher.

The "father of lobster" responded to these comments, mentioning that turning off Fast Mode could reduce costs by 70%. Furthermore, since OpenClaw was acquired by OpenAI, only three members remain in charge of the project, running 100 instances on Codex.

These instances automatically handle various issues in the software development process, such as code submissions, bug fixes, and feature updates.

But looking at OpenClaw updates alone, does it really need $1.3 million to support it? He also mentioned that he is working on some startup projects other than OpenClaw, and that he is exploring a question: how would software be built if token costs were not important?

That's a good question. But after spending $1.3 million, he still hasn't gotten an answer.

This could be the most expensive question mark of 2026.

Even those with immense computing power seem to have no idea what these tokens can be used for.

Executives at major tech companies, looking at the huge GPU procurement costs on their financial statements, urgently needed to prove to their boards that the money hadn't been wasted. Since "reconstructing the real business flow" was too difficult, too slow, and required too much courage, they opted to settle for second best and assess "token consumption."

Employees weren't even asked "How do you think tokens should be used?" at the beginning; they were asked "How much have you used this week?"

When a tool's performance is measured by "consumption" rather than "output," it ceases to be a tool. It becomes fuel, its sole purpose being to be burned. What it powers after it's burned is of no real concern to anyone.

Because if you really ask, many people will find that the tokens they burned, just like those laid off at the beginning of the year, didn't get them anything in return.

What we are experiencing is a game where everyone pretends to understand the rules. Companies pretend to know how to use them, employees pretend to use them diligently, and investors pretend to see returns.

The only thing that's real is the ever-increasing bills.

Tokens will eventually find their true purpose and become a genuine "new form of productivity." But before that day arrives, before we burn through hundreds of millions of tokens, we can ask ourselves if it's really necessary.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

1.

OpenAI's two arch-rivals, Anthropic and Musk, finally formed an alliance at the beginning of the month after putting aside their prejudices.

Prior to this, Anthropic and Musk had a strained relationship: in February of this year, Musk accused Anthropic on his X account of being "woke," "evil," and "misanthropic," saying the company was "anti-civilization."

In retrospect, this attack was not due to Musk's unconventional personality, but rather because something Anthropic did touched a nerve with him, and there was a reason behind it.

Prior to this, xAI internally used Cursor, but at the beginning of this year, employees discovered that the Claude model was suddenly unusable in xAI's Cursor corporate account.

Wu Yuhuai, the co-founder who was still working at xAI at the time, said this in an all-staff email: "Anthropic has updated its policy, requiring Cursor not to provide its main competitors with the ability to call Claude models."

At the time, Wu Yuhuai wrote a rather interesting sentence in his letter:

"This is both bad news and good news. Our productivity will be affected, but it also urges us to develop our own coding products and models."

Why did xAI's top management believe that developing their own coding products was crucial at the time?

What happened next is well-known. The entire founding team of xAI absconded, and Musk, in a fit of anger, used his financial power to deliver the final blow to Cursor:

At the end of last month, SpaceX and Cursor jointly announced an unprecedented strategic partnership to train AI models for programming and knowledge-based jobs; in addition, SpaceX also acquired the right to acquire Cursor for $60 billion or pay the latter $10 billion in cooperation fees.

Note the key qualifier " programming ," which will be used later in the call-back section.

2.

Recently, I watched a video of Theo Browne, an early investor in Cursor, a vocal critic of Anthropic, and the founder of T3.

I originally clicked in to see him criticize AstraZeneca and SpaceX for their underhanded dealings, but unexpectedly, I found a unique yet extremely reasonable analysis about the SpaceX + Cursor collaboration:

Leaving aside the 60 billion acquisition, just considering the 10 billion cooperation fee— Theo stated in the video that he believes "even if it's just exchanging Cursor's user data, this 10 billion is worth it."

Soon after, netizens discovered Kimi's model ID in the publicly available code snippets, and screenshots spread throughout the developer community, forcing Cursor VP Lee Robinson to clarify: "Composer 2 did indeed originate from an open-source platform. Ultimately, only about 1/4 of the model's computing power came from the platform, while the remaining 3/4 was trained by us."

A few hours later, Cursor co-founder Aman Sanger also posted an apology: "It was a mistake not to mention the Kimi base at the beginning."

Five days later, Cursor released the complete Composer 2 technical report, showing that the base was indeed Kimi K2.5, the licensor was Firworks AI, and the general process was to train on K2.5 and then continue with large-scale reinforcement learning (RL).

The key point is that Composer 2's RL runs in a real Cursor session, using the exact same tools and harnesses as the production deployment.

Cursor calls this process "real-time reinforcement learning," which means deploying the model's checkpoint directly to the Cursor production environment to observe user responses, collect data, and aggregate it into reward signals—it can iterate the model version as quickly as every 5 hours, and then continue to deploy it to Cursor, repeating the cycle.

The most extreme example is Cursor's Tab feature for automatic code completion, which processes more than 400 million requests every day. Whenever a user enters characters or moves the cursor, the model predicts the next action. If the prediction confidence is high, it displays suggestions, and the user presses Tab to accept the auto-completion.

This feature utilizes online reinforcement learning, a unique feature in the industry. Cursor can update Tab's model capabilities to users at an extremely high frequency (as fast as every one and a half to two hours), collecting on-policy data directly within the product for training.

This high-frequency, near-real-time feedback loop allows Tab to learn extremely subtle user intentions. Cursor revealed that this method reduced Tab's rejection rate by 21% and increased its acceptance rate by 28%.

Returning to the Composer model itself, after things were clarified, some Kimi employees deleted their previous sarcastic tweets, and the official Kimi account posted congratulations.

A coding agent application layer company valued at $60 billion (based on Musk's figures) that doesn't build its own model base can still use its own data flywheel to pull out proprietary programming models that go beyond the base model.

So rather than saying Cursor crashed, it's more accurate to say that this is a perfect example of the importance of coding agent products.

In another article about real-time RL, Cursor wrote: "(Training a programming model) The biggest challenge lies in modeling the user. In Composer's production environment, there is not only a computer executing commands, but also people supervising and guiding it. Simulating a computer is easy, but simulating the people using it is difficult."

This statement is gradually becoming a consensus among leading model vendors in the field of programming models. If you look at benchmark lists and general user reviews, you'll find that the top vendors are all investing heavily in their own coding agents/programming products. The only difference lies in who is closer to the user.

Taking relatively authoritative ranking lists such as SWE-bench and LLM-Stats as examples, models such as Claude, GPT, Gemini, and Kimi basically dominate the top ten. They are all model vendors that have developed their own coding agent products (including CLI, IDE, and desktop clients that integrate coding agents).

A few counterexamples appear on some lists, such as Meta ( Muse Spark) and DeepSeek, which have not developed their own coding agent.

However, you'll find that these counterexample models struggle to rank on more authoritative benchmarks that are closer to real-world scenarios and avoid contamination. For instance, DeepSeek scores 70% and ranks ninth on SWE-bench bash only, but its score drops to around 15% on SWE-bench Pro.

OpenRouter's real-world traffic data can explain this discrepancy: the platform's 2025 report shows that over 80% of Claude token consumption was used for programming and technical tasks, while DeepSeek token consumption was primarily focused on casual conversation and role-playing.

Vendors without their own coding products may be able to rank high on some coding task benchmarks, but they will be exposed as incompetent on more challenging real-world engineering benchmarks and in real-world traffic where users consume tokens to vote.

Not only Cursor, but Anthropic also explicitly revealed in a paper published in November 2025 that it was doing the exact same thing: "We train on Anthropic's own real production programming environment." That is, Anthropic feeds back the interaction data of its employees using Claude Code to the Claude model for training.

5.

In the evolution of AI, the definition of production factors has undergone a profound shift. While the three traditional core elements—computing power, research, and training data—continue to grow in total quantity, they have become severely imbalanced in structure.

Today's major AI giants have significantly increased their capital expenditure (CapEx) on computing power, making computing infrastructure the dominant theme in current public discourse. However, in reality, especially within the field of programming, as model vendors exploit publicly available code data on the internet, such as GitHub repositories and Stack Overflow, in a "draining the pond to catch all the fish" manner, the boundaries of models in code generation and logical reasoning are gradually becoming apparent.

This is why industry consensus is gradually shifting towards a rising new strategic high ground:

For any model vendor that wants to master top-level coding capabilities, building its own coding agent product is no longer an optional business path, but a core lifeline to ensure the continuous evolution of the underlying model.

As APPSO argued earlier, simply learning from publicly available data is like only learning the outcome of successful people, without understanding the path to success. This is definitely not what a proper understanding of success should be. In a real programming environment, knowing what errors occurred, how they occurred, and how to correctly understand and efficiently implement requirements—understanding the correct process—is far more valuable than obtaining the correct result itself.

Only by possessing their own coding products can model vendors obtain high-quality "process supervision" signals, thereby ensuring they maintain a technological moat in the next stage of competition in coding/inference capabilities.

Otherwise, they would have to do what SpaceXAI did and spend money to partner with coding agent product companies.

However, not all model manufacturers are as wealthy as Musk, and the division of power, alliances, and territorial struggles among giants starting in 2026 will become even more intense. When a model manufacturer lacking its own coding products finally realizes this, it may not have enough partners to choose from, and the price of cooperation will also rise accordingly.

The situation of major US model manufacturers is generally well-known, so we won't go into detail here. APPSO has also noted that most of the mainstream domestic model manufacturers and AI giants have already made inroads into coding agent products.

Domestic giants are mainly working on native AI IDEs or IDE plugins: ByteDance launched TRAE very early last year, Alibaba has Qoder, Tencent has CodeBuddy, and Baidu has Comate, etc.

Among AI companies, Moon's Dark Side was the first to develop an independent coding agent product, mainly Kimi Code with a CLI interface. However, Kimi had previously revealed that CLI would not be the final form of native programming products .

Another approach is for model vendors to provide their own API services and coding plans. This way, regardless of the AI ​​development environment used by the user, the model vendor can obtain process data that closely approximates the native coding product through server-side API records.

However, this is only a close approximation, not an exact resemblance. The core issue is that the request-response logs of the server-side API still differ significantly from the deeply inherited product interaction patterns.

Vendors with their own built-in products (such as Cursor, Claude desktop, and Codex) have the most direct explicit feedback signals, while the API side provides relatively vague implicit inferences. Simply put, the API side can see user requests and responses, but it has no idea whether the user ultimately adopted the code, whether the code runs successfully, or what bugs it caused. They cannot understand this crucial label of the user's final behavior, thus failing to achieve the highest quality reinforcement learning.

Metaphysically speaking, language is the world, and code is the solution. Code can express the vast majority of tasks in this world, and it also acts as an amplifier, allowing top talents to amplify their productivity many times over.

Only the best coding models deserve the best talent. If leading model vendors don't value coding, they will inevitably fall out of the top tier.

Of course, in reality, every model manufacturer will value coding—but rather, under the new paradigm, those products without their own controllable native coding agent are very likely to gradually fall behind manufacturers that do have such products.

Just a few days ago, MiniMax also released a major update to its desktop client product: the Mavis feature with a brand-new multi-agent orchestration architecture, which also significantly improves the client's support for coding tasks.

Previously, MiniMax only launched a desktop version, but did not include native coding and agent features.

Following that, on May 15, Alibaba officially released Qoder 1.0 – this product was officially upgraded from an IDE to a complete Agent product (Alibaba's official name is the Intelligent Agent Self-Development Workbench).

At the same time, xAI's Grok Build CLI has finally been officially launched.

That's right, it's the coding agent that xAI came up with themselves after their accounts were banned by Anthropic and Cursor earlier this year.

And now, there are several more ready-made cases.

It seems that everyone agrees that Cursor, Codex, and Claude desktop clients are on the right track.

6.

The same applies if we extend the discussion from coding to the agent itself.

While some trajectory data for coding tasks can indeed be found in publicly available corpora (such as GitHub commit records/PRs, although the quality is not high), trajectory data for agent tasks, including but not limited to moving and clicking the mouse, manipulating the touchscreen, and filling in input boxes, cannot be found in publicly available corpora.

Therefore, we see that even in the smallest implementation path of agent operation—the browser plugin, something that seems not high-end at all—almost every model vendor makes its own.

OpenAI launched Operator back in January 2025—rather than calling it an "AI-automated browser," it's essentially a large-scale data collection device. Every user who tries Operator is providing OpenAI with on-policy data for free.

Subsequently, OpenAI spawned ChatGPT Agent and a new version of the Codex desktop application; the same applies to Anthropic; recently, Kimi quietly created a project called WebBridge, which is essentially a browser plugin.

Even Deepin, the Chinese model giant that has been most restrained in its actions over the past two years, has recently begun to show interest in Agents.

In a previous interview, CEO Liang Wenfeng mentioned that mathematics and code are natural testing grounds for AGI, somewhat like Go, a closed and verifiable system that has the potential to achieve high intelligence through self-learning.

The subtext of this statement is that DeepSeek has always treated coding and agents as research and testing grounds, rather than commercialization.

However, in March of this year, DeepSeek released more than a dozen agent-related positions at once, including the first-ever Model Strategy Product Manager (Agent-focused). The job description at the time covered "leading the design of agent evaluation systems and training data solutions," and required "deep use of products such as Claude Code and Manus."

APPSO noted that DeepSeek recently posted job openings for positions such as Agent Product Manager and Harness Product Manager—clearly, DeepSeek is going to create an independent, native Coding/Agent product.

Previous reports indicated that DeepSeek V3.2 incorporated nearly two thousand synthetic agent training environments and over eighty thousand complex instructions during its training process. However, it seems that synthetic training data can only take DeepSeek this far; the remaining portion—the real successes and failures of real users in real-world environments—can only be obtained through their own agent products.

DeepSeek has been developing its models and products with extreme restraint for three years ( only adding multimodal capabilities to its official website last month). However, today, it's becoming increasingly difficult for DeepSeek to achieve state-of-the-art (SOTA) performance in coding tasks, and even those it previously achieved are soon surpassed.

When the main force could no longer support the flywheel by relying on research, DeepSeek finally took action.

7.

Finally, let's return to the story at the beginning.

According to The Information, citing sources familiar with the matter, while accepting Musk's $60 billion acquisition offer/$10 billion cooperation, Cursor stated that it will not collaborate with xAI to develop new models, but will instead focus on optimizing its own Composer model.

This could mean that even if Cursor is bought or acquired by Musk, it will still need to retain the core of its data flywheel.

The ownership of data itself is the most crucial hidden point of contention.

When all the top model manufacturers have made their own products, and all the top product manufacturers have started training their own models, the already blurred line between "model companies" and "product companies" seems to be disappearing more and more…

This game has only just begun.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.