Pogbagram

I assigned a task, and the agent activated the plan mode, outlining 7 steps.

I approved it, and it started running. After three steps, it stopped and reported: "I have completed steps 1, 2, and 3, and the results are as follows… Shall we continue with steps 4, 5, 6, and 7?"

I said continue. It ran two more steps, then stopped again: "I've completed 4 and 5, and the results are these and which… Shall we continue with 6 and 7?"

After a whole night, when I asked the agent to do some long-term tasks, there were no long-term effects; the dialog box just kept showing "Continue".

This has been my experience for a long time, using various agents to get things done.

This experience is illogical. While "stopping to confirm" is a good work habit when working with AI, in many tasks I never actively asked it to stop, but it would still stop anyway.

In its latest technical blog post, MiniMax attributes this behavior of its agent products to "contextual anxiety." The core issue is that the model itself is ambiguous about when a very long task is considered complete. Simply put, it's not that they can't do it, but that they're afraid to. They're afraid of making a mistake at every step, which is why they stop halfway through and ask questions.

Today, the MiniMax Agent desktop client underwent a major update. A new mode called Mavis has been added (actually, it's an abbreviation for "MiniMax as a Jarvis").

It's well known that having one agent act as the boss and a group of agents as employees—this traditional multi-agent framework is nothing new. However, MiniMax points out that previous mainstream multi-agent frameworks essentially relied on cue word arrangement to let the model play a "role-playing" role. But this approach doesn't last long, as it encounters problems such as context anxiety, long-term task degradation, and self-checking issues, as mentioned earlier.

Multi-agent systems require a reliable infrastructure that is continuously running and maintained, and where multiple agents do not "collude." This is what MiniMax does.

Real-world testing experience: Let the agent "nitpick" the other party

MiniMax calls its Agent Team infrastructure the Team Engine, which has three core roles: Leader, Worker, and Verifier. As the names suggest, one manages, one performs the work, and one verifies.

The most crucial difference is that the Worker and Verifier are in an "adversarial" relationship, and neither can get away with it.

A while ago, APPSO was researching a topic: "All model vendors with ambitions in coding/agent should develop their own independent coding/agent products."

(That's right, MiniMax was a negative example before, but unexpectedly, it proved itself even before the article was published!)

So we ran this problem again on MiniMax's Agent Team.

This task was divided into 5 workers. After each worker completes its task, it will organize the results and submit them to the leader (displaying the status as "Mavis sent to General" or "General sent to Mavis", etc.).

A worker had been running for 12 minutes without returning any results. APPSO noticed that the leader was getting impatient, so it sent a bash command to check its status:

After all 5 workers have completed their tasks, the leader generates 5 verifiers—displayed in the task list as agents wearing "yellow hats":

The verifier quickly found the error! One of the verifiers discovered a clear data error in the corresponding worker's deliverables and issued a "failure" penalty. Immediately afterwards, the corresponding worker restarted (displayed as running, indicated by a small blue circle).

Click into the corresponding worker's workspace to observe its thought process: "The verifier rejected my previous deliverables based on the following three errors… I need to go back and re-verify the key facts and check and correct the specific numerical issues…"

And I must say, the agents are all "uncompromising" with each other, making them really reliable in their work.

This back-and-forth occurred dozens of times in the five 1v1 agent battles. During the process, Mavis also said that he "learned something new" and updated his memory.

While the previous task is underway, we will begin a new in-depth study, analyzing the tourism market during the May Day holiday based on authoritative data and delivering a multi-dimensional analysis report.

This research is far more complex than the previous task. Moreover, because of the ongoing confrontation, the Agent Team spends significantly more time on in-depth research than a typical single agent.

However, the final report was indeed much cleaner and more credible compared to other AI in-depth research deliverables.

APPSO has been preparing for many offline events recently, and planning and devising solutions has always been a challenge. We've also entrusted this task to Mavis to see how it goes.

I need to plan an offline AI developer salon in Guangzhou. Please provide me with as many venues as possible suitable for tech events with hundreds or thousands of attendees, along with approximate quotes, and information on similar events. Then, please help me plan the theme, promotion, and operation of this AI event, compiling all of this into a rigorous business plan format, as well as a beautifully designed website that matches the theme.

The planning process alone took longer than previous in-depth research tasks. Mavis replied, "This task is large-scale and requires multiple agents to work in parallel—site research, competitor analysis, theme planning, business plan, and website development."

Mavis's strength lies in its ability to continuously add new requirements:

In addition to the long report, it would be best if you could also draft a preliminary formal contract, including contracts for cooperation with the venue, cooperation with invited guests, and other possible contracts, as well as preliminary financial statements. Also, please provide a PowerPoint presentation to showcase this plan, the more detailed the better.

Upon receiving new requirements, the Agent Team further refined the plan and launched more workflows. In the end, we launched as many as nine parallel tasks.

If we open Mavis's thought process, we can see a large number of messages sent between agents. These agents work under a dedicated Team Engine, transmitting each other's status; some are waiting, some are executing, and some are verifying.

Look at this Verifier, doesn't it resemble a nitpicking "client"?

The final number of files delivered by the entire task reached an astonishing 10 or more, including xls, ppt, html web pages, and corresponding .md versions.

Next, let's talk about another major feature of Mavis: it can connect to chat platforms and supports multitasking.

Similar to OpenClaw and Hermes Agent, which MiniMax already supports, Mavis can also assign tasks through WeChat and Lark, two IM platforms. The integration process is extremely simplified; simply click the settings button, scan the QR code, and name the application, and you can use Mavis within WeChat/Lark.

When a typical agent product connects to an IM, and we assign it a task that takes a long time to complete, it often means that after the message is sent, we can no longer consult it on other issues.

One reason is that these agents cannot open multiple dialog windows simultaneously; another reason is the limitation of the agent's working mode. Running multiple tasks in a single session can easily lead to contextual confusion and pollution.

MiniMax's solution is to decouple the logic of "instant response" and "execution".

I had APPSO research the recent oil price surge in Lark; after the task started, I also had it research the important products released by Silicon Valley AI giants in the past month.

Mavis didn't stop the previous task, but instead told me that the new task was already completed, while the task about rising oil prices was still being processed.

This is another key design principle of Mavis: the benefits of context isolation.

Each Agent Team, and each agent within the team, only sees a summary of information relevant to their own mission, and only reads the full text when details are needed.

This approach has two advantages: firstly, it keeps token costs under control, preventing the context from easily overflowing even with a large team; secondly, it prevents context pollution, ensuring that incorrect information encountered by the agent during searches won't wipe out the entire team.

In the most extreme scenario, we tried assigning him 8 tasks in a very short time using Lark, and there were no instances of context confusion.

The whole experience is a lot like working with a colleague with extremely high cognitive bandwidth: not only can they reply to messages instantly, but they can also work in the background without being interrupted. If you want to know the progress, you can just ask directly without worrying about disturbing their "flow state".

Agents handling different sessions only see information relevant to their own tasks and do not share an ever-expanding conversation history.

In short, Mavis achieves end-to-end context isolation, from the IM channel to the task hub, and then to each molecular agent in the molecular task.

Finally, while answering questions about the new AI products released this month by major AI companies and important embodied intelligence products, it also successfully completed the main thread of the oil mission, giving us a detailed report that even mentioned the recent news that Japanese potato chip packaging is going to turn black and white.

After testing, did you notice that Mavis's arrangement strategy is actually somewhat similar to the "Three Provinces and Six Ministries" skill that was popular for a while?

What each character does, when it starts, and when it hands over will be determined by the state machine at the engine level, rather than by the black box of the model making its own decisions.

In short, this means using engineering-level controllability, rigor, and determinism in multi-agent work orchestration to fundamentally address the uncontrollability and randomness of the model.

This approach completely solves the classic problem of past agents/models "acting as both referees and players".

Uniform credit limits, ample agent availability.

After testing Mavis, let's talk about another equally important thing MiniMax did that affects all paying users: this time, the Token Plan and Agent Plan have been merged.

After the merger, whether it's for ordinary users' "daily use," such as communicating and using the Agent on the official website and in the app, or accessing the official API to call other tools (such as coding products or OpenClaw/Hermes Agent), a unified plan can now be used. Furthermore, both M2.7 and subsequent flagship models, as well as multimodal models for music, video, and voice, are all included in this single plan.

All credit limits are shared, and users can decide how to spend them. MiniMax also offers a bonus: users who previously subscribed to two plans simultaneously will receive an extra month of membership.

Why do this? From the user's perspective, it's actually quite reasonable.

To put it simply, in the Agent era, users' motivation to pay comes from the demand for "model computing power". As the models improve in coding, agent and multimodal capabilities, the scenarios for these demands will only become more diverse and will naturally occur in model vendors' products (official website, independent products, CLI) as well as outside of products (independently deployed agents that access external APIs).

This is actually a problem that all major AI giants are facing: OpenAI currently separates user subscriptions and API billing, as does Anthropic; as for smaller agent startups, they use their own subscription fees to pay the underlying API fees instead of users paying for them.

This time, MiniMax took the lead in dismantling the internal walls of its product matrix. APPSO believes that in today's highly commoditized market where users always flock to the newest and cheapest model APIs, this unified package strategy actually helps model manufacturers maintain user loyalty.

Let's go back to the product itself.

As mentioned earlier, APPSO is writing an article about "model vendors who are serious about coding/agents must develop their own coding/agent products." MiniMax can be said to have arrived late, but it's not far off.

Today, Mavis is not the first product to bet on a multi-agent architecture. In the past six months, companies such as ChatGPT, Manus, and Genspark have all joined this "multi-agent" war.

After completing the actual test, APPSO's impression was that Mavis performed better and had a more stable architecture than its competitors in terms of "the product running an extremely complex/long-term task on its own." While other products' multi-agent approaches were limited to prompt word arrangement and task splitting, Mavis implemented adversarial hard constraints at the engineering level—the resulting difference was quite significant.

However, while this architecture looks promising, there's an unavoidable reality: it's expensive.

MiniMax introduced the concept of "Cost of Consensus" in its technical blog. In layman's terms, while several agents "check and balance" each other, making the process and results more reliable, the process of reaching consensus has a cost, with token consumption being several times that of a single agent; moreover, just like arguing, getting into a heated argument can lead to straying from the topic, and the accuracy may even decrease instead of increasing.

According to MiniMax's analysis, its Agent Team architecture specifically has three types of costs:

First, there's the handover cost. Information needs to be reorganized when it's transferred between agents. Each handover requires "translating" the information into a form that the next agent can use, which consumes tokens.

Secondly, there's the cost of sharing (context information). Context isolation is designed to control this cost to some extent. However, even if each agent only looks at the "summary" passed from other agents, as the size of the agent team increases, storing and distributing the summaries will still incur costs.

Thirdly, there's the cost of aggregation. APPSO has always wanted to emphasize this point: don't assume that a workflow with hundreds or thousands of skills and an extremely complex "three provinces and six departments" system is the ultimate solution—it's often not. In fact, you might be falling into the trap set by token vendors… You may have made the work more detailed, but you also need to spend more tokens to aggregate and organize the final results.

These costs combined mean that having multiple agents is never a simple matter of "the more agents the better".

However, from another perspective: the more complex the information exchange in a task, the higher its inherent value often is. A thorough research report requiring multiple verifications and repeated checks, and a casual question, shouldn't be measured by the same logic. Mavis is expensive because of its meticulousness, and those meticulously handled tasks are worth the price.

They would rather spend more money to ensure everything goes perfectly than do a shoddy job; this is what high-value users behind complex tasks value.

Of course, the MiniMax team also did some engineering design to avoid token waste caused by program redundancy.

MiniMax's advice to users is that Agent Teams are for "expensive and complex" tasks; they are a strategic option, not the default. Users should assess the task's complexity, workflow length, risk, and the value of experience reuse—the higher these factors are, the more worthwhile it is to use Agent Teams. Conversely, a single agent or even a regular chat can be used.

Does having more agents necessarily mean more intelligence? Not at all. But the significance of Mavis is that it allows truly complex, knowledge-intensive tasks to be handled by a proven engineering system with adversarial mechanisms, verification, clear division of responsibilities, and reward/punishment systems, instead of letting the model make decisions on its own.

It may not necessarily make AI smarter, but it will definitely make it harder for AI to slack off—which is a long-standing problem for large models themselves.

After all, in real interpersonal work, we don't really need our colleagues to be very smart… just don't be lazy or try to be clever, that's often enough, isn't it?

By Du Chen and Zhang Zihao

#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.

iPhone 17 Pro price officially reduced by 1000 yuan

Last night, Apple unexpectedly launched a 618 promotional event for the iPhone 17 Pro series on Tmall's "Apple Store Official Flagship Store". The Pro and Pro Max models were reduced by 1,000 yuan, with starting prices of 7,999 yuan and 8,999 yuan respectively.

This event also covers iPad, Apple Watch, AirPods and other series. Non-iPhone products can use surprise coupons ranging from 50 to 100 yuan, which can be combined with 88VIP consumption coupons, store discounts and national subsidies and other multiple benefits.

ChatGPT's integration into Siri fell short of expectations; OpenAI reportedly plans legal action against Apple.

According to Bloomberg, the partnership between Apple and OpenAI has shown clear signs of strain, and OpenAI is working with external law firms to assess possible legal action.

Sources familiar with the matter revealed that OpenAI's legal team is collaborating with external law firms to explore several options, including sending a breach of contract notice to Apple, but no formal lawsuit has been filed yet. OpenAI still hopes to resolve its dispute with Apple out of court, and related legal action is expected to proceed only after Musk's related lawsuits have concluded.

Apple's collaboration with OpenAI began in June 2024, when Apple integrated ChatGPT into Siri and iOS at WWDC, covering scenarios such as text generation, visual intelligence, and the image generation app "TuLeFun".

With a valuation of $4.65 billion, Tian Yuandong, former director of Meta Fair, has officially launched his startup.

Yesterday, Tian Yuandong, former director of Meta FAIR Research Institute, officially announced the co-founding of AI startup Recursive Superintelligence, exactly six months after confirming his departure from Meta.

He posted on X that he rejected offers from several leading AI companies, including OpenAI, xAI, and Anthropic, and chose to start his own business: "While I'm still young, I'm going to be a co-founder of a new startup."

Recursive also announced the completion of a $650 million funding round, valuing the company at $4.65 billion. The round was led by GV (Google Ventures) and Greycroft, with participation from AMD Ventures and Nvidia.

The company's core focus is on "recursive self-improvement," which means building an AI system that can automatically discover knowledge and continuously optimize itself.

Claude has reinstated access to Lobster, prompting developers to say it's a disguised price increase.

Anthropic officially announced yesterday that starting June 15th this year, it will introduce a separate monthly quota for the Agent SDK for Claude subscription users. As a result, third-party Agent tools such as OpenClaw will be allowed to access Claude through subscription accounts again.

According to the new regulations, calls to the Agent SDK will be completely separated from the user's regular subscription limit and replaced by a dedicated monthly quota pool. The quota will be refreshed according to the billing cycle, and any unused portion will not be carried over to the next month.

Pro and Team Standard seats receive a $20 monthly credit; Max 5x and Team Premium seats are $100; Max 20x and Enterprise Premium seats can be up to $200.

This quota is dedicated to Claude Agent SDK projects (Python or TypeScript), claude -p non-interactive commands, Claude Code GitHub Actions integration, and third-party applications certified by the Agent SDK.

This new rule is seen as Anthropic's formal solution to the ban policy implemented on April 4th of this year.

At that time, Boris Cherny, head of Claude Code, announced that Claude Pro and Max subscriptions would no longer cover calls to third-party tools such as OpenClaw, arguing that the usage patterns of such tools were putting unsustainable pressure on Anthropic's computing infrastructure.

The developer community reacted coldly. Some developers bluntly stated that the $20 Agent credit was consumed extremely quickly under high-intensity tasks, with the actual usable amount being far lower than the subscription plan before the ban, and characterized this adjustment as a "disguised price increase."

Tencent denies rumors that its top AI executive will leave the company.

Yesterday, Tencent officially issued a statement refuting rumors circulating online that its "AI leader is about to leave the company." Tencent's PR director, Zhang Jun, stated, "This is outrageous; they've even fabricated all sorts of ridiculous scenarios."

The rumor originated from multiple posts on social media platforms, claiming that the AI ​​chief of a major company was preparing to resign because the company's green business unit had taken away computing resources, and specifically alluding to Yao Shunyu, the head of Tencent's AI division.

The Jensen Huang Foundation donated $108 million in AI computing power.

According to Reuters, the foundation established by Nvidia CEO Jensen Huang and his wife Lori purchased AI computing power from cloud computing company CoreWeave and donated it to several universities and other non-profit research institutions for scientific and artificial intelligence research.

According to a document submitted yesterday, the donated computing power is currently valued at $108.3 million. Nvidia also stated that it plans to provide free engineering services to some of the recipient organizations.

CoreWeave is a cloud computing company focused on AI applications, and the GPUs it provides to its customers are designed and manufactured by Nvidia. In January of this year, Nvidia invested $2 billion in CoreWeave, briefly becoming the company's second-largest shareholder.

InStone CEO responds to Luna's high price: 5299 yuan is the US price.

Yesterday, InStone CEO Liu Jingkang responded to the pricing issue of its flagship dual-camera gimbal camera, Luna Ultra, stating that the rumored price of 5299 yuan for the single unit is true, "but that's the price in the United States. I wouldn't dare to sell it at that price in China, no matter how bold I am."

Previously, the blogger "Digital Chat Station" revealed that the machine was expected to be priced at 5,299 yuan and the all-in-one package was expected to be priced at 6,499 yuan, which immediately sparked doubts online about the "high price".

HarmonyOS is in talks to develop Maserati's new energy vehicle.

According to CloudInsight, Huawei HarmonyOS, JAC Motors, Stellantis Group and its Maserati brand are in talks to jointly develop new energy vehicles under the Maserati brand.

According to the plan, the four-party cooperation model is highly similar to the "Five Realms" of HarmonyOS: Huawei leads the product definition and provides core technologies, JAC jointly develops and is responsible for production and manufacturing, and Maserati provides styling design and brand endorsement.

The collaborative models will be divided into domestic and overseas versions. The domestic version will belong to the Zunjie brand, while the overseas version will be branded as Maserati. The first model is currently in the styling design stage and is planned to be mass-produced in the second half of next year.

BMW Brilliance's annual water conservation data released: Reclaimed water replaces tap water, reducing total consumption for three consecutive years.

BMW Brilliance released its annual water conservation data yesterday, showing that the water consumption per unit at its Shenyang production base will be reduced to 1.4 cubic meters per unit in 2025, a year-on-year decrease of 10.8%, marking the eighth consecutive year of steady decline.

In terms of total water consumption, the total water intake of BMW Brilliance's Shenyang plant decreased to 752,000 cubic meters throughout the year, a reduction of over 40% compared to 2023, marking the third consecutive year of decline. The use of reclaimed water exceeded 680,000 cubic meters, an increase of 71.9% year-on-year. Officials stated that the water saved could meet the annual domestic water needs of nearly 9,300 households in Shenyang.

BMW Brilliance's Dadong Plant North Plant has achieved zero tap water consumption in all production processes, replacing it entirely with reclaimed water; the cooling towers in the painting workshops of the Tiexi Plant and Lida Plant have also expanded to use municipal reclaimed water, and the irrigation of the plant's green areas had already achieved 100% use of municipal reclaimed water by 2024.

Intel becomes McLaren's official computing partner

Intel announced yesterday a multi-year strategic partnership with the McLaren Racing team, officially becoming the official computing partner of the McLaren MasterCard Formula 1 team, Arrow McLaren Indy team, and McLaren F1 Sim Team.

Under the cooperation agreement, Intel Xeon processors and Core Ultra processors will provide computing power support for McLaren's core performance workloads, covering areas such as computational fluid dynamics, aerodynamic analysis, vehicle dynamics simulation, and race strategy analysis.

Intel's technology will also be applied to scenarios such as edge computing, digital twins, and AI predictive modeling.

Honda suffered its first annual loss since its listing.

Yesterday, Honda released its consolidated financial results for the fiscal year ending March 2026. Revenue for the period was 21,796.6 billion yen, a slight increase of 0.5% year-on-year; net loss attributable to shareholders was 423.9 billion yen , compared to a net profit attributable to shareholders of 835.8 billion yen in the same period of the previous fiscal year, turning from profit to loss.

The main reason for the loss was a significant adjustment to its electrification strategy. During the fiscal year, Honda decided to cancel the development and launch plans for several electric vehicle models in North America, resulting in impairment losses, disposal losses, and contract compensation totaling approximately ¥1.5777 trillion, all of which were recorded in the automotive business. The motorcycle business was the only segment to achieve positive profitability this fiscal year.

Looking ahead to next year, the company expects revenue of 23.15 trillion yen and net profit attributable to shareholders to recover to 260 billion yen.

LinkedIn announced layoffs affecting its engineering, product, and marketing departments.

According to Bloomberg, LinkedIn, the professional networking platform, announced layoffs yesterday, affecting multiple functional departments including engineering, product, and marketing.

In a memo to employees, LinkedIn CEO Daniel Shapero stated that the company needs to provide greater value to users and achieve higher profitability, and this organizational restructuring aims to position the company for future development.

A LinkedIn spokesperson later confirmed the changes, stating that the adjustments were part of "routine business planning," but did not disclose the specific number of layoffs.

Anthropic CFO: AI has written over 90% of the code; the human role is shifting to "supervisor".

According to a Business Insider report yesterday, Anthropic CFO Krishna Rao revealed on the podcast "Invest Like the Best" that Claude currently handles over 90% of the company's coding work and is deeply involved in financial processes, driving a shift in white-collar work from "execution" to "supervision".

Rao also revealed that the company has been using Claude to generate financial statements, and the monthly financial review process is now 90% to 95% complete before manual intervention. He mentioned that the generation time for some internal reports has been reduced from several hours to 30 minutes.

Regarding staffing, Rao described the introduction of AI as a productivity "accelerator." He stated that because of AI's involvement, employees were able to shift more of their energy from information gathering to decision-making, which actually led Anthropic to expand its hiring scale.

We hired more people because there's no job that can't be done.

Rao also described a new work model: employees are increasingly taking on the role of "managers" of AI systems, with teams deploying multiple batches of AI agents to process projects in parallel. He summarized this trend as "everyone becoming a manager to some extent," and believes that the resulting productivity potential "is only just beginning to emerge."

He also emphasized that Anthropic values ​​"talent density" rather than "talent scale," believing that a high density of top AI research talent combined with the strongest models is the truly competitive combination.

Reports indicate that internal testing of GPT-5.6 has begun, and the flagship model's "ultra-fast mode" will be launched this week.

Renowned leaker Leo revealed yesterday that GPT-5.6 development is in full swing and is expected to be officially released as early as next month. Developers discovered two internal codenames—ember-alpha and beacon-alpha—as well as a routing mapping record in OpenAI's internal Codex logs.

The logs show that the vast majority of model calls still point to GPT-5.5, but one entry explicitly points to GPT-5.6, suggesting that the Codex environment may be testing a new model.

In addition, the whistleblower Chetaslua claimed that OpenAI plans to launch a "super-fast mode" for its flagship model this Thursday, which could improve response speed by 2 to 3 times, primarily targeting latency-sensitive tasks.

Honor 600 series scheduled for release on May 25th

Honor announced yesterday that its Honor 600 series will be officially launched on May 25th at 19:00.

Official teaser information shows that the Honor 600 series focuses on "4K flash mirrorless Live" imaging, equipped with a 200-megapixel ultra-clear large-sensor main camera, and built-in 8600mAh Qinghai Lake large battery.

On the eve of Google I/O, Gemini 3.2 Flash and several "Agent" models were leaked ahead of time.

According to Sources, Google plans to officially release its next-generation lightweight model, Gemini 3.2 Flash, at the Google I/O conference on May 20. Its overall performance is roughly equivalent to that of GPT-5.5.

Abacus.AI CEO Bindu Reddy revealed that the Gemini 3.2 Flash reportedly achieves approximately 92% of the performance of GPT-5.5 in coding and inference tasks, while its inference cost is only one-fifteenth to one-twentieth of the latter, with most query latency below 200 milliseconds.

According to leaker Leo (@synthwavedd), Google DeepMind apparently accidentally pushed a batch of unreleased models to its production API, including:

The Gemini 3 Flash, Gemini 3 Pro Image, Gemini 3.1 Flash Image, Gemini 3.1 Flash Lite, Gemini 3.1 Flash TTS, Gemini 3.1 Pro, and the music generation models Lyria 3 Clip and Lyria 3 Pro, as well as a video generation model codenamed "Omni".

It is worth noting that all of the above models have variant versions with the suffix "agent".

Tencent open-sources its AI Agent memory system, reducing token consumption by up to 61%.

Tencent Cloud's database team yesterday officially open-sourced TencentDB Agent Memory, a local-first memory engine for AI agents, supporting integration with OpenClaw and Hermes Agent. The core design of this system lies in splitting the memory into two independent structures:

• The long-term memory portion is divided into L0 original dialogue, L1 atomic facts, L2 scene blocks, and L3 user profile, which are then progressively accumulated layer by layer.
• Short-term task memory externalizes the lengthy tool log to a refs file, writes a summary of steps to jsonl, and preserves the task structure and node indexes using a Mermaid canvas.

Official benchmark tests show that after integrating OpenClaw, the token consumption for WideSearch tasks decreased from 221.31M to 85.64M, a reduction of 61.38%, while the task pass rate increased by 51.52%. On the long-term memory evaluation set PersonaMem, the accuracy improved from 48% to 76%.

GitHub: github.com/Tencent/TencentDB-Agent-Memory

lululemon launches lightweight sun protection jacket series

lululemon recently released a lightweight sun-protective jacket series, featuring UPF 40+ sun protection, which can block more than 97.5% of ultraviolet rays, while also being lightweight and breathable.

• For women, the self-packing sun protection jacket is made of Glyde fabric, which has triple performance of UV protection, rainproof and windproof. The Goal Smasher sports sun protection jacket is designed for running, with elastic splicing structure and partial opening design, emphasizing a cool experience when running flexibly.
• For men's styles, Stash and Dash The self-packing sun protection jacket features multiple pockets and can also be packed away as a portable pouch; another lightweight version of the sun protection jacket also uses Glyde fabric, and the back ventilation design provides stable protection for changeable weather.

Pop Mart denies entering the traditional home appliance market; LABUBU refrigerators reported first-quarter revenue of approximately 6 million yuan.

Recently, Pop Mart's Chief Operating Officer, Si De, stated in the Q1 2026 business call that the company has no plans to expand into traditional home appliances. In Q1 of this year, the sales revenue of LABUBU refrigerators was approximately RMB 6 million, accounting for a very small percentage of the total revenue.

According to Si De, Pop Mart's home appliance business is currently in a very early stage. In the future, it will combine small home appliances with lifestyle, IP and trendy toy culture to create an important supplementary category.

Sidde stated that the company will not accelerate the development of its home appliance business, but hopes to refine its products and supply chain in a more stable way, "taking more time and patience to continuously improve this category."

CASETiFY collaborates with a Japanese illustrator to launch the "nokonoko uchinoko" series.

CASETiFY announced yesterday that it has launched a new collaboration series, "nokonoko uchinoko," with Japanese illustrator Saori Iijima. The series is inspired by the imagery of soft, baked bread and uses gentle brushstrokes to depict everyday emotions.

Saori Iijima is an independent illustrator from Japan, known for her delicate and soft art style. She has previously collaborated with several brands. CASETiFY regularly collaborates with global artists and IPs, and this artist series is one of their regular collaborative projects.

Netflix's "East of Eden" trailer released, set to premiere this fall.

According to Douban Movie, Netflix released a Chinese-subtitled trailer for the series "East of Eden" yesterday, and the official announcement confirmed that the series will be released this fall, consisting of 7 episodes.

The series is adapted from John Steinbeck’s classic novel of the same name published in 1952. Set in the Salinas River Valley of California, it tells the story of the intertwined fates of the Trask and Hamilton families across generations, spanning from the American Civil War to the end of World War I.

The Big Bang Theory spin-off series "Stuart Failed to Save the Universe" has released its release date and trailer.

The trailer for "Stuart Failed to Save the Universe," a spin-off of "The Big Bang Theory," has been released, announcing its release date.

The story centers on Stuart Bloom, the owner of a comic book store.

He accidentally damages a device built by Sheldon and Leonard, triggering a multiverse-level apocalyptic crisis. He is then forced to work with his girlfriend Denise, geologist Bert, and quantum physicist Barry Kripke to repair reality, encountering characters from various parallel universe versions of The Big Bang along the way.

K-POP: Witcher Girls Global Tour Officially Announced

Netflix and AEG Presents recently announced that they will adapt the animated film "K-POP: The Witcher" into a world tour, stating that they will bring elements of the film to the stage in a "dynamic and memorable way".

K-POP: The Witcher won two awards at this year's Academy Awards: Best Animated Feature and Best Original Song. The winning track "Golden" became the first K-pop song in history to win the Academy Award for Best Original Song.

The film's soundtrack also won a Grammy Award and topped the Billboard Hot 100 chart, becoming the best-selling soundtrack worldwide last year.

#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.

Many people are probably familiar with this scenario:

When chatting about work, discussing travel plans, or even just having casual conversations with friends on WeChat, you might suddenly want AI to help summarize the key points and make a plan. However, you end up having to copy and paste the chat history piece by piece into the AI's chat window.

For example, I took a short nap at noon, and when I woke up I found hundreds of messages in the group chat. If I wanted to get the latest gossip, I had to scroll through them one by one…

Wouldn't it be wonderful if we could forward chat history to AI with just one click?

This feature has finally arrived, but unfortunately it's not in WeChat:

Now, after updating WeChat and Yuanbao to the latest version, you can choose to send chat history directly to Yuanbao for summarization.

Although both belong to Tencent's ecosystem, there has always been a subtle sense of distance between WeChat and Yuanbao—finally, this collaboration between the two has brought them closer together.

iFanr also tried out this feature as soon as possible to see how well it works.

The product is great, but the effect is just so-so.

To forward WeChat chat history to Yuanbao, the method is very simple. First, select the chat history you want to forward, then select "Forward to other applications" to send it directly to Yuanbao. You can add some prompts.

The whole process is somewhat like having AI help organize the scattered content from discussions and chats. Therefore, the most suitable scenario for this function is the discussion and arrangement information in various work groups.

One advantage of directly processing chat logs is that Yuanbao can accurately match discussions with speakers, making it ideal for processing meeting minutes and work arrangements.

Travel planning is also a high-frequency scenario, because itineraries are usually discussed and decided through conversations. At this time, you can directly hand it over to Yuanbao to handle, and it can also help recommend travel destinations.

There's another slightly unorthodox use: If you're too busy with work to check group chats and are worried about missing out on trending topics or discussions, you can select all the chat history and let Yuanbao help you check the group chats and summarize the information.

Or if you've been discussing for a long time and still don't know what to eat for your get-together, you can select the discussion process and let Yuanbao help with recommendations.

We tried many scenarios, and in most cases, it met our expectations. However, overall, there are still some details of the experience that are not perfect.

Currently, Yuanbao is best at handling text-based chat logs. However, if the chat logs contain files, especially multiple files, the results are less reliable, and the files often cannot be read.

Even if the task can be successfully read, it doesn't necessarily mean that the task can be completed according to the prompts: sometimes even sorting invoices might not be filled out correctly, and asking him to create a mind map based on a document often fails.

You know, organizing invoices from chat logs is a really common scenario!

I can clearly feel a gap: sending chat history to AI is a valuable scenario, but the result of Yuanbao's delivery always makes me slightly disappointed, and I can't help but wonder if it would be better if it could be forwarded to Doubao or ChatGPT?

The deeper issue is that the chat log forwarding interaction always gives me the feeling of "taking a detour".

Prior to this, Yuanbao already existed as a contact in WeChat, but processing chat history still required a whole set of plug-in processes that involved "jumping" multiple times.

Perhaps due to WeChat's strict privacy protection rules, Yuanbao can only handle chat history in the form of "temporary chat" and will not leave any history records. Once closed, it cannot be retrieved.

For WeChat, such a feature seems "optional".

If WeChat wants to integrate AI, they might have better ways to achieve this, such as creating an "AI group member" that can answer questions or process chat history; and an agent that is always present in the chat list that can directly engage in conversation and receive chat history.

But this is not so easy, for reasons that are already well-known: WeChat needs to be simple and restrained, because every small change will affect a billion users.

On the same day that the Yuanbao feature was launched, Tencent released its first-quarter financial report for this year. At the shareholders' meeting, Ma Huateng said that developing agents in WeChat requires "longer-term considerations. Everyone needs to be patient. This is not something that can be rushed."

Tencent's AI trump card is WeChat.

Ma Huateng also mentioned Tencent's AI strategy:

A year ago, we thought we were on a boat, but then we found out it was leaking. Now we feel like we can stand on it but can't sit down. We still hope the boat can go faster.

According to the financial report, AI has become a major investment for Tencent. In the first quarter, R&D investment reached 22.54 billion yuan, up 19% year-on-year; capital expenditure was 31.94 billion yuan, mainly used for the expansion of cloud infrastructure such as computing power and storage.

This afternoon, some self-media outlets reported that "Tencent's AI chief executive officer is about to leave the company," but Tencent quickly denied the rumor this afternoon, sending a clear signal to the outside world: Tencent AI is not to be underestimated.

According to the latest AICPB rankings, Tencent Yuanbao ranks tenth among domestic AI applications with 17.73 million visits, far behind Doubao (163 million) and Qianwen (44.24 million).

In the field of AI, Tencent and Apple are in a similar situation: they have top-tier product entry points and user scale, but their AI models themselves cannot yet support such expectations.

Generative AI has been a hot topic for three years. Since the end of last year, the battle between AI applications has entered a "scenario-based" competition. Not only have new forms like "Lobster" OpenClaw emerged, but the traditional ChatBot interaction mode has also integrated life services. Companies like Alibaba's Qianwen and ByteDance's Doubao are actively combining their advantages in e-commerce and other ecosystem services with their own AI assistants.

This should have been Tencent's comfort zone, since the most powerful and versatile lifestyle application in China, WeChat, is in Tencent's hands.

However, it's clear that WeChat is taking a very cautious approach to "AI integration." Tencent cannot afford to disrupt the experience of WeChat's massive user base due to immature technology or features.

However, for Tencent, given its lagging AI performance compared to its competitors, the value of WeChat as a trump card remains extremely important. Therefore, while WeChat and Yuanbao haven't fully embraced each other, they've begun cautiously testing the waters.

The reasons for this, besides WeChat's relatively closed ecosystem logic, include two unanswered questions: How exactly should WeChat be "AI-ized"? And what role should Yuanbao play in this process?

The Information reported that even though Tencent is investing heavily in accelerating the development of its own hybrid model, WeChat still believes that the current level of hybrid is not outstanding enough, and has even begun testing third-party AI models for WeChat Agent, which brings challenges in terms of privacy and integration.

According to sources, WeChat has been developing its own AI Agent model technology and plans to release a WeChat Smart Agent in the middle of the year. This agent can be integrated with the entire mini-program ecosystem to provide services such as ride-hailing and shopping, and will be directly integrated into WeChat through contacts and chat. However, as mentioned earlier, this matter is too significant and has too wide an impact, so it cannot be rushed.

Once this product is launched, it will likely become the AI ​​gateway with the most users nationwide.

WeChat needs AI, but it may not need "Yuanbao" (a virtual currency), which is Tencent's most delicate problem right now.

#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.

If you want to destroy a world-famous painting on the internet today, the fastest way is not to physically destroy it, but simply to label it: "This was painted by AI."

Recently, user X, @SHL0MS, conducted a rather quirky social experiment. He uploaded an original painting of "Water Lilies" by French Impressionist master Claude Monet, deliberately tagging it with the platform's "Made with AI" label, and adding the following text:

"I just generated an image in the style of Monet using AI. Please describe in as much detail as possible why this painting is inferior to Monet's original works?"

Faced with the blogger's "entrapment" tactics, online art purists instantly sprang into action. Led by the blogger's deliberate emphasis on "details," the comment section was quickly flooded with lengthy, professional analyses.

Rather than being a hilarious yet deeply moving farce, this farce also reveals a cruel truth: in the AI ​​era, it's not just AI that's starting to experience serious hallucinations, but also humans themselves.

After a genuine Monet painting was disguised as an AI artwork, art experts across the internet were completely fooled.

As of press time, the original tweet had attracted 4 million online views and began to circulate on major social media platforms.

There's nothing suspicious about the painting itself. It depicts Monet's iconic water lily pond, with dappled sunlight shimmering on the surface, the edges of the brushstrokes blending into varying shades of green, and the whole painting enveloped in a soft, misty atmosphere. If you were to stumble upon it in the Louvre, you'd most likely just exclaim, "Monet is indeed Monet," and then pull out your phone to take a picture.

But now, it hangs there with the label "Made with AI".

As expected, netizens did not disappoint the blogger's expectations. To prove they possessed an artistic aesthetic that surpassed AI, countless amateur Leeuwenhoek-like figures instantly flooded the comments section. Everyone took up microscopes and began dissecting this "AI garbage" in every detail: some confidently pointed out the fatal flaws in the composition: "This is all a mess, with absolutely no sense of space."

Some astute observers seized upon the flaw in the colors: "The colors are reversed; blue water lilies are displayed on the green water surface." Others criticized the roughness of the details: "The lack of texture, sharp edges, wrinkles, gaps, creases, bevels, and three-dimensionality are typical characteristics of plastic artworks."

One critic harshly commented: "The depth of field and color choices in the image are completely inconsistent. The reflections of the trees and water lilies are mixed together, completely disregarding spatial depth and contrast. The mixed parts of water lilies and algae in the background are blurry, just like most AI works."

One person gave what seemed to be the most incisive assessment: "You can feel from the bottom of your heart that this painting lacks real passion; it's a soulless cyber junk." A top-tier "Jiahao" even wrote a lengthy analysis of hundreds of words.

Looking at these earnest and logically consistent criticisms, you almost want to applaud humanity's keen perception of art, until the truth is revealed: this painting, which was ridiculed by the entire internet, is actually a masterpiece painted by the Impressionist master Claude Monet.

Anyone with even a basic understanding of art history should know that Monet was diagnosed with severe cataracts in both eyes in 1912. As his vision deteriorated rapidly, the world he saw lost its cool tones and became blurry and mottled.

In his later years, he created 250 oil paintings in the "Water Lilies" series, which were originally intended to record the melting world of light and shadow in his eyes with extremely abstract, unrestrained, and even "imprecise" brushstrokes.

If someone had judged him by standards like "inaccurate brushstrokes" or "lack of passion," Monet would probably have only been able to manage a wry smile.

Labels can be used to define justice; even in the AI ​​era, there are "cheap wine experiments."

Why do netizens criticize Monet's original works as if they were created by AI?

On Reddit's Singularity subreddit, this incident sparked thousands of heated discussions. One user astutely pointed out the underlying psychological mechanism: this is simply another classic experiment of cognitive bias.

This brings us to a famous psychology experiment that took place at the University of Bordeaux in France in 2001.

At the time, researcher Frédéric Brochet invited 54 senior wine tasting experts. He dyed a bottle of inexpensive white wine red with tasteless food coloring and then asked the experts to taste it.

An interesting experiment was conducted at the University of Bordeaux in France in 2001. Frédéric Brochet, a psychology PhD and professor who also makes his own wine, gave 54 wine experts two glasses of wine and asked them to compare the taste.

The result was astonishing: these experts, who were usually extremely picky about flavors, picked up their glasses and wrote down "rich berry aromas," "nutty finish," and "heavy tannins"—all characteristics of red wine. Not one of them realized that it was actually a white wine.

If you tell them it's cheap wine, they'll taste its sourness; if you tell them it's from a top-tier winery, they'll taste the patina of time.

Today, the words "AI-generated" are the label stuck on cheap liquor bottles.

When this painting is labeled as AI-generated, people subconsciously presuppose that it is cheap, mechanical, and soulless. Therefore, they are not looking at the painting with their eyes, but rather using the biases in their minds to "search" for those preconceived flaws.

The claim by netizens that "I can feel deep down that it has no soul" is nothing more than metaphysical nonsense packaged within a rationalist framework. As one Reddit user succinctly observed: "If someone tells you it's AI, it's soulless; if they tell you it's human-drawn, it's full of passion. The true quality of the art has become irrelevant in this discussion."

We must acknowledge an awkward truth: the vast majority of people—including those eloquent cyber judges online—simply lack the ability to distinguish between top-tier art and AI masterpieces. If you couldn't understand Monet before AI existed, you still won't understand him now.

A large-scale witch hunt targeting AI

The Monet incident is by no means an isolated case; it reflects an extremely dangerous and morbid trend on the internet today: "anti-AI witch hunts."

In today's era of rapid advancements in generative AI, countless real human artists are desperately trying to prove their innocence every day. Their paintings are labeled as AI and attacked by angry netizens simply because they are too realistic, the lighting is too perfect, or conversely, the fingers are drawn a little rough or the proportions are slightly off.

The most infamous tragedy on the internet is undoubtedly the Ben Moran incident at the end of 2022.

This human digital artist posted an illustration titled "A Muse in Warzone" on Reddit, a forum with 22 million users. The moderator permanently banned him, citing that "this is an AI-generated image."

When Ben Moran tearfully submitted his line art, layers, and a screen recording of his dozens of hours of drawing process in an attempt to prove his innocence, the moderator gave an extremely arrogant reply: "I don't believe you. Even if this is really your work, its style is too much like AI; it's worthless. You'd better change your style."

Look, this is the absurd reality we live in now. Real human artists can neither paint too perfectly (that's done by AI), nor make basic mistakes (that's an illusion created by AI), and they can't even have the same painting style.

This prejudice is not only present among the general public, but is also spreading to professional fields.

Brandon Sanderson, the American fantasy novelist, once conducted a blind test experiment. Researchers placed passages he wrote by hand alongside passages generated by AI that imitated his writing style, and asked a group of professional writers and peers to distinguish between them.

As a result, even writers who make a living by writing cannot accurately distinguish which piece was written by AI.

Even more interestingly, in some unlabeled literary experiments, top literary critics, in blind tests, actually preferred AI-generated literary works, believing them to be more compelling; however, once…

Label these works as "AI-generated," and the same critics immediately change their tune, starting to nitpick about the lack of human emotional resonance.
The fact that Monet's original works were treated as AI cyber junk may seem like just an internet frenzy, but the social crisis hidden behind it is something that cannot be laughed off.

As one netizen put it: "What I fear most is not how powerful AI becomes, but that it is destroying the foundation of trust in society. In a society where no one trusts anything, we will find it impossible to move an inch."

The most terrifying future may not be the awakening of AI and the extermination of humanity, but the complete bankruptcy of social trust. When people discover that their senses are no longer reliable, when "seeing is believing" becomes history, we will completely retreat into our emotional and positional comfort zones.

If something is something I don't like or that doesn't conform to my understanding, I can rightfully accuse it of being "AI-generated"; if something is something that suits my taste, even if it's crudely generated by AI, I will praise its "realism".

We seem to be rejecting AI, yet unknowingly we are turning ourselves into machines most easily predicted and manipulated by algorithms. Humans are becoming more and more like the AI ​​we imagine ourselves to be: inputting labels, outputting preset conclusions, with the real thinking process in between being omitted.

When faced with the unknown, humanity's illusions and unfounded beliefs surpass those of AI.

#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.

DJI's Pocket series hadn't been updated for three years, but when it was finally updated, two units arrived—

For the past month or so, the Pocket 4 and Pocket 4P have been the most talked-about cameras. Tonight at the Cannes Film Festival, the Pocket 4P made its debut as a pocket cinema camera in the European cinema hall of fame, and its official release is not far off.

At the same time, DJI also lowered the price of the Pocket 3, thus creating a clearly defined Pocket series product line:

• Pocket 3: An entry-level handheld gimbal vlogger, priced at 2299 yuan.
• Pocket 4: A versatile handheld gimbal camera, 2999 yuan
• Pocket 4P: A semi-professional pocket cinema camera, expected to be priced around 4000 yuan.

Many readers who are holding cash and waiting to see what happens share the same question as iFanr:

Which of these three Pocket phones should I choose? Do I still need a Pocket phone if I already have an iPhone?

Based on real-world experience, ifanr will help you understand it all at once.

Pocket: Three Generations Under One Roof – Which One to Choose?

Let's talk about the Pocket 4P in detail.

Although more detailed specifications for the Pocket 4P have not yet been released, specifications are the least important topic compared to this form factor update.

From a product logic perspective, the Pocket 4 has already brought single-lens handheld gimbal cameras to a relatively mature stage. At this point, adding another camera becomes a logical step.

Most smartphones offer wide-angle views for their main camera and single-lens Pocket camera. Wide-angle lenses are good for capturing the environment and framing grand landscapes. However, when you walk into a crowded area with your device, the drawbacks of wide-angle lenses become apparent: the image becomes too cluttered.

At this point, a telephoto lens can achieve spatial compression, eliminating irrelevant background elements.

DJI cleverly positioned this extra camera at 3x, a very ingenious number. It's closer to the human eye's field of vision after focusing than the main camera. From our hands-on experience, the Pocket 4P's 3x lens also provides a relatively natural bokeh effect, making it ideal for shooting portraits or close-ups that emphasize the subject.

However, for users, the extra camera is not necessarily a positive benefit: since its inception, the core selling point of the Pocket series has been "seamless portability". Adding a lens means increasing weight, taking up internal space, and may even sacrifice some battery life.

In fact, this concern has helped DJI clarify the distinction between 4P and 4.

The Pocket 4P is designed for "creators." They know exactly what they want to shoot, and the footage will eventually be transferred to a Mac for meticulous editing in DaVinci Resolve or FCPX. For them, sacrificing a little weight is entirely worthwhile in order to achieve native focal length framing and highly consistent footage specifications within an extremely compact size.

For most ordinary people, the Pocket 4 is the lighter and more rational choice.

Although the previous generation Pocket 3 achieved the legendary feat of selling tens of millions of units, it was not a flawless camera – its USB 2.0 capability resulted in slow data transfer, low pixel count, and insufficient dynamic range, all of which were problems.

The Pocket 4 offers 37 megapixels, and the higher pixel count means improved built-in lossless ISZ. On the Pocket 4, the 2x focal length, obtained through sensor cropping, is sharper and more than sufficient for everyday recording.

In addition, it has 107GB of built-in storage, resulting in better dynamic range. Issues such as forgetting to bring a memory card, insufficient transfer speed, and inadequate dynamic range are all resolved at once.

It can be said that the Pocket 4 is the complete version of the current single-lens handheld gimbal camera.

At this point, the differences between the two Pocket 4 series units are quite clear:

If you prioritize productivity, emphasize the image quality and bokeh brought by the native optical focal length, and need a 3x lens to obtain a wider range of field of view, then the Pocket 4P is undoubtedly the better answer;

If you simply want a camera that can easily fit in your pocket, has excellent image stabilization, and can shoot videos and Live Photos independently of your phone, then the Pocket 4, with its higher resolution and user-friendly 2× ISZ, is more than enough.

Finally, there's Pocket 3.

With Pocket 4 addressing many pain points, is there still a reason for Pocket 3 to exist?

Of course, the biggest contribution of new devices is often to bring down the price of older flagship models.

Between the releases of the Pocket 4 series, we observed that the Pocket 3 has undergone an official price reduction, and with various subsidies on some platforms, it has dropped to 2099.

It's not impossible that in the future we might be able to buy a Pocket 3 for a price starting with 1.

For those with limited budgets or beginners just starting out with the idea of ​​making videos, this is the most cost-effective "gatekeeper".

Finally, according to the speculation of the iFanr editorial team, the Pocket 4P will most likely be priced at around 4,500 yuan, but it is also possible that DJI will roll out the price and set it at 3,999 yuan.

At this point, if we arrange the Pocket series according to price, we will find a clever design—

From the Pocket 3 in the 2000 yuan range to the Pocket 4 in the 3000 yuan range, and finally to the Pocket 4P in the 4000 yuan range, DJI is bringing the same combination of Pro, Air, and Mini features from its drones to its pocket cameras.

This lineup has already proven its success in the drone market. Now, Pocket will also be forming a similar lineup.

With an iPhone, do I still need a Pocket?

After publishing my hands-on article about the Pocket 4P, I received questions from more than one friend:

I already have an iPhone, do I still need to buy a Pocket?

Having clarified the internal positioning of the Pocket family, this is a real question we must face: since everyone already has a powerful iPhone, why spend money to buy a single-function gimbal camera?

This is actually a philosophical question about "device attributes".

After multiple iterations, the iPhone's foundation in dynamic imaging is undeniably strong, featuring smooth zoom with triple cameras, highly consistent colors, excellent image stabilization, and even the Apple Log has been upgraded to its second generation.

The problem is that mobile phones are not products specifically designed for photography.

Besides taking photos, iPhones also need to handle complex daily tasks, requiring constant distraction. Work messages and voice calls may interrupt your shooting at any time. You need to consider whether the phone has enough battery left to keep you connected to the world, and you also need to be careful that shooting doesn't fill up the already precious storage space…

Pocket, on the other hand, is a camera designed entirely for shooting: it can't make phone calls, reply to WeChat messages, or browse social media. Nothing unrelated to shooting will appear on your Pocket.

Not to mention, the Pocket also has the inherent advantages of a mechanical gimbal, which makes it structurally more stable than the iPhone's image stabilization, more comfortable to hold, produces more natural physical bokeh, and offers a wider range of video specifications than the iPhone.

So, if I already have a primary iPhone, should I buy a Pocket now?

If I need one right now, which one would be the best fit for me?

You can check the following questions to perform a pre-purchase assessment:

1. In the past three months, have you ever had your filming interrupted by a sudden phone call or WeChat message?
2. When shooting videos, are you restricted by insufficient storage space on your phone?
3. In the sweltering summer, have you ever experienced your phone overheating, dropping frames, or even crashing while taking photos or recording videos?

If your answer to these questions is yes, then purchasing either the Pocket 3 or the Pocket 4 is a good choice.

Specifically, if you need to take high-definition live photos, or if you prefer to use your phone's 2x camera, then choose the Pocket 4; otherwise, the Pocket 3 is sufficient.

On top of that, if you want to use the Pocket for commercial shoots, concert shoots, or serious creative work, then the Pocket 4P is the perfect all-in-one choice.

Of course, if your answers to the first few questions are all negative, then you should hold onto your wallet tightly.

The iPhone's image stabilization and image quality are more than enough to meet your needs. You can shoot and edit instantly, share directly via AirDrop, and the smooth ecosystem experience remains irreplaceable by any third-party device.

From an industry perspective, ifanr suggests you wait a bit longer.

Currently, DJI isn't the only player in the pocket gimbal camera market—Instal Luna, which is also "half-hidden" like the Pocket 4P, is also eager to enter the fray and grab a share of the market.

Because products like the Pocket 4P have seriously threatened the security of mobile phone manufacturers.

Many people are willing to buy flagship imaging phones costing 6,000 or even 8,000 yuan, largely for their telephoto lenses. Now, for a lower price, the Pocket 4P offers a comparable telephoto lens, better gimbal stabilization, and a more professional overall solution.

No mobile phone manufacturer can turn a blind eye to this.

According to iFanr, vivo and OPPO have already put handheld gimbal camera product planning on the agenda, while Xiaomi and Huawei are also continuously paying attention to this market. From the second half of 2026 to the first half of 2027, we can expect to see even more handheld gimbal cameras emerge.

Now that there are more players in the market, we as consumers can actually be more composed. You can hold onto your money and wait a bit longer.

We believe that even more exciting products will emerge in this young category.

#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.