Fitbit Air makes judgment the interface

ChatGPT, Claude, Codex, and Workbody are turning product management's visible artifacts into drafts, which makes the harder interface the judgment behind them. woshipm's author Liu Xing Liao Chanpin argues that AI tools have moved into daily product work, taking on PRDs, prototypes, data analysis, process sorting, and code writing. A requirement document that once took half a day can now be generated in a few minutes, in Liu's account. That changes what competence looks like.

The old proof of work was often a document, a chart, or a prototype. If tools can produce those on demand, the valuable question shifts upstream: what should be built, when should it interrupt the user, and where could confident automation cause harm? Liu's sharper claim is that AI lowers the threshold for doing product work while raising the threshold for the product manager profession. The job moves from producing artifacts to defining context, boundaries, and trust.

AI turns product outputs into drafts, not decisions

AI is making the old evidence of product work easier to manufacture. Woshipm's author Liu Xing Liao Chanpin argues that tools such as ChatGPT, Claude, Codex, and Workbody have entered product managers' daily routines, taking over execution-level tasks that once signaled competence: PRDs, prototypes, data analysis, process sorting, and code writing.

The change is clearest in time compression. Liu says requirement documents that once took half a day can now be generated in a few minutes. Chang Sixing, who ran AI through every stage of his own product manager work for half a year, gives the same pattern at task level: a weekly report fell from 20 minutes to 3 minutes, and a competitor feature list fell from half a day to 30 minutes.

That does not make the output decisive.

Chang says Feishu Minutes can produce meeting notes with sections for decisions, to-do items, and risks, with accuracy above 95%. He also says AI can draft the "background and goals" section of a PRD from meeting recordings, requirement emails, and related data, leaving him to make minor edits. User stories changed too: one module used to take 2 hours to split, while AI now produces a first draft in 15 minutes and he spends 45 minutes adjusting it.

Even SQL is becoming a draft surface. Chang says complex SQL used to take 30 minutes to write; now AI generates it from natural language, and he finishes in 5 minutes after changing table names and fields. The artifact arrives sooner, but the remaining work has shifted toward checking assumptions, fit, and consequence.

Chang's own accounting is the useful warning. His time spent writing documents fell from 40% to 15%, and the freed 25% now goes into validating AI outputs and making deeper decisions. He argues the outputs are above average but below excellent: complete in structure, thin in insight and sharp judgment. Liu's broader point lands there too. AI lowers the threshold for product work, but raises the threshold for the profession.

The PM skill floor rises when tools raise the output floor

The clean answer is not to ignore tools. If AI can now produce acceptable execution work, tool use becomes the floor, not the edge. Chang Sixing argues in woshipm that a well-prompted AI already beats a junior product manager with 3 months of experience on execution-level tasks. He also says that since 2024, hiring for product manager assistants and junior product specialists has clearly declined because AI performs more than 50% of the work those roles used to carry.

That does not make tool-chasing a career plan.

The harder lesson is that juniors must stop treating AI output as proof of product ability. Woshipm's Liu Xing Liao Chanpin argues that AI replaces execution rather than judgment, lowering the threshold of work while raising the threshold of the profession. His point comes from implementation, not abstraction: he says he has explored AI in government services over the past two years, with his team trying mainstream large-model capabilities across intelligent Q&A, intelligent guidance, service-handling assistance, and AI Agent applications.

In those projects, Liu argues, the model is only part of the solution. The business is the difficult part. No model can decide where AI should intervene, when automatic processing is appropriate, when human review is necessary, how the knowledge base should be maintained, or whether the user needs a chatbot instead of an assistant that completes tasks.

That reframes the junior practice path. First, study scenarios until the process has edges: user intent, exception cases, handoff points, and failure costs. Then build workflows with AI as the execution layer. Finally, review the output as a product decision, not a writing sample.

Chang says his own information-input workflow has an AI replacement rate of about 70%, and he recorded a course with 55 lessons and 280 subsections covering AI toolchain practice, problem-definition methodology, and cross-departmental promotion. The ordering matters. Tools are learned, but value comes from defining the problem, defending the process, and knowing when the machine should stop.

Active assistance has to pass a scenario threshold

Active assistance should not pass review because it can produce a plausible notification, summary, or prompt. It has to pass a scenario threshold. Woshipm's author frames that threshold with demand-review questions: whether the function works only in a specific situation, whether the information must attach to the user's current action, whether AI should wait for commands or observe, and whether the beneficiary is also the person affected.

The diary push-notification case shows why this matters. According to woshipm, a customized diary push feature increased push open rates and daily active users after launch, but after two weeks it did not increase the number of people writing diaries. The feature brought users back. It did not make them want to write.

That is the difference between activation and fit.

A review process for AI features should start by asking what moment the feature claims. AirPods, woshipm's author argues, worked because they occupied walking, commuting, or both-hands-occupied situations where audio control mattered. For glasses, the same author says the physical advantage over phones appears only when the user's hands are occupied. A feature that sounds useful at a desk may lose its reason to exist when the phone is available.

The second check is attachment. Woshipm identifies the question as whether information must be connected to the action happening now. Many AI glasses, the author says, have no display and rely on voice to read information aloud. If the information can wait, or if hearing it interrupts the action, the AI has created noise rather than assistance.

The third and fourth checks are more sensitive. Most current AI glasses, woshipm argues, still follow a user-asks, device-answers model, while Meta is pushing toward glasses that silently remember what they see. That changes the affected party. Photography, video recording, and real-time understanding may feel powerful to the wearer, but threatening to nearby people. Review cannot stop at usefulness for the owner.

From visible product outputs to context, boundaries, and trust

Screenless devices do not remove the interface problem

Removing the screen changes the interface problem; it does not make the problem disappear. According to geekpark, Fitbit Air has no screen, no watch face, and no buttons. Users can double-tap the body to make a side LED show approximate battery level, while the main body weighs 5.2 grams. GeekPark's Zhang Yongyi argues that it is not meant to be a watch, and that the point of the screenless wristband is to remove enough interaction that users can forget it exists.

That only works if something else explains the product.

Fitbit Air removes time interfaces, map interfaces, notification interfaces, and exercise interfaces, according to geekpark. Zhang's stronger point is that the device is a sensor terminal for continuously feeding data to AI, not a smartwatch with its screen taken away.

The risk sits there: after the screen disappears, the model speaking for the device has to be reliable enough to carry the user's trust.

Apple shows why this is hard. ifanr cites Mark Gurman describing the current Apple Watch as a mini iPhone worn on the wrist, a useful phrase because the Watch began with visible interaction as its premise. In 2015, The New Yorker recorded that Apple wanted the Apple Watch to help people check notifications and messages on the wrist. Its square face suited lists such as contacts and schedules, and the Digital Crown appeared because two-finger zooming on a small screen was awkward.

Oura and Whoop took the opposite route. ifanr says both removed screens and gave up complex interaction. They focused on wearability and battery life, but they did not leave users with raw signals. Whoop packages indicators into a daily recovery score. Oura uses scores across three dimensions to describe readiness for the day.

The same tradeoff appears in sspai's iPhone Air account. The author moved from iPhone 15 Pro after 7 years of Pro-series iPhones, with iOS 26's Liquid Glass in the background.

The device became a content consumption device. It became a social connection device. It became an AI control device, while weak battery life and heating still mattered. Thin hardware can fade back, but judgment cannot.

Use AI for outputs; reserve human judgment for context, intervention, and harm

• You need meeting notes, weekly reports, competitor tables, PRD drafts, user-story splits, or SQL first drafts. Use AI aggressively for the first pass. The sources describe Feishu Minutes producing meeting notes with decisions, to-do items, and risks at accuracy above 95%; weekly reports falling from 20 minutes to 3 minutes, or being drafted in 30 seconds; competitor feature lists falling from half a day to 30 minutes; user-story splitting falling from 2 hours to a 15-minute first draft plus 45 minutes of human adjustment; and SQL falling from 30 minutes to 5 minutes after edits. Treat the output as structured material to inspect, not as final judgment.
• The product question is where AI should intervene, when automatic processing is appropriate, when human review is necessary, how to maintain accurate knowledge, or whether the user needs a chatbot or an assistant that completes tasks. Do not delegate the decision to the model. The woshipm source argues that in real AI implementation projects, the model is only part of the solution and the difficult part is the business itself; no model can directly answer those boundary and responsibility questions. Use AI as input, but make the intervention rules, review points, and operating process explicit yourself.
• A feature improves engagement metrics but does not change the underlying user behavior you care about. Do not mistake artifact success for product success. The diary push-notification case increased push open rates and daily active users after launch, but after two weeks it did not increase the number of people writing diaries. The source's lesson is that the feature brought users back but did not make them want to write.
• You are evaluating AI glasses, screenless wearables, or other ambient devices. Judge them by scenario thresholds, not feature count. The woshipm source argues that the gap between dispensable and indispensable products is a scenario threshold: whether the function only works in a specific scenario, whether information must attach to the user's current action, whether AI waits for commands or observes on the user's behalf, and whether the beneficiary is also the person affected. This is especially important for glasses, where photography, video recording, and real-time understanding can feel like a superpower to the wearer but a threat to nearby people.
• You are designing health wearables or AI health coaching. Prioritize continuous data, unobtrusive wear, and model reliability over screens and complex interaction. Fitbit Air has no screen, no watch face, no buttons, no independent GPS, and removes time, map, notification, and exercise interfaces; GeekPark frames it as a sensor terminal that continuously provides data to AI rather than a traditional smartwatch with the screen removed. But use caution: the same source warns that AI coach features can introduce hallucination, forgetting, misunderstanding context, and confidently wrong answers into health scenarios.

Health AI exposes the cost of wrong confidence

Health is where proactive assistance becomes most tempting. Apple has pushed Apple Watch toward all-day health monitoring and wants users to wear it during sleep for continuous sleep and wrist temperature data, according to ifanr. Jeff Williams told ifanr in 2022 that people cannot actively watch their health at all times, so monitoring has to slip into daily life. GeekPark's Zhang Yongyi makes the same sensor argument from another angle: once basic accuracy is good enough, continuous and consistent data becomes more valuable.

That is the appeal of Fitbit Air's Health Coach. GeekPark says it tries to turn collected health data into directly understandable suggestions and can proactively offer next actions. The hardware is also positioned for reach: Fitbit Air costs $99.99 and includes three months of Google Health Premium, while WHOOP starts at $199 per year and includes hardware in the membership relationship.

But health advice is not a notification category. A confident mistake can change training, sleep, recovery, and anxiety.

Zhang's warning is the useful one: AI coach features can bring large-model failure modes into health scenarios, including hallucination, forgetting, misunderstanding context, and confident wrong answers. Fitbit Air also has no independent GPS, which matters because missing context can be as dangerous as bad inference. A model that speaks as if it knows why a run looked wrong, while lacking part of the location record, is not merely summarizing data. It is performing judgment.

The competitive pressure is clear. Apple Watch accounts for nearly 60% of global smartwatch sales by revenue, while IDC and Counterpoint data show global smartwatch shipment growth slowing to single digits. The same data put smart rings at 49% growth and AI glasses at 110% growth. Lighter devices create more chances for continuous sensing, but fewer visible moments to question the machine.

So the product question shifts from output quality to permission. The woshipm author's demand-review rule applies directly: ask whether the person enjoying the feature is the same person affected by it. For health wearables, that means consent boundaries, bystander rules when sensors expand beyond the wearer, and audits for missing context as much as hallucination. GeekPark's Zhang argues that after the screen disappears, the reliability of the model speaking for Fitbit Air will decide how far it can go. That is the interface now.

Treat saved document time as a budget, not a trophy. Chang Sixing says AI cut his writing-documents share from 40% to 15%; the useful part is what happened to the freed 25%. He moved it into checking outputs and making deeper decisions, according to woshipm.

Watch for the moment when a generated artifact looks complete before the scenario is understood. In 2022, Chang worked on a data analysis product for small and medium-sized e-commerce sellers, then spent 3 weeks in the offices of 5 sellers. The finding changed the product: 80% of sellers checked only sales, order volume, and advertising spend every day. The answer was not more "intelligent interpretation." It became abnormality alerts plus one-click attribution.

Use AI for the 70% replacement zone. Spend the remaining work on deciding what should interrupt a person, and what should stay quiet.

For readers outside China

• Availability: The source material covers Chinese commentary on Apple, Fitbit, WHOOP, Oura, AI glasses, Feishu Minutes, ChatGPT, Claude, Codex, Workbody, and government-service AI implementations, but it does not provide a full availability map outside China. The sspai author saw an 8000 yuan iPhone Air price on Apple's mainland China website, later borrowed a Hong Kong version after iPhone Air went on sale overseas, and bought iPhone Air for 5500 yuan. Fitbit Air is described as officially on sale for more than two months, but buyers still needed to pay an extra 200~300 yuan on second-hand platforms. Availability for Feishu Minutes, Workbody, the cited AI glasses, and the government-service deployments outside China is not disclosed in sources.
• Pricing: Prices disclosed in the sources are limited. iPhone Air appeared at 8000 yuan on Apple's mainland China website in the sspai account, and the author later bought iPhone Air for 5500 yuan. Fitbit Air is priced at $99.99 and includes three months of Google Health Premium; the same source says buyers still needed to pay an extra 200~300 yuan on second-hand platforms. WHOOP starts at $199 per year and includes hardware in the membership relationship. Other pricing is not disclosed in sources.
• Closest Western equivalents: Feishu Minutes is closest in role to AI meeting-note tools that turn recordings into decisions, to-do items, and risks; specific Western product comparisons are not disclosed in sources.; Workbody is named alongside ChatGPT, Claude, and Codex as an AI tool entering product-manager workflows; a closer Western equivalent is not disclosed in sources.; Fitbit Air is positioned against WHOOP and Oura-style screenless or low-interaction health wearables focused on continuous data rather than smartwatch interaction.; AI glasses are discussed as a category whose current high-frequency functions are mostly first-person shooting and voice interaction; the sources mention Meta as tackling sensing and social-perception thresholds.
• Data residency: Data residency, cross-border data transfer, storage location, and enterprise privacy terms are not disclosed in sources. The sources do note that health and ambient devices raise trust issues: Fitbit Air's value depends on continuous health data flowing to AI, and GeekPark warns that AI health coaches can hallucinate, forget, misunderstand context, or give confidently wrong answers. For AI glasses, woshipm highlights a social harm boundary: the wearer may benefit from photography, video recording, and real-time understanding while nearby people experience the device as a threat.

Sources

• woshipm AI降低了产品工作的门槛，却提高了产品经理的门槛 https://woshipm.com/share/6437224.html
• woshipm 评审需求时，我只问四个问题 https://woshipm.com/share/6428346.html
• sspai 我，与「唯一」的 iPhone Air https://sspai.com/post/112880
• ifanr 下一块 Apple Watch，苹果连屏幕都不想要 https://ifanr.com/1674558
• woshipm AI到底能不能取代产品经理？我拿自己做了半年实验 https://woshipm.com/ai/6429434.html
• geekpark Fitbit Air 手环深度体验：「无屏」只是手段，模型能力才是护城河 https://geekpark.net/news/368562