Rise turns AI latency into a workflow cost

Rise, Xingjie Intelligence's Era-series immersive-generation model, makes AI's economic question concrete: its listed price is 1 cent per second, alongside a 500-millisecond imperceptible delay. According to geekpark, it followed Genesis in less than one month.

The model call is only the opening cost.

woshipm's author argues that a useful measure is the full cost of an acceptable business outcome, not procurement or model usage alone. Verification can turn a cheap output into an expensive workflow. Rework can do the same, while maintenance and depreciation add to the workflow's cost. Failures do as well. Waiting adds cost, as can human takeover. The real test is where responsibility lands when the output is wrong.

36kr's Project Recommendation pairs AI-assisted generation with structured collection. It also uses human review. That pattern frames the synthesis: AI changes work only when its speed fits the job. Oversight must fit as well. Its operating burden must also fit the job.

Price the completed task, not the model call

Rise makes the model-call price easy to see. According to geekpark, Xingjie Intelligence released Rise less than one month after Genesis, pricing the Era-series immersive-generation model at 1 cent per second and claiming a 500-millisecond imperceptible delay. That price and delay describe generation. They do not describe the completed task.

An acceptable outcome begins only when somebody can use it. The result must be usable.

The woshipm author argues that companies should price that outcome in full, including procurement and model calls. Waiting and verification also belong in the calculation. Rework also consumes resources. Maintenance creates further expense over time. Depreciation and failures can add to that burden, while human takeover sets the real cost when automation cannot finish safely. Without a stopping condition, an apparently cheap task can keep absorbing effort after its useful value has ended.

The publishing flow described by 36kr makes the distinction concrete. Project Recommendation combines structured information collection with AI-assisted content generation. Human review follows. Project content must still undergo that review. The model may reduce the effort involved in producing a draft, but the workflow has not produced an acceptable item until review is complete. Rise's low delay can matter where waiting holds up that handoff; its 1-cent-per-second rate can matter where generation runs at scale. Neither figure alone establishes the cost of a publishable result. The relevant unit is the task that survives review. It must also survive rework or any required human takeover.

Start where mistakes are cheap to catch

A useful AI pilot begins with work whose output can be checked before it causes wider damage. The benchmark is not a general mandate to use a model. It is a narrow task with a known baseline: can people complete more of it, more accurately, while a named person checks the cases the model cannot safely decide? woshipm cites a knowledge-worker experiment in which AI users completed more tasks in less time, and quality was higher when the work stayed within the model's capability boundary. Accuracy fell when tasks crossed that boundary.

That boundary should define the pilot, not appear as an exception after deployment.

The measurable upside can be substantial, but it is uneven. A study of 5,179 customer-service workers cited by woshipm found an AI assistant raised average issues resolved per hour by 14%. Novice and low-skilled workers improved by 34%, while experienced workers saw little improvement. That pattern gives a practical test: start where the assistant can support repeatable work and where gains for less experienced staff are visible. Then specify the human-verification points. McKinsey survey high performers more often defined when people had to verify model outputs, according to woshipm.

A workflow-native product should make those boundaries legible. 36kr describes KuWork, Kuailu Technology's enterprise AI work-partner platform, as aiming to fit AI into real business workflows rather than leaving it in standalone chat or question-answering tools. Its stated capabilities span task management, information processing, collaborative office work, and AI assistance. The distinction matters because a deployment can be judged on completed work, review handoffs, and error detection-not on how often employees open a chatbot.

AI approaches: workflow integration, control, and cost exposure

Ask whether the AI changes the product architecture

Architecture determines whether AI changes the work itself or merely sits on top of it. According to geekpark, Xingjie Intelligence says conventional real-time video generation splits action understanding from image generation into separate models, producing latency above 4 seconds. Its alternative is an end-to-end design: the system is built around the real-time task rather than assembled from distinct stages. That design, the company claims, changes both responsiveness and the economics of running the product.

The claimed difference is stark. Xingjie Intelligence puts traditional two-stage compute at 0.7 yuan per second and its end-to-end approach at 0.01 yuan per second.

That makes architecture a workflow question, not an implementation detail. At 0.01 yuan per second, geekpark reports, a two-hour AI live interactive session costs 72 yuan; the earlier range was 5,000 to 8,000 yuan. The point is not that every AI product needs an end-to-end model. It is that a product can become practical only when its model design removes a bottleneck embedded in the task.

By contrast, woshipm describes a common B2B SaaS pattern: choose an industry foundation model, add prompts or fine-tuning with industry data, then package the result as a module. A medical SaaS founder characterized an AI medical-record feature as a major company's medical foundation model with an adaptation layer. That may still be useful, but it does not automatically create a defensible change in product architecture.

Miss Zhuozhu's test is more demanding: scarce industry data may matter more than algorithms or computing power. A SaaS company needs a data moat, in-house AI capability, or deep industry know-how. Without one of those advantages, the AI feature may remain a replaceable model call behind a new interface.

Budget for scarce compute without mistaking activity for value

Compute scarcity can turn an apparently cheap AI feature into a constrained operating resource. According to woshipm, Gavin Baker expects confirmed projects to consume all available computing capacity through 2028, with no surplus left over. He also warns that token prices may rise if demand keeps exceeding supply; some forecasts anticipate a tenfold increase. That possibility makes capacity planning a workflow question, not a procurement footnote.

More generated material does not prove more value.

The woshipm author argues that AI consumerism often celebrates model counts and usage volume. It also celebrates output volume and project launches. It overlooks revenue. It overlooks profit and cycle time.

Quality can receive too little attention. Risk and customer outcomes can also receive too little attention.

Cheap generation can also shift work downstream, so someone must read the output before verifying it. That person must correct it and make the decision that carries responsibility. If organizational judgment does not expand with generation, extra model activity can make a company busier rather than more productive.

That is why utilization needs an outcome attached to it. 36kr reports that Yuanyao Zhihui designs for "human-machine isomorphic AI efficiency management" and focuses on cost problems after companies adopt large numbers of models. It also addresses efficiency issues and organizational-management problems involving agents or digital employees. Its stated aim is to show how AI resources are used. It also aims to show how much value they produce. It further aims to show how collaboration between people and AI can improve. The useful budget is therefore not a quota for calls or tokens. It is a way to decide which scarce compute commitments reduce meaningful work, and which merely create more output awaiting human attention.

Choose AI by workflow fit, not by model presence

• A task sits within AI's demonstrated capability boundary and has a measurable throughput problem, such as customer service. Run a bounded pilot with outcome measures such as issues resolved per hour, quality, cycle time and rework. A study of 5,179 customer-service workers found an AI assistant increased average issues resolved per hour by 14%, with a 34% improvement for novice and low-skilled workers; experienced workers saw little improvement.
• The workflow requires reliable decisions or produces outputs that must meet a defined quality boundary. Deploy only with explicit human-verification points and named human responsibility for exceptions. Organizations classified as AI high performers were more likely to redesign workflows and to specify when model outputs had to be verified by humans; accuracy declined when knowledge-worker tasks exceeded AI's capability boundary.
• A company is considering an AI feature mainly because competitors are adding one, while its core product is unprofitable, renewals are declining, or customers do not pay for basic functions. Do not divert the core team into an AI rebuild. Put resources back into the core product until there is a specific operating bottleneck and evidence that customers will pay. One CRM SaaS company reportedly shut down after spending nine months on an AI sales assistant that customers did not buy, while its core product iterations fell behind.
• A B2B SaaS provider wants to package a model call, prompt layer or fine-tuned model for a new vertical. Proceed only if the company has a data moat, in-house AI development capability, or deep industry know-how. Treat each vertical as a separate cost and validation problem: the source material argues that industry data is scarce, outcomes can be hard to verify, and success in one industry may not transfer to another.
• A use case depends on real-time interactive video, where a multi-second response breaks the experience. Consider an end-to-end immersive-video architecture only after testing the claimed latency, quality and operational limits in the intended setting. Xingjie Intelligence describes Rise as priced at 1 cent per second with a 500-millisecond imperceptible delay, but its CEO also says current immersive-video technology still differs from a true world model in physical simulation, memory, multi-user participation and visual detail.

Put responsibility into the operating contract

When probabilistic AI enters work that customers treat as certain, responsibility cannot remain implied. The operating contract needs a named human reviewer who owns the output boundary and basic quality, reflecting 36kr's account of human review in content workflows.

That owner should decide what the system may produce without escalation, what must be checked before release, and what it must never decide alone. The boundary is not a disclaimer. It is a deployment rule.

A workable approval path can begin with project-information review, then test whether the work meets content requirements before it moves into production and publication. That sequence follows the process described by 36kr. Records should retain the submitted project information, the applicable content requirements, the AI output, and the reviewer's decision, so a later dispute can be traced to an accountable action rather than an unexplained model result.

Human takeover also needs a trigger, not a vague expectation that someone will intervene when needed. woshipm's Miss Zhuozhu argues that B2B customers need certainty while AI outputs are probabilistic; the contract should therefore route uncertain cases to the responsible person before a customer receives an answer. It should do the same when a case is consequential or crosses an established boundary.

Verification cannot rely on a dashboard metric alone. Miss Zhuozhu also argues that conversion rates and decision quality are hard to attribute to AI because too many variables affect them. Review should therefore examine the actual output against its permitted use, rather than treating an apparent business result as proof that the system acted correctly.

The same caution applies when an interface looks more capable than its underlying model. According to geekpark, Wang Yuxin identifies gaps between immersive-video technology and a true world model in physical simulation, memory, multi-user participation, and visual detail. A contract should assign humans to the gaps the system cannot reliably cover.

Start with a workflow where information already moves between people and where the handoffs create visible collaboration cost. Ask whether an agent can use participants' needs and context to improve that handoff. It can also use their relationships, rather than adding another place for employees to ask questions.

Make the test specific. According to 36kr, KuWork combines task management with information processing. It also supports collaborative office work through AI assistance. Lianlian AI positions agents as a hub for matching information. It assists communication and connects resources. A proposed deployment should show which action changes the existing path of work. If it cannot identify the information flow or collaboration link it will alter, treat it as a standalone tool rather than a workflow advantage.

For readers outside China

• Availability: Availability outside China is not disclosed in sources. The clearest public access detail is 36Kr Project Recommendation's free submission link, https://36kr.com/seek-report-new, and its contact email, aireport@36kr.com. The sources do not say whether KuWork, Lianlian AI, Yuanyao Zhihui, Xingjie Intelligence's models, or 36Kr's service are available internationally.
• Pricing: The only product pricing disclosed is for Xingjie Intelligence's Rise: 1 cent per second. The company claims that a two-hour AI live interactive session costs 72 yuan, compared with a previous cost of 5,000 to 8,000 yuan. Pricing for the other services is not disclosed in sources.
• Closest Western equivalents: No specific Western equivalent is identified in the source material.; KuWork is described as an enterprise AI work-partner platform for task management, information processing, collaborative office work and AI assistance.; 36Kr Project Recommendation is described as a structured project-submission and publication service using AI-assisted drafting and human review.
• Data residency: The source material does not cover hosting locations, cross-border data transfer, data residency, retention, model-training use of customer data, or enterprise security controls. Organizations handling regulated, confidential or personal data would need to obtain those details directly from the provider before deployment.

Sources

• 36kr 让项目从"被介绍"，走向"被理解与连接" https://36kr.com/p/3964434609069320
• geekpark 形界智能：沿用两段式架构，是市场对 AI 实时视频的最大误判 https://geekpark.net/news/369264
• woshipm 当所有人都在担心AI泡沫，a16z却说真正的风险完全相反 https://woshipm.com/ai/6457837.html
• woshipm 拥抱AI，为什么死得更快：To B SaaS的AI陷阱 https://woshipm.com/ai/6440913.html
• 36kr 第16期｜新锐项目观察：10个正在快速冒头的优质标的 https://36kr.com/p/3963287670766729
• woshipm 企业 AI 化是否正在成为企业版的消费主义陷阱？ https://woshipm.com/ai/6444907.html