硅谷101

E223 | The Year of AI Application Explosion: Discussing Model Evolution and Commercialization

1/30/20261:06:42

Editorial Note

<img src="https://imagev2.xmcdn.com/storages/a723-audiofreehighqps/B7/79/GAqhfD0NByqOAAHpNQRBU__s.jpg" alt=""> 2025 is being hailed as the "Year One of AI Applications." Large models are no longer just technical toys in laboratories; they are truly moving onto production lines, into enterprise workflows, and into the daily lives of consumers. AI has charged into various industries and started to "monetize"! This episode's conversation reveals a clear signal of the times: large models are shifting from "technical enthusiasm" to "business pragmatism." Whether it is Alibaba promoting industry-wide penetration through the Qwen (Tongyi Qianwen) model, or companies like Insta360 and Usay Technology integrating AI into smart hardware and enterprise data analysis on the front lines, it all points to one fact: the value of AI lies not in how "smart" it is, but in how "useful" it is. At the current stage, with costs falling, tools maturing, and enterprise awareness rising, future competition may not lie in who has the largest model, but in who best understands enterprise scenarios, can best package business value, and can iterate most sustainably. If you are wondering: ● How will AI land in my industry? ● Is it actually cost-effective to invest in AI? ● Which AI capabilities will become standard in the future? This episode might provide you with plenty of inspiration. In this episode, we invited Xu Dong, General Manager of Alibaba Cloud's Qwen Model Commercialization, along with two enterprise representatives on the front lines of AI implementation—Professor Qi Lu, Director of the Insta360 Research Institute, and Lyu Yingjie, Co-founder & CEO of Usay Technology. Together, we deconstruct the logic, costs, and hidden challenges of "breaking ground" on the AI commercial front. 【Host】 Hongjun, Founder of Silicon Valley 101, Podcast Host 【Guests】 Xu Dong, General Manager of Alibaba Cloud Qwen Large Model Business Professor Qi Lu, Director of Insta360 Research Institute Lyu Yingjie, Co-founder & CEO of Usay Technology 【You Will Hear】 Model Technological Progress and Commercialization 03:23 Video generation enters the era of "controllable production" 06:57 How AI comics and short dramas are making money 10:12 Progress in language models in 2025 16:27 AI + smart hardware: You can now buy coffee via smart glasses 19:18 Two directions for large model commercialization: productivity and user experience 22:13 Cloud-edge collaboration: 70% of common tasks processed locally 23:51 Shift in B2B user focus: From model performance to efficiency and cost 26:45 Driving factors behind the exponential decline in inference costs 29:31 Why develop "open-source" large models? How to balance open-source and closed-source models? 33:33 Evolution of evaluation metrics: From Tokens to business value 35:40 Foundation model R&D enters a stage of meticulous refinement, focusing on practical Agent capabilities How Insta360 Uses AI 38:54 AI editing: From "stitching" to "understanding" 40:49 Challenges facing AI: Understanding vague user intentions 43:02 How to extract and process material from panoramic cameras More Than Just Cost Reduction—Increasing Efficiency 48:57 What AI can do in the consulting industry 54:16 What capabilities to look for when choosing a foundation model 56:05 Using AI to help enterprises discover marketing opportunities 01:00:24 Competitiveness as a "middle layer" 01:03:34 B2B Observation: AI is making enterprises more willing to pay for SaaS 【Production】 Producer: Hongjun Post-production: AMEI Operations: Zhu Jie 【BGM】 All Parts Equal - Airae Poisson d'Avril - Ludvig Moulin 【Find Us Here】 WeChat Official Account: 硅谷101 (Silicon Valley 101) Listening Channels: Apple Podcast | Spotify | Xiaoyuzhou | Ximalaya | Qingting FM | Lizhi FM | NetEase Cloud Music | QQ Music Other Platforms: Search "硅谷101播客" on YouTube | Bilibili Contact us: <a href="mailto:podcast@sv101.net" rel="nofollow">podcast@sv101.net</a>Special Guests: Professor Qi Lu, Lyu Yingjie, and Xu Dong.

Voices

Keywords

Chapter 01Read Full

From 'Visual Toys' to 'Digital Studios': The 2025 Revolution of Controllable Video Generation

视频生成进入“可控生产”时代<ul><li>AI 模型正从单一语言处理演进为语言、视觉、音频三位一体的矩阵。</li><li>视频生成已跨越特效阶段，进入规模化生产：5人团队日产6000条视频成为可能。</li><li>“可控性”是当前技术的核心突破，支持人物、物体与背景的高度一致性。</li></ul>

AI 漫剧与广告：商业化落地的“第一桶金”<ul><li>国内短剧市场规模已超电影，AI 漫剧成为结合最紧密的应用场景。</li><li>AI 广告生成单条成本已降至 25-50 元，形成良性商业闭环。</li><li>广告主与电商卖家通过批量生成素材，极大提升了投放转化率。</li></ul>

2025 模型进化：更聪明、更快速、更精准<ul><li>2025 年关键词：稀疏结构 (MoE)、高推理能力、指令遵循。</li><li>响应速度 (TPS) 将从 30-50 提升至 100 以上。</li><li>AI 开始表现出“逻辑偏好”，能够执行包含跨软件操作的复杂指令。</li></ul>

Phase 1: Model Matrix

A Triadic Architecture

Qwen (Language)Cognitive Core
Wanx (Vision)Pixel Reconstruction
AudioLifelike Interaction

Phase 2: Efficiency

2025: The Efficiency Surge

2024 (TPS 30)

2025 (TPS 100+)

Sparse architectures will triple response speeds while significantly reducing operational costs.

Phase 3: Business

The Commercial Sweet Spot

AI Short Dramas

Controllable Costs: Motion Comics Enter the Mainstream

AI Advertising

5-Person Team / 6,000 Videos per Day

Production Efficiency: Traditional vs. AI

Metric	Traditional Pipeline	AI-Augmented Pipeline
Labor Cost	Extensive Production & Post-Production Teams	Lean Team (approx. 5 people)
Output Capacity	1–5 Premium Videos / Day	Approx. 6,000 Videos / Day
Controllability	Production Defines the Outcome	Replicable Characters, Objects, and Sound

“Enhanced control means AI is no longer about random generation; it has evolved into a professional-grade digital studio.”

“In 2025, AI will transition from ‘partial involvement’ to ‘full-cycle production’—a triumph not just of speed, but of certainty.”

Chapter 02Read Full

Hardware Awakening: When Large Models Gain Eyes, Ears, and Limbs

物理世界的‘交互闭环’：智能眼镜能买咖啡了？AI 硬件正在经历从简单的语音识别（ASR）到深层语义理解的跨越。通过视觉与文本模型的结合，智能硬件已能实现从‘看到需求’到‘完成支付’的完整闭环。

商业化的十字路口：提升生产力 vs 优化用户体验大模型的商业化分为两个核心维度：企业侧通过流程再造提升‘生产力’；消费侧通过硬件交互重塑‘用户体验’。其中，端侧模型（计算在本地）的崛起成为关键转折点。

01 / Hardware Evolution

Sensory Augmentation: From ‘Recognition’ to ‘Understanding’

Smart glasses are evolving beyond mere capture; possessing the dual perception of visual and textual models, they can ‘interpret’ a menu and complete payments in a closed loop, much like a human.

02 / Business Logic

Dual-Engine Drive: Productivity and Experience

• Enterprise:Reshaping workflows, result-oriented.
• Consumer:Reshaping interaction, gateway-oriented.

03 / Tech Ratio

Edge-Cloud Integration: The 7:3 Golden Ratio

Edge (70%)

Cloud (30%)

In the future, 70% of general tasks—such as voice and basic vision—will be processed locally to ensure both privacy and responsiveness.

AI Glasses: End-to-End Coffee Ordering Workflow

Visual Capture:Snapping a photo of a colleague’s coffee or a roadside menu

Logical Reasoning:Tailored recommendations based on your preferences and historical patterns

Execution:Seamless API integration for automated ordering and frictionless payment

Chapter 03Read Full

From “Intelligence Wars” to “Economic Realism”: The Second Act of LLM Commercialization

商业化的真谛：客户不再为“花架子”买单企业级用户对AI的需求已进入‘严肃生产’阶段，关注点全面转向TPS（并发处理能力）、海量输入下的响应速度以及极端的成本控制。

Focus Shift

From Efficacy to Performance

Where we once prized a model's ability to write poetry, we now demand speed, stability, and affordability. It is the evolution from a laboratory prototype to a mass-market vehicle.

Cost Reduction / The Playbook

Model Sparsification (MoE)
Hardware-Software Synergy (Silicon + Cloud)
Precision Scheduling (Idle Capacity Optimization)

Ecosystem / The Strategic Blueprint

Open Source as a Moat

Nurturing the world’s brightest minds—developers and students—within the Qwen (千问) ecosystem is a priority that transcends mere profit.

The First Principle of Model Selection: Fit Over Power

Intent Recognition / Data Extraction Opt for 0.5B - 7B Small Models

Exceptional cost-efficiency, lightning-fast response

Complex Reasoning & Creation Opting for 72B+ Models

COST COMPONENT ANALYSIS / Inference Cost Structure

* Data derived from logical deduction; for visualization purposes only.

Cost is the ultimate metric; inference costs for large models are dropping by an order of magnitude nearly every six months.

Chapter 04Read Full

From “Billing by the Word” to “Understanding the Mind”: AI Enters the Era of Refinement

评价标准的维度跃迁：Token 之后是什么？探讨 AI 评估体系的去泡沫化：从量化字符消耗转向量化任务结果。大模型研发也已进入拼细节、拼 Agent 工具调用能力的新阶段。

01 / Metric Evolution

From Counting to Outcomes

Early evaluations prioritized token consumption; the future belongs to task completion. It is like hiring a chef: once paid for the volume of prep work, now judged by the quality of the feast.

02 / Meticulous Cultivation

The Rise of Agents

The Perfection of Tool-Calling Capabilities
Precision Execution of Complex Instructions
The Infinite Refinement of Engineering Details

03 / Visual Revolution

Capturing Intent

AI editing is no longer about mechanical splicing; it utilizes semantic understanding to distinguish between “documentary aesthetics” and “highlight reels,” achieving true aesthetic alignment.

AI+

The “Refined” Path of Model Evolution

The definition of intelligence is shifting: if intelligence could be scored, today’s intensive refinement is about raising that score in real-world business applications.

Chapter 05Read Full

From Panoramic Vision to Intent Decoding: How AI is Reshaping Digital Editors and Shopping Assistants

AI 剪辑的终极命题：读懂你的‘弦外之音’剪辑不仅仅是拼接画面，更是对用户模糊意图的精准捕捉。影石通过自研全景理解模型，试图在海量 360° 素材中自动识别高光时刻，降低普通人的创作门槛。

AI 进军零售业：从流水线客服到‘金牌咨询’语忆科技展示了 AI 如何在消费领域实现‘意图标签化’。通过识别客服对话中的肤质、反馈和情绪，AI 不仅提高了服务准确率，还通过自动化归因重塑了企业的管理绩效。

Sector 01 / Visual Intelligence

Insta360: Panoramic Reconstruction

Using proprietary panoramic understanding models to distill highlights from raw 360° footage, resolving the 'paradox of choice' in 'shoot first, edit later' workflows—a step toward a video-based 'World Model.'

Sector 02 / Intent Recognition

Yuyi: Consumer Attribution

Tagging customer service dialogues. Accuracy has been elevated from 70% to Agent-level, enabling real-time identification of skin types and allergy feedback, while automatically attributing return reasons to the responsible departments.

Sector 03 / Economic Impact

Optimizing Cost and Efficiency

Video Processing Costs: High hardware overhead has been compressed through technology to approximately ¥10 per segment. Retail Operations: A six-person attribution team has evolved into fully automated AI performance evaluation.

Panoramic AI Processing Workflow

360° Panoramic Footage (Massive Raw Data)

→

Proprietary Multimodal Model (Highlight Extraction / Framing Suggestions)

→

General LLMs (Flat Content Orchestration / Narrative)

* Why not simply use open-source models? Because global panoramic data is scarce, and open-source models lack a nuanced understanding of 360-degree perspectives.

The barrier in AI editing is not merely visual comprehension, but the elimination of the “translation cost” when a user expresses intent.

Chapter 06Read Full

From Efficiency to Expansion: How AI Reshapes Business Intuition

意图识别：AI 正在成为电商的“读心者”<ul><li>基模选择： 语忆科技选择通义千问（Qwen）是看中其在复杂电商文档处理及中国消费者语义理解上的卓越表现。</li><li>核心壁垒： “中间层”不仅仅是接口转发，更通过留存行业垂直数据，训练出比基座模型更懂业务的“行业专家”模型。</li></ul>

01 / Model Selection

Intent Recognition

Tongyi Qianwen (Qwen) was selected as the foundation for its precision in capturing the nuanced intent of Chinese consumers within the e-commerce landscape.

02 / Commercial Growth

From Retention to Expansion

Automated Cost-Saving (2023) Limited Ceiling

AI-Driven Opportunity (2024+) Infinite Growth

03 / The End Game

The Decision-Maker’s Insight

As AI tools become commoditized, the capacity for human-led decision-making—informed by AI-generated insights—defines a company's ultimate ceiling.

Visual Analysis: AI Business Evolution Paper Texture Aesthetics

BUSINESS LOGIC TRANSITION

Cost Reduction (Stock)

Revenue Growth (Incremental)

Chart: The strategic shift in corporate AI investment objectives

The true ceiling of AI is not defined by labor savings, but by its ability to uncover business opportunities that were previously invisible.

Chapter 07Read Full

From 'Software Subscriptions' to 'Utility Bills': How AI is Reshaping the B2B Ledger

从功能付费到算力付费AI 时代的商业变革在于成本的显性化。当软件背后是真实的算力支出，中国客户正逐渐接受‘按量计费’的新逻辑，这为 SaaS 行业带来了前所未有的高增速机会。

Fundamental Principle

A Generational Leap in Business Logic

Traditional SaaS relies on 'feature subscriptions,' whereas AI SaaS shifts toward 'cost-based pricing.' Given the tangible computing power consumed by AI, clients are more inclined toward a utility-style, pay-as-you-go model.

Metrics

Data Granularity 1,000,000+

Total processing volume across voice, social media, and intent recognition.

#High-Growth Commercial Model

Step 01

Raw Data Collection

Acquiring massive, fragmented data across voice, social media, and interactive channels.

Step 02

Advanced AI Model Refinement

Intent Recognition, Attribution Analysis, and Persona Profiling

Step 03

Volume-Based Value Delivery

Scaling costs by processing volume to achieve commercial closure

Visualized by Pod-Design-System · Minimalist Paper Aesthetics

Business Model Comparison: Traditional SaaS vs. AI SaaS

Dimension	Traditional SaaS	YuYI Tech AI Model
Core Billing Metric	Feature access / Seat count	Data processing volume (Usage)
Value Perception	Access to tools	Tangible outcomes and strategic insights
Cost Structure	Near-zero marginal cost	Computing & API: Explicit Costs

AI has taught enterprises to pay for software as they would for a utility.

Chapter 01

From 'Visual Toys' to 'Digital Studios': The 2025 Revolution of Controllable Video Generation

Phase 1: Model Matrix

A Triadic Architecture

Qwen (Language)Cognitive Core
Wanx (Vision)Pixel Reconstruction
AudioLifelike Interaction

Phase 2: Efficiency

2025: The Efficiency Surge

2024 (TPS 30)

2025 (TPS 100+)

Sparse architectures will triple response speeds while significantly reducing operational costs.

Phase 3: Business

The Commercial Sweet Spot

AI Short Dramas

Controllable Costs: Motion Comics Enter the Mainstream

AI Advertising

5-Person Team / 6,000 Videos per Day

“

视频生成进入“可控生产”时代

AI 模型正从单一语言处理演进为语言、视觉、音频三位一体的矩阵。
视频生成已跨越特效阶段，进入规模化生产：5人团队日产6000条视频成为可能。
“可控性”是当前技术的核心突破，支持人物、物体与背景的高度一致性。

Read Insight

泓君开场并介绍嘉宾 03:23

Hello everyone, and welcome to Silicon Valley 101. I’m Hongjun. I remember when we discussed large models two years ago; they were fraught with issues and felt distant from our daily work and lives.

Today, how do the evolution of AI capabilities and commercialization intertwine?

We are joined by Xu Dong (Dongge), General Manager of Alibaba Cloud’s Tongyi business, to discuss the latest progress in the commercialization of large models.

徐栋介绍通义大模型家族 04:50

Hello everyone. Tracing back our foundation models, the Tongyi family is currently built upon three core pillars:

Qwen (Tongyi Qianwen): Our large language model and a leader in the open-source community, representing logic and linguistic intelligence.
Wanx (Tongyi Wanxiang): Powering image and video generation and editing, it serves as the prototype for a future 'World Model'.
Audio (Tongyi Audio)Debuted at last year’s Apsara Conference (云栖大会), this audio-only model pursues hyper-realism and a sophisticated understanding of multilingual dialects.

Note | Editor’s Note: This ‘multimodal’ paradigm implies that AI is no longer a mere chatbot, but an integrated entity endowed with sight, hearing, voice, and intellect.

泓君追问技术进化关键点 06:15

In your view, what have been the most significant technological evolutions in AI models over the past six months?

徐栋解析视频生成的质变 06:30

I believe the most critical advancement is that video generation models have matured from 'visual novelties' to the stage of 'controllable production.' Where AI was once a tool for playful effects, we are now seeing:

Production Efficiency: Traditional vs. AI

Metric	Traditional Pipeline	AI-Augmented Pipeline
Labor Cost	Extensive Production & Post-Production Teams	Lean Team (approx. 5 people)
Output Capacity	1–5 Premium Videos / Day	Approx. 6,000 Videos / Day
Controllability	Production Defines the Outcome	Replicable Characters, Objects, and Sound

Today’s models are producing longer videos with increasing coherence. More importantly, the consistency of characters, objects, and vocal timbres is now preserved, offering vast potential for creative expansion.

“Enhanced control means AI is no longer about random generation; it has evolved into a professional-grade digital studio.”

徐栋解释角色扮演与分身技术 09:15

For instance, we can now achieve ‘Character Replication.’ A mere five-second selfie allows the model to mirror both image and voice, elevating our control over generated content to a new dimension.As input parameters multiply, AI executes instructions with newfound precision.

“

AI 漫剧与广告：商业化落地的“第一桶金”

国内短剧市场规模已超电影，AI 漫剧成为结合最紧密的应用场景。
AI 广告生成单条成本已降至 25-50 元，形成良性商业闭环。
广告主与电商卖家通过批量生成素材，极大提升了投放转化率。

Read Insight

泓君探讨商业应用案例 10:05

Given these advancements, what compelling commercial use cases have emerged?

徐栋分析短剧与广告市场 10:20

The domestic short-form drama market has actually surpassed that of traditional cinema. We’ve observed thatShort-form dramas are transitioning from live-action to AI-integrated production—most notably in the surging trend of ‘Manju’ (Motion Comics).Motion Comics have become a standard-bearer for AI-content integration. AI advertising is another sector that has established a robust commercial cycle.

泓君询问成本对比 11:15

The primary concern remains cost: exactly how much cheaper is AI advertising compared to live-action production?

徐栋给出行业成本数据 11:40

In the current market,A standard 15-second AI-generated advertisement currently retails for approximately 25 to 50 RMB.Compared to the overhead of locations, talent, and post-production inherent in live-action shoots, this pricing is exceptionally competitive. For e-commerce vendors and digital platforms, AI enables the mass production of creative assets, allowing them to optimize performance through high-volume testing.

“

2025 模型进化：更聪明、更快速、更精准

2025 年关键词：稀疏结构 (MoE)、高推理能力、指令遵循。
响应速度 (TPS) 将从 30-50 提升至 100 以上。
AI 开始表现出“逻辑偏好”，能够执行包含跨软件操作的复杂指令。

Read Insight

泓君展望 2025 年技术突破 13:10

Looking ahead to 2025, what will define the core advancements in language and speech models?

徐栋解析技术演进趋势 13:25

I see three pivotal directions:

Sparse Architectures (MoE): Model structures are becoming increasingly sparse, driving higher speeds and lower costs.
The TPS Surge: While response speeds of 30–50 TPS were once considered impressive,exceeding 80–100 is highly probable in the near future,shifting the user experience from “waiting for results” to “instant interaction.”
Long-context and Precise Command: Models are evolving beyond simple responses to handle expansive contexts and orchestrate external tools like CRM and ERP systems for complex task execution.

泓君追问 AI 为何变“听话”了 15:10

Why the sudden consensus that 2025 will see an explosion in AI applications? Is it because “controllability” has improved?

徐栋解释逻辑推理与指令遵循 15:30

The core driver is Reasoning. Models are no longer merely predicting the next token based on probability; they are exhibiting a distinct “logical preference.” Through meticulous data curation and enhanced instruction-following, today’s AI functions like a highly perceptive assistant. Given complex, multi-step instructions, it can navigate context to make autonomous decisions, finding the ideal balance between generalization and accuracy.

“In 2025, AI will transition from ‘partial involvement’ to ‘full-cycle production’—a triumph not just of speed, but of certainty.”

Chapter 02

Hardware Awakening: When Large Models Gain Eyes, Ears, and Limbs

01 / Hardware Evolution

Sensory Augmentation: From ‘Recognition’ to ‘Understanding’

Smart glasses are evolving beyond mere capture; possessing the dual perception of visual and textual models, they can ‘interpret’ a menu and complete payments in a closed loop, much like a human.

02 / Business Logic

Dual-Engine Drive: Productivity and Experience

• Enterprise:Reshaping workflows, result-oriented.
• Consumer:Reshaping interaction, gateway-oriented.

03 / Tech Ratio

Edge-Cloud Integration: The 7:3 Golden Ratio

Edge (70%)

Cloud (30%)

In the future, 70% of general tasks—such as voice and basic vision—will be processed locally to ensure both privacy and responsiveness.

“

物理世界的‘交互闭环’：智能眼镜能买咖啡了？

AI 硬件正在经历从简单的语音识别（ASR）到深层语义理解的跨越。通过视觉与文本模型的结合，智能硬件已能实现从‘看到需求’到‘完成支付’的完整闭环。

Read Insight

泓君观察智能硬件趋势 16:27

At the recent CES, I observed a compelling trend: from earbuds to smart glasses, manufacturers are racing to integrate AI. Smart glasses, in particular, have evolved beyond mere tools for filming and editing to become the ultimate vehicle for AI.

徐栋解释硬件理解力的进化 17:05

Merging hardware with AI is not a new concept; ASR (Speech Recognition) has been a staple for years. However, the fundamental shift brought by this wave of large models is:It has become more "human-like."

Note | Editor’s Note: While ASR (Automated Speech Recognition) previously only transcribed speech to text, today’s large models can grasp the underlying intent.

In the past, hardware could only "recognize"; today, it can "perceive" and "comprehend" to deliver the desired results. Our Tongyi Qianwen model, for instance, fuses visual and textual understanding to imbue hardware with the capacity to think.

泓君提出具象场景疑问 17:45

"So, is the scenario of ordering coffee directly through your glasses now a reality?"

徐栋确认场景落地并详解流程 17:50

"To be precise, it has been possible for a while. Choosing the blend, sugar levels, the shop, and the delivery time—even the final payment—can all be completed through a perfectly natural interaction."

AI Glasses: End-to-End Coffee Ordering Workflow

Visual Capture:Snapping a photo of a colleague’s coffee or a roadside menu

Logical Reasoning:Tailored recommendations based on your preferences and historical patterns

Execution:Seamless API integration for automated ordering and frictionless payment

This approach offers a level of convenience that transcends the smartphone. It presents proactive interface cards, intelligently curated based on your proximity, preferences, and past choices.

AI hardware is no longer a cold utility; it is an intimate digital portal endowed with memory.

“

商业化的十字路口：提升生产力 vs 优化用户体验

大模型的商业化分为两个核心维度：企业侧通过流程再造提升‘生产力’；消费侧通过硬件交互重塑‘用户体验’。其中，端侧模型（计算在本地）的崛起成为关键转折点。

Read Insight

徐栋分析行业渗透与商业逻辑 19:20

The current surge of Large Models is industry-wide. Their commercial logic can be distilled into two primary categories:

Productivity Logic (B2B):Focusing on internal enterprise workflows. By applying outcome-oriented models to every stage, overall efficiency is elevated.
User Experience Logic (B2C):Targeting smartphones, automobiles, and smart hardware. Large models are fundamentally redefining the human-machine interface.

Our collaborations are deepening. While the initial focus was primarily on the cloud, smartphone manufacturers and automakers are now pivoting toward“Cloud-Edge Synergy”.

泓君总结消费端愿景 20:45

From a consumer perspective, the future implies that every device—be it a phone or a vehicle—will inherently possess deep intelligence.

徐栋阐述端侧模型的技术挑战 21:10

Precisely. As personal “portals,” smart devices are uniquely positioned to handle high-privacy, time-sensitive tasks.

Note | Editor’s Note: Privacy and latency are the primary catalysts driving AI from the cloud to the edge.

We released small-scale models like 0.5B specifically to operate on compute-limited mobile and automotive platforms while optimizing power consumption.

泓君追问端云分工比例 22:15

Which scenarios are handled locally? Which are deferred to the cloud? What is the projected ratio between the two?

徐栋给出核心数据结论 22:30

Today’s on-device models outperform the flagship models of just two years ago.

Task Type	Processing Tier	Projected Share
General Interactions (Voice & Vision Recognition)	Local (On-device)	Approx. 70%
Complex Logic & Large-scale Data Retrieval	Cloud	Approx. 30%

This implies that over 70% of future daily interactions will be handled locally by on-device models.

The First Principles of “Edge-First”

When model miniaturization crosses a tipping point (such as Qwen-0.5B), local silicon takes the lead. This enables:
1. Zero Latency: Eliminating the friction of network transit.
2. Absolute Privacy: Personal dialogues and biometric data never leave the device.
3. Cost Efficiency: Relieving manufacturers of the burden of cloud inference and bandwidth overheads.

Chapter 03

From “Intelligence Wars” to “Economic Realism”: The Second Act of LLM Commercialization

Focus Shift

From Efficacy to Performance

Where we once prized a model's ability to write poetry, we now demand speed, stability, and affordability. It is the evolution from a laboratory prototype to a mass-market vehicle.

Cost Reduction / The Playbook

Model Sparsification (MoE)
Hardware-Software Synergy (Silicon + Cloud)
Precision Scheduling (Idle Capacity Optimization)

Ecosystem / The Strategic Blueprint

Open Source as a Moat

Nurturing the world’s brightest minds—developers and students—within the Qwen (千问) ecosystem is a priority that transcends mere profit.

“

商业化的真谛：客户不再为“花架子”买单

企业级用户对AI的需求已进入‘严肃生产’阶段，关注点全面转向TPS（并发处理能力）、海量输入下的响应速度以及极端的成本控制。

Read Insight

泓君提问客户关注点 23:51

When collaborating with Alibaba, what do you believe is the most critical priority for clients?

徐栋分析关注点迁移 24:10

There is a distinct shift currently underway.

The past six months: Clients focused primarily on 'performance'—the raw intelligence of the model and its capacity to execute tasks.
Today: Clients are entering“Mission-critical production.”They have validated the model’s utility; the focus now is on cost-performance.

Specifically, two metrics have taken center stage:

TPS (Concurrency Performance): For instance, when facing 100,000 simultaneous queries, can the model keep pace? Is the response instantaneous?
Cost: AI usage has scaled a hundredfold; without significant cost reduction, enterprise adoption remains unsustainable.

Note | TPS (Transactions Per Second) measures throughput. In the AI domain, this typically refers to tokens generated or requests processed per second—akin to a vehicle’s top speed.

泓君观察行业成熟度 25:30

It appears that clients now possess a sophisticated understanding of AI. The industry is no longer in its “introductory phase,” but has entered a mature stage of discerning demands and vertical applications.

徐栋补充小模型趋势 26:15

Precisely. Over the past three months, we have seen a surge in demand for **“Small Models” (SLMs)**.

The First Principle of Model Selection: Fit Over Power

Intent Recognition / Data Extraction Opt for 0.5B - 7B Small Models

Exceptional cost-efficiency, lightning-fast response

Complex Reasoning & Creation Opting for 72B+ Models

Clients now favor 'Hybrid Deployment': utilizing LLMs for complex logic and proprietary small models for routine tasks to minimize total cost.

Cost is the ultimate metric; inference costs for large models are dropping by an order of magnitude nearly every six months.

徐栋拆解成本下降的三大动力 28:10

Why is the cost falling so sharply? It hinges on three primary levers:

Dimension	In Plain Terms	Key Technical Terms
Model Algorithms	Teaching models to 'economize' by activating only a fraction of their neural network.	MoE (Sparsification), MTP
Hardware Synergy	Bespoke chips and software designed to eliminate wasted compute power.	Hardware-Software Integration, Chip Optimization
Resource Scheduling	Serving users by day, running batch processes by night—maximizing utility during idle hours.	Cloud-native scheduling, smoothing the peaks and valleys of demand.

Take our upcoming version in R&D: despite having 80B parameters, it only activates 3B of energy for any given task. It is as if a giant were working on the rations of a child.

泓君对比谷歌与行业 30:00

This brings Google to mind. They have managed to bring costs down to a fraction of OpenAI’s because they possess their own silicon and cloud architecture. Thishardware-software synergycreates a formidable edge in the market.

徐栋阐述开源战略价值 31:15

Many ask: Why does Alibaba remain committed to open source?

In truth, it is our moat.

The Talent Pipeline:When today’s elite researchers and students grow accustomed to coding and researching with Tongyi Qianwen (Qwen), they become our organic user base upon graduation.
Ecosystem Feedback:Developers proactively adapt our models for mobile or Mac. This global, decentralized testing and iteration is a feat closed-source models simply cannot replicate.

泓君追问商用与开源平衡 32:30

How then do we balance open source with commercialization? If the open-source version is available, who pays for the service?

徐栋解释商业闭环 32:50

Open source is the draw—the foundation. However, enterprises require an advanced tier: something more intuitive, with higher concurrency and superior capabilities. This is where our proprietary APIs and cloud services come in.

Note | Think of open source as a free sapling. While the tree is yours, if you wish to enjoy the fruit immediately (high-performance APIs) or require a professional gardener to maintain the orchard (enterprise-grade management), you will naturally opt for the paid service.

Moreover, the progress of open-source peers like DeepSeek is propelling the entire industry forward. As these models prove their viability, more enterprises are encouraged to adopt AI, effectively expanding the collective market.

COST COMPONENT ANALYSIS / Inference Cost Structure

* Data derived from logical deduction; for visualization purposes only.

Chapter 04

From “Billing by the Word” to “Understanding the Mind”: AI Enters the Era of Refinement

01 / Metric Evolution

From Counting to Outcomes

Early evaluations prioritized token consumption; the future belongs to task completion. It is like hiring a chef: once paid for the volume of prep work, now judged by the quality of the feast.

02 / Meticulous Cultivation

The Rise of Agents

The Perfection of Tool-Calling Capabilities
Precision Execution of Complex Instructions
The Infinite Refinement of Engineering Details

03 / Visual Revolution

Capturing Intent

AI+

“

评价标准的维度跃迁：Token 之后是什么？

探讨 AI 评估体系的去泡沫化：从量化字符消耗转向量化任务结果。大模型研发也已进入拼细节、拼 Agent 工具调用能力的新阶段。

Read Insight

泓君探讨评估 AI 应用的标准 33:33

A prevailing topic of discussion this year is: how do we truly measure an application’s success?

The industry standard has been to quantify usage via Token consumption. Do you believe Tokens will remain a viable metric in the future?

徐栋分析 Token 指标的局限性 33:55

I believe it serves as a valuable benchmark for the current phase; indeed, many corporate valuations are rooted in it. However, in the long run, relying solely on Tokens has its limitations. For instance, the discrepancy in cost and commercial value between the smallest and largest models can be as much as a hundredfold.

Note | Tokens are the fundamental units of AI processing, much like word counts in an essay. But a high word count does not equate to a superior composition.

While Tokens remain a tool for monitoring usage volume, the future will place greater emphasis onTask Outcomes.

Evaluation Metrics	Past (The Token Era)	Future (The Task Era)
Core Metrics	Character Consumption	Task Success Rate / Commercial Value
Model Preference	The 'Bigger, Costlier, Better' Axiom	Context-Optimized Models
Delivery Format	Text Generation	Problem Solving via Agents

The definition of intelligence is shifting: if intelligence could be scored, today’s intensive refinement is about raising that score in real-world business applications.

泓君询问大模型研发热度 35:10

With the conversation now revolving around commercialization, is investment in foundational model R&D still significant? Has the hype subsided?

徐栋揭秘大模型研发现状 35:30

That is a misconception. Investment in model R&D has actually intensified. Every company and lab is now lamenting a 'compute shortage.' It is simply less overt than before—less about sudden, world-shaking model releases.Large models have entered the era of 'intensive cultivation.'Although the paradigm remains unchanged, there is a vast amount of work occurring within the engineering details. For example:

**Data Refinement Capabilities**: How to feed models higher-quality 'fuel'.
**Evaluation Metrics**: Precision in assessing model performance.
**Targeted Enhancements**: Elevating coding proficiency and the execution of complex instructions.

徐栋定义未来模型竞争点 36:40

The future belongs to **Agentic** capabilities. Models must continuously interact with their environment, engage in self-reflection, and formulate hypotheses. Our strategic focus for Tongyi Qianwen (Qwen) includes:

Precision in tool use.
Processing long-form context.
Nuanced understanding of complex instructions.

The convergence of these capabilities marks a paradigm shift for business implementation.

The “Refined” Path of Model Evolution

泓君引出影像 AI 嘉宾 38:00

Welcoming Lyu Yingjie, Co-founder of Yuyu Technology, and Professor Qi from Insta360.

Professor Qi, could you share how Insta360 integrates AI into its products?

亓鲁介绍 AI 自动剪辑战略 38:30

As a visual imaging company, our mission at Insta360 is to empower users to capture and share their lives more effectively. A prime example of our collaboration with Alibaba isCloud-based automated editing. A user captures hundreds of clips over a day; how can we swiftly identify that 1% of highlight-worthy content? That is our critical challenge.

泓君点赞产品体验 39:15

What I admire most about Insta360 is its ability to “make the camera disappear”—it is a truly remarkable feature.

What is the true challenge of automated editing?

亓鲁解析“剪辑意图”的捕捉 39:30

The difficulty lies not in splicing clips together, but in capturing the user’s creative intent. Preferences vary from person to person:

Some seek a **documentary style**: authentic, comprehensive, and narrative-driven.
Others want **highlight reels**: fast-paced, perfectly synced, and high-impact.

This requires a deep understanding of musical sentiment, cinematic language, and transition styles.

Note | AI must go beyond merely 'recognizing' objects—people, pets, landscapes—to 'feeling' the emotional resonance of the footage.

Our collaboration with Alibaba spans perceptual understanding and video generation models, evolving AI from 'seeing the frame' to 'understanding the heart.'

Chapter 05

From Panoramic Vision to Intent Decoding: How AI is Reshaping Digital Editors and Shopping Assistants

Sector 01 / Visual Intelligence

Insta360: Panoramic Reconstruction

Sector 02 / Intent Recognition

Yuyi: Consumer Attribution

Sector 03 / Economic Impact

Optimizing Cost and Efficiency

“

AI 剪辑的终极命题：读懂你的‘弦外之音’

剪辑不仅仅是拼接画面，更是对用户模糊意图的精准捕捉。影石通过自研全景理解模型，试图在海量 360° 素材中自动识别高光时刻，降低普通人的创作门槛。

Read Insight

泓君提出用户痛点 40:49

Suppose I prompt an AI tool: 'Clip every moment I wipe out while skiing.' How far can current models actually understand?

影石代表解释技术难点 41:15

The challenge today lies not in the raw footage, but in “Intent Recognition”.

The Layer of Understanding: Algorithms must determine what is happening within the frame (e.g., identifying skiing, a fall, or a highlight).
Intent Alignment: User descriptions are often ambiguous. If a product requires surgical precision—such as specifying the exact angle of a fall—the user experience becomes cumbersome.

We need AI to take vague instructions and automatically synthesize disjointed clips into a coherent, narrative-driven story.

The barrier in AI editing is not merely visual comprehension, but the elimination of the “translation cost” when a user expresses intent.

影石代表介绍全景相机壁垒 42:30

Insta360’s core competitive moat lies in its“Panoramic Understanding Model”.

Panoramic AI Processing Workflow

360° Panoramic Footage (Massive Raw Data)

→

Proprietary Multimodal Model (Highlight Extraction / Framing Suggestions)

→

General LLMs (Flat Content Orchestration / Narrative)

* Why not simply use open-source models? Because global panoramic data is scarce, and open-source models lack a nuanced understanding of 360-degree perspectives.

We first extract “highlights” through our proprietary model, then leverage General LLM APIs for final orchestration. This effectively touches upon the“World Model”in its nascent stage, as we are navigating a complete three-dimensional spatial scene.

泓君询问落地成本 44:50

That sounds expensive. How much does it cost a user to process a single clip? What are the drivers for cost reduction?

影石代表分析成本结构 45:15

Currently, processing a cloud-based clip costs the user roughly **10 RMB**.

Note | Editor’s Note: While 10 RMB may seem steep for a single edit, AI offers immense leverage compared to the hours of labor required for manual production.

Two primary drivers are reducing costs:

High-ratio compression: Despite the volume of panoramic footage, it is inherently redundant; technical optimization minimizes the frames processed by the model.
Compute orchestration: As Scaling Laws optimize inference costs, the unit price of cloud-based automated editing will continue to improve.

“

AI 进军零售业：从流水线客服到‘金牌咨询’

语忆科技展示了 AI 如何在消费领域实现‘意图标签化’。通过识别客服对话中的肤质、反馈和情绪，AI 不仅提高了服务准确率，还通过自动化归因重塑了企业的管理绩效。

Read Insight

语忆科技代表分析市场环境 47:15

Post-2023, China’s traffic dividend has peaked. Brands are shifting focus from mere acquisition toward“Consumer Insights”. Enterprises possess vast repositories of customer service data, which previously remained dormant—value left unextracted.

语忆科技代表对比传统与 AI 画像 48:50

Traditional user profiling focuses on demographics like age and location, factors that rarely correlate with actual purchasing behavior.

Dimension	Legacy Tags (Traditional Models)	AI Intent Recognition (Agent)
Core Data	Age, Gender, Geography	Skin profile, allergy reports, specific needs
Accuracy	Below 70%	High accuracy (viable for strategic decision-making)
Commercial use	Broad-spectrum advertising	Precision recommendations, product refinement, risk mitigation

语忆科技代表分享经典案例 51:20

We developed for a leading FMCG (Fast-Moving Consumer Goods) enterprise a“Attribution Model”. Previously, they employed **six people** specifically to manually log return reasons and assign departmental accountability. Today, AI autonomously analyzes every interaction—distinguishing between logistics, packaging, and quality issues—and integrates directly with the performance management system.

Note | Editor’s Note: This exemplifies AI’s capacity to automate repetitive white-collar labor—streamlining a team of six down to zero.

泓君总结应用层次 53:05

So, AI in retail transcends basic customer support, operating across three distinct layers: pre-sales conversion, post-sales resolution, and internal management?

语忆科技代表深化客服场景 53:40

Exactly. Particularly in pre-sales, AI is evolving from simple reactive responses to sophisticated logic processing. For instance, a customer might ask: “I’m based in Zhejiang; do I qualify for the national subsidy? I’m buying Item A—can you suggest a secondary item to help me meet the discount threshold?”

Large models can retrieve policy data and calculate the optimal coupon combinations in real-time. This kind of“Composite Sales Intelligence”transcends the capabilities of traditional customer service bots.

Chapter 06

From Efficiency to Expansion: How AI Reshapes Business Intuition

01 / Model Selection

Intent Recognition

Tongyi Qianwen (Qwen) was selected as the foundation for its precision in capturing the nuanced intent of Chinese consumers within the e-commerce landscape.

02 / Commercial Growth

From Retention to Expansion

Automated Cost-Saving (2023) Limited Ceiling

AI-Driven Opportunity (2024+) Infinite Growth

03 / The End Game

The Decision-Maker’s Insight

As AI tools become commoditized, the capacity for human-led decision-making—informed by AI-generated insights—defines a company's ultimate ceiling.

Visual Analysis: AI Business Evolution Paper Texture Aesthetics

“

意图识别：AI 正在成为电商的“读心者”

基模选择： 语忆科技选择通义千问（Qwen）是看中其在复杂电商文档处理及中国消费者语义理解上的卓越表现。
核心壁垒： “中间层”不仅仅是接口转发，更通过留存行业垂直数据，训练出比基座模型更懂业务的“行业专家”模型。

Read Insight

泓君询问大模型协作模式 54:16

How do you integrate foundation models like Tongyi Qianwen into specific business workflows? Beyond acting as an 'external knowledge base,' what core value do these large language models provide?

嘉宾阐述 AI 的知识处理能力 54:30

YuYi Technology (语忆科技) utilizes large models to process highly complex internal documentation. Even for something as visual as a product detail page, AI can distill the information into a structured knowledge repository.Large models do more than read text; they synthesize complex multimodal data into actionable 'knowledge assets' for the enterprise.

泓君追问筛选模型标准 54:55

When selecting a foundation model, what is the single most critical capability you prioritize?

嘉宾解析意图识别优势 55:10

Our priority is **Intent Recognition**. Through extensive testing, we found that Tongyi Qianwen demonstrates an exceptional grasp of authentic consumer intent within the e-commerce landscape.

Note | Editor’s Note: Intent recognition refers to an AI’s ability to read between the lines—understanding, for instance, that a query about shipping times may stem from underlying delivery anxiety.

泓君探讨市场逻辑 55:40

As the bridge between large models and downstream clients, would you say enterprises adopt AI to 'reduce costs' (existing markets) or to 'drive growth' (incremental markets)?

嘉宾对比 2023 与 2024 市场变化 56:05

It is an evolving, dynamic process.

2023:The focus was on **automation** and labor replacement—reducing a six-person audit team to one, for instance. Yet, there is a hard ceiling to the gains from cost-cutting.
H2 2024:More enterprises are pivoting toward **value enhancement**. They are leveraging AI to uncover latent consumer needs and identify new contextual opportunities for legacy products.

BUSINESS LOGIC TRANSITION

Cost Reduction (Stock)

Revenue Growth (Incremental)

Chart: The strategic shift in corporate AI investment objectives

The true ceiling of AI is not defined by labor savings, but by its ability to uncover business opportunities that were previously invisible.

泓君询问真实销量案例 57:40

Are there specific brand case studies? Did their sales and conversion rates truly improve following the AI integration?

嘉宾分享高端家电案例 58:00

Certainly. A leading domestic appliance brand utilized our AI VOE (Voice of Expert) analysis for its premium line—with units priced above 20,000 RMB—to capture nuanced feedback from high-end clientele. Leveraging these AI-driven insights, they refined their marketing narratives, livestreaming scripts, and customer service protocols. After eighteen months of collaboration, overall sales surged, and the average order value (AOV) increased by over 6%.

Note | This is a rare case of simultaneous growth in both volume and price: the product became more expensive, yet it sold even better.

泓君询问收费模式 59:30

That’s compelling. Did they subsequently allocate a dedicated budget for AI?

嘉宾解释收费逻辑 59:45

They did. Today, every business unit within the brand has integrated AI recognition capabilities. While the focus is currently on speech recognition, the corporate propensity to pay for AI capabilities has been firmly established.

泓君探讨中间层竞争力 60:20

As the 'middle layer' bridging foundation models and clients, what is your true competitive edge? Is it industry expertise or data?

嘉宾揭秘行业垂直模型 60:40

While we utilize foundation models, we also develop and train **our own vertical-specific models**.

Competitive Dimensions	Strategic Execution	Client Value
Data Feedback Loops	Capturing Industry-Specific Intent Data	Models with deeper industry expertise than Tongyi Qianwen
Industry Know-how	Acting as an external specialist	Integrating AI seamlessly into workflows
Business Metrics	Aligning conversion and return rates	Delivering tangible financial impact

泓君提出未来“军火竞赛”担忧 62:00

If every enterprise eventually arms itself with AI, effectively leveling the technological playing field, will the competitive advantage simply erode?

嘉宾强调“人”的核心价值 62:30

That is a profound question. I believe the core value of AI will always rest in human hands. On a macro level, AI-processed reports may appear identical from one company to the next; however, the strategic decisions made based on that data remain a true test of human insight.AI is the weaponry, but the victory still rests with the commander.

嘉宾分享对 SaaS 行业的观察 63:15

Furthermore, AI has been instrumental in reshaping the mindset of Chinese B2B enterprises toward software investment. Where hiring was once preferred over software, the realization that AI solves specific growth challenges has notably increased the willingness to pay.

Chapter 07

From 'Software Subscriptions' to 'Utility Bills': How AI is Reshaping the B2B Ledger

Fundamental Principle

A Generational Leap in Business Logic

Metrics

Data Granularity 1,000,000+

Total processing volume across voice, social media, and intent recognition.

#High-Growth Commercial Model

Step 01

Raw Data Collection

Acquiring massive, fragmented data across voice, social media, and interactive channels.

Step 02

Advanced AI Model Refinement

Intent Recognition, Attribution Analysis, and Persona Profiling

Step 03

Volume-Based Value Delivery

Scaling costs by processing volume to achieve commercial closure

Visualized by Pod-Design-System · Minimalist Paper Aesthetics

“

从功能付费到算力付费

AI 时代的商业变革在于成本的显性化。当软件背后是真实的算力支出，中国客户正逐渐接受‘按量计费’的新逻辑，这为 SaaS 行业带来了前所未有的高增速机会。

Read Insight

亓鲁阐述商业闭环逻辑 01:03:34

If the end client does not pay, we cannot sustain the service, as we bear our own underlying operational costs.This 'no payment, no service' logic aligns with the mature paradigms of the US SaaS industry.Industry clients are increasingly reaching a consensus:AI inherently incurs computational costs, and the market is beginning to embrace cost-based pricing.

Note | Editor’s Note: This marks a watershed moment for B2B business models. Traditional software is an Asset; AI services are a Utility.

泓君追问具体收费标准 01:04:12

What is the current pricing benchmark? Is it the fixed $20 or $30 monthly subscription seen overseas, or has the Chinese B2B market developed more specific structures?

亓鲁拆解语忆科技收费模型 01:04:30

Our pricing logic is anchored in data processing volume:

Profiling & Recognition:Deep processing features including audience profiling, intent recognition, and attribution analysis.
Billing Unit:For instance, a billing package might comprise 1 million voice records or 1 million social media interactions.
Cost Conversion:Estimated AI processing costs are translated into corresponding “usage package” fees.

Business Model Comparison: Traditional SaaS vs. AI SaaS

Dimension	Traditional SaaS	YuYI Tech AI Model
Core Billing Metric	Feature access / Seat count	Data processing volume (Usage)
Value Perception	Access to tools	Tangible outcomes and strategic insights
Cost Structure	Near-zero marginal cost	Computing & API: Explicit Costs

We are grateful for the commercial transformation catalyzed by AI, which has propelled Usense Technology (语忆科技) to exceptional growth within the industry.

AI has taught enterprises to pay for software as they would for a utility.

泓君总结并致谢 01:06:45

A comprehensive insight from Usense Technology (语忆科技). Our thanks to Mr. Qi Lu.

Note | Editor’s Note: This concludes our interview. The synergy of AI and marketing transcends mere efficiency; it is a fundamental redefinition of value exchange at the bedrock of business.

That concludes today’s program. If you enjoyed the session, please subscribe via the following platforms:

Audio Platforms: Apple Podcasts, Spotify, Xiaoyuzhou (小宇宙).
Video Platforms: Bilibili (B站), YouTube.

Thank you for listening. We look forward to seeing you in the next episode.

Related Episodes