On June 17, 2025, Andrew Ng delivered a keynote address at the "AI Startup School" in Silicon Valley. As a leading scholar and practitioner in AI, Ng has deeply cultivated the implementation of AI technology and entrepreneurial incubation, forming a comprehensive ideological system grounded in rich experience.

His speech centered on the value distribution within the AI technology stack, the logic of entrepreneurial decision-making, the evolution of organizational capabilities, and insights into the core of AI development. Here are his key viewpoints:

1. AI Technology Stack: Application Layer at the Core, Agentic Workflow as the Vital Bridge

Ng divided the AI technology stack into distinct layers: semiconductor companies at the base, hyperscale cloud service providers above them, and AI foundational model companies at the apex. While media attention often focuses on underlying technologies, Ng clearly stated from a commercial perspective that "the true opportunity for AI businesses lies in the application layer." This assertion stems from the fundamental logic that only when the application layer generates substantial revenue can it sustain the development of foundational models, cloud computing, and semiconductor technologies. He emphasized, "Opportunities exist across all technological levels, but the application layer stands as the nexus connecting technology and the market, boasting the highest value density."

Ng particularly highlighted the revolutionary significance of agentic workflow: "In medical diagnosis and legal document projects AI Fund has participated in, the adoption of agentic workflow often determines project success." The traditional linear AI interaction model (input prompts → obtain results) is being disrupted. Agents, through the cycle of "outline → research → writing → evaluation → revision," though slightly slower than linear output, significantly enhance result quality. This workflow innovation introduces a new layer in the technology stack—the Agentic Orchestration Layer, serving as a crucial bridge between models and applications.

2. Focus on Specific Product Ideas: Efficient Validation and Rapid Iteration as the Key to AI Era Entrepreneurial Success

Focus solely on specific product ideas. Here, "specific" refers to the extent to which engineers can immediately commence development based on a clear requirements description. This is a principle Ng consistently upholds at AI Fund. He illustrated this with a case comparison: "Optimizing medical resources with AI" is a vague idea, leading engineers to develop disparate products; whereas "Developing software that enables patients to book MRI appointments online" is a concrete plan engineers can immediately program, accelerating development.

The value of specific ideas manifests in three dimensions: first, a clear direction enables teams to proceed at full speed; second, conclusions can be swiftly drawn regardless of validation results; third, excellent specific ideas often originate from domain experts' long-term thinking and deep problem understanding.

Taking Coursera's establishment as an example, Ng had researched the online education field for years, contemplating how to construct an educational technology platform that truly addresses issues. After extensive deliberation, he realized that experts with years of field experience often make high-quality decisions intuitively. He reminded entrepreneurs, "If you change direction after every user interaction, it indicates you haven't formulated a high-quality specific idea. At this juncture, you need to introduce domain experts to guide direction."

Successful startups should prioritize validating clear hypotheses. With limited resources, they must focus on a single direction and pivot swiftly if it proves infeasible. The primary development risk lies in market acceptance, and AI programming assistance tools are transforming the traditional feedback loop.

Software development encompasses rapidly building prototypes and maintaining mature codebases. The former is for idea validation, where AI can boost efficiency over tenfold with minimal code reliability requirements. The latter can tolerate imperfect maintenance, with AI improving efficiency by 30-50%.

Today, startups can screen directions by constructing numerous prototypes. Due to low validation costs, failed prototypes are acceptable. Concurrently, AI programming tools like GitHub Copilot, Cursor, OpenAI Codex, etc., continually enhance development efficiency. The intergenerational tool gap significantly impacts code value attributes, reducing software engineering costs and facilitating codebase refactoring.

3. Technological Decision-Making and Programming Ability: From "One-Way Door" to Full-Stack Programming Necessity

In Bezos' theory of "One-Way Door Decisions" (hard to reverse) and "Two-Way Door Decisions" (easily changed), past technology stack and software architecture choices belonged to one-way doors, difficult to alter. However, due to factors like AI, although they're not fully two-way doors, changes to technology stacks and codebases have become easier, even allowing rewrites.

Simultaneously, despite AI's ability to write code, understanding programming remains vital. Historically, programming tool simplification has expanded the developer community. Today, it's more crucial for every position to learn programming. Full-stack programming ability within a team can boost performance, such as team members controlling AI-generated images through precise prompts. The core lies in learning to articulate needs clearly to computers. Guiding AI to write code will remain a highly effective tool for the foreseeable future.

4. Organizational Capability Upgrades: Product Management Transformation and Efficient Feedback Mechanisms

The leap in engineering efficiency necessitates organizational capability upgrades. Ng has observed a significant trend: product management is gradually becoming a bottleneck—the past model of "1 product manager interfacing with 6-7 engineers" has been disrupted, with some teams even adopting configurations of "2 product managers interfacing with 1 engineer." This isn't a resource misallocation but because after AI tools boost engineer efficiency, product design and engineering management speeds can't keep pace with technological realization.

In an environment where engineering development speed accelerates, product managers proficient in programming or engineers with a product mindset excel. Startup leaders must establish mechanisms for swiftly obtaining feedback.

Ng summarized a tactical system for product feedback, ranging from fast to slow and rough to precise: the fastest is intuitive judgment after domain experts personally experience it; slightly slower is seeking feedback from three to five friends or colleagues after trying it; still slower is inviting three to ten strangers to try it and collect opinions; slower yet is sending prototypes to 100 test users; and the slowest but most precise is A/B testing. Except for the first method, decisions can't be based solely on surface data, especially for A/B testing, which requires deep analysis of poor functional performance reasons and improving product intuition through in-depth data analysis. Through such deep reflection, all data can update mental models, enhancing rapid decision-making quality.

5. Team Competitiveness and Acceleration Laws: Efficiency, Feedback, and Technological Sensitivity

Understanding AI technology is crucial for enhancing work efficiency. As AI is an emerging technology with few mastering its essence, teams proficient in AI have a competitive edge. In technological decision-making, such as choosing technology for customer service chatbots, wrong choices can result in tenfold efficiency losses, making correct judgments vital for startups. Continuously tracking AI advancements is beneficial. Combining numerous generative AI tools and modules can create new applications, fostering creativity akin to Lego blocks.

Entrepreneurial success strongly correlates with team execution speed. Acceleration laws include focusing on specific and feasible ideas, accelerating decision-making, leveraging AI programming assistance tools, establishing efficient user feedback mechanisms, and consistently tracking technological trends.

6. Rational Understanding of AI Development: Value, Risks, and Social Responsibility

Ng presented clear viewpoints on many AI development key issues. Regarding humans and AI, he believes Artificial General Intelligence (AGI) is overhyped. For the foreseeable future, humans will retain unique values AI can't replace. Individuals proficient in AI tools and adept at collaborating with AI will be more competitive and needn't fear replacement.

He criticized many AI field exaggerations, such as claims that "AI will lead to human extinction," "replace all jobs," or "require nuclear energy data centers or space GPUs," lacking technical basis. In reality, AI is creating new jobs and transforming existing ones, with substantial optimization potential in ground-based computing facilities.

Regarding AI's essence and entrepreneurial logic, Ng emphasized AI is a tool like electricity, its safety depending on usage. More attention should be paid to "responsible AI." He opposes exaggerating extreme laboratory cases into sensational stories, especially using them to attack open-source software. He warns against technological monopolistic behavior under the guise of "security" and advocates jointly maintaining a free and open innovation ecosystem.

For entrepreneurs, the core is creating products users genuinely love, first solving the "product-market fit" problem. Currently, the application layer abounds with untapped opportunities, so there's no need to excessively worry about models or functions being swiftly replicated.

In AI tools and specific domain applications, Ng noted agentic workflow can already integrate various technological modules like prompt engineering and retrieval-augmented generation. Developers needn't excessively worry about token costs in the initial stage. He suggests considering technological module replaceability in architectural design and maintaining flexibility in technological choices to ensure rapid iteration when adding more functions.

In education, future education will evolve toward high personalization, but this is a gradual process. The claim that "AGI will radically transform education" is exaggerated, necessitating continuous exploration of educational workflow and AI agentic workflow integration.

Regarding social impact and knowledge popularization, Ng proposed developers should uphold the principle of "ensuring products make people's lives better." AI Fund has halted several potentially negative impact projects while striving to make AI benefits accessible to all.

He believes it's crucial for the public to understand deep learning. Knowledge popularization should keep pace with technological development, and vigilance is needed against companies establishing technological monopolies by exaggerating AI risks (like California's SB-1047 bill). Protecting open-source software is essential to avoid technological inequality.

END