Andrew Ng on the AI era: Validation speed is the competitiveness of enterprises, and there are still numerous opportunities in the application layer.

On June 17, 2025, Andrew Ng gave a speech at the "Artificial Intelligence Startup School" in Silicon Valley. As an authoritative scholar and practitioner in the field of AI, Andrew Ng has been deeply involved in the implementation of AI technology and the incubation of startups. With his rich experience, he has formed a theoretical system that combines in - depth theory with practical guidance.

During this speech, he mainly focused on topics such as the value distribution of the AI technology stack, the logic of startup decision - making, the evolution of organizational capabilities, and insights into the essence of AI development. The following are his core viewpoints:

1. AI Technology Stack: The Application Layer is the Business Core, and the Agent Workflow is the Key Bridge

Andrew Ng divides the AI technology stack into several clear levels: semiconductor companies at the bottom, large - scale cloud service providers above them, and AI foundation model companies further up. Although the media spotlight often focuses on underlying technologies, from a business perspective, he clearly points out that "the real business opportunity in AI lies in the application layer." This judgment is based on a simple yet profound logic: only when the application layer generates sufficient revenue can it further support the development of underlying foundation models, cloud computing, and semiconductor technologies. Andrew Ng emphasizes: "There are opportunities at each technology level, but the application layer is the hub connecting technology and the market, with the highest value density."

He especially emphasizes the revolutionary significance of the agent workflow: "In projects such as medical diagnosis and legal document processing participated by AI Fund, whether to adopt the agent workflow often determines the success or failure of the project." The traditional linear AI interaction model (input prompt → get results) is being disrupted. Agents use a cycle of "outline - research - write - evaluate - revise", which, although slightly slower than linear output, significantly improves the quality of results. This workflow innovation is giving rise to a new level in the technology stack - the Agentic Orchestration Layer, which serves as a key bridge between models and applications.

2. Focusing on Specific Product Ideas, Efficiently Validating and Rapidly Iterating are the Key Paths to Startup Success in the AI Era, Avoiding Premature Optimization and Resource Waste

Only focus on specific product ideas. Here, "specific" means that engineers can start development immediately based on a clear requirement description. This is a principle Andrew Ng has always adhered to in AI Fund. He uses two cases for comparison: "Optimizing medical resources with AI" is a vague idea, and engineers may develop completely different products; while "developing software for patients to make online appointments for MRI equipment" is a specific plan, and engineers can start programming immediately and quickly advance the development.

The value of specific ideas is reflected in three dimensions: first, a clear direction allows the team to fully advance the development work; second, whether the validation result is successful or not, a conclusion can be quickly reached; third, excellent specific ideas usually come from the long - term thinking and in - depth understanding of a problem by domain experts.

Andrew Ng takes the establishment of Coursera as an example. He studied the online education field for several years and repeatedly pondered how to build an educational technology platform that can truly solve problems. After long - term thinking, he realized that experts in a certain field for many years can often make high - quality decisions quickly based on intuition. He reminds entrepreneurs: "If you change direction after every user communication, it means you haven't formed a high - quality specific idea. At this time, you need to introduce domain experts to guide the direction."

Successful startups should focus on validating clear hypotheses. When resources are limited, they need to focus on a single direction. If it's not feasible, quickly change direction. The biggest risk in development lies in market acceptance, and AI programming assistance tools are changing the traditional feedback loop.

Software development is divided into quickly building prototypes and maintaining mature codebases. The former is used to validate ideas, and AI can improve efficiency by more than 10 times, with low requirements for code reliability. The latter can tolerate imperfect maintenance of mature codebases, and AI can improve efficiency by 30 - 50%.

Nowadays, startups can screen directions by building a large number of prototypes. Since the validation cost is low, failed prototypes are acceptable. At the same time, AI programming tools such as GitHub Copilot, Cursor, and OpenAI o3 continuously improve development efficiency. The generational gap between tools has a significant impact, the value attribute of code changes, the cost of software engineering decreases, and it becomes easier to refactor codebases.

3. Technical Decision - Making and Programming Ability: The Necessity of Moving from "One - Way Doors" to Company - Wide Programming

In Jeff Bezos' theory of "one - way door decisions" (difficult to reverse) and "two - way door decisions" (easily changeable), in the past, the selection of technology stacks and software architectures was a one - way door and difficult to change. Now, due to factors such as AI, although it's not completely a two - way door, changes to technology stacks and codebases have become easier, and they can even be rewritten.

Meanwhile, even if AI can write code, it's still important to understand programming. Historically, the simplification of programming tools has actually expanded the developer community. Nowadays, every position should learn programming. When all team members have programming abilities, team performance can be improved. For example, team members can control AI to generate images through precise prompts. The core is to learn to clearly express requirements to the computer. Guiding AI to write code will be the most effective tool for a long time in the future.

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

The leap in engineering efficiency is forcing an upgrade of organizational capabilities. Andrew Ng has observed a significant trend: product management has gradually become a bottleneck - the past model of "one product manager for 6 - 7 engineers" has been disrupted, and in some teams, there is even a configuration of "two product managers for one engineer." This is not a misallocation of resources but because after AI tools have improved engineer efficiency, the speed of product design and engineering management cannot keep up with the pace of technology implementation.

In an environment where engineering development speed is accelerating, product managers who understand programming or engineers with product thinking perform better. Startup leaders need to establish a mechanism for quickly obtaining feedback.

Andrew Ng has summarized a tactical system for product feedback from fast to slow and from rough to precise: the fastest is for domain experts to make intuitive judgments after personal experience; slightly slower is to invite three to five friends or colleagues to try and provide feedback; even slower is to invite three to ten strangers to try and collect opinions; relatively slow is to send prototypes to 100 test users; the slowest but most accurate is A/B testing. Meanwhile, except for the first method, decisions should not be made based solely on surface data. Especially in A/B testing, the reasons for poor function performance need to be deeply analyzed, and product intuition should be improved through in - depth data analysis. Through such in - depth reflection, all data can be used to update the mental model to improve the quality of rapid decision - making.

5. Team Competitiveness and Acceleration Rules: Efficiency, Feedback, and Technical Sensitivity

Understanding AI technology is crucial for improving work efficiency. Since AI is an emerging technology, there are few people who master its essence. Teams that understand AI have a competitive advantage. In technical decision - making, such as the technology selection when building a customer service chatbot, a wrong choice may lead to a 10 - fold loss in efficiency. Correct judgment is crucial for startups. Continuously following the latest developments in AI is beneficial. The combination of numerous generative AI tools and modules can create new applications, bringing rich creativity like Lego bricks.

Startup success is strongly related to the team's execution speed. The acceleration rules include focusing on specific and feasible ideas, increasing decision - making speed, using AI programming assistance tools, establishing an efficient user feedback mechanism, and continuously tracking technological trends.

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

Andrew Ng has given clear viewpoints on many key issues in AI development. Regarding the relationship between humans and AI, he believes that Artificial General Intelligence (AGI) has been over - hyped. For a long time in the future, humans will still have unique values that AI cannot replace. People who master AI tools and are proficient in collaborating with AI will be more competitive, and there is no need to worry about being replaced.

Meanwhile, he criticizes many exaggerated claims in the AI field, such as "AI will lead to human extinction," "AI will replace all jobs," "Nuclear - powered data centers or space GPUs are required." These claims lack technical basis. In fact, AI is creating new jobs and changing the nature of existing jobs, and there is still a huge room for optimization in ground - based computing facilities.

Regarding the essence of AI and the logic of startups, Andrew Ng emphasizes that AI is a tool similar to electricity, and its safety depends on the way it is used. More attention should be paid to "responsible AI." He opposes exaggerating extreme laboratory cases into sensational stories, especially opposes using this to attack open - source software, and is vigilant against technology monopoly behavior in the name of "safety" to jointly maintain a free and open innovation ecosystem.

For entrepreneurs, the core is to create products that users truly love and first solve the problem of "product - market fit." Currently, there are a large number of blank areas and undeveloped opportunities in the application layer, and there is no need to overly worry about the rapid replication of models or functions.

In terms of AI tools and specific domain applications, Andrew Ng points out that the agent workflow can already integrate multiple technology modules, such as prompt engineering and retrieval - augmented generation. Developers don't need to overly worry about token costs in the initial stage. He suggests considering the replaceability of technology modules during architecture design to maintain flexibility in technology selection, ensuring rapid iteration when adding more functions.

In the education field, future education will develop towards high - level personalization, but this is a gradual process. The claim that "AGI will completely change education" is exaggerated, and continuous exploration of the combination of educational workflows and AI agent workflows is needed.

Regarding social impact and knowledge popularization, Andrew Ng proposes that developers should adhere to the principle of "ensuring that products make people's lives better." AI Fund has stopped several projects that may have negative impacts and should also ensure that the benefits of AI are shared by everyone.

He believes that it is crucial for the general public to understand deep learning. Knowledge popularization should keep up with technological development. At the same time, be vigilant against some enterprises establishing technology monopolies by exaggerating AI risks (such as the California SB - 1047 bill) and protect open - source software to avoid technological inequality.

This article is from the WeChat official account "Big Bowl Thinking," written by Big Bowl and republished by 36Kr with permission.