Written by: Techub News整理

Introduction

In July 2025, one of the pioneers in the AI field, Coursera co-founder, and renowned AI expert Andrew Ng spoke at Y Combinator's Startup School event, titled "Building Faster with AI." As the founder of AI Fund (a venture studio focused on AI, co-founding approximately one startup every month), Andrew Ng is not only observing from an investment perspective but is also personally involved in the entire process of coding, customer communication, feature design, and pricing decisions. Against the backdrop of rapidly changing AI technology, he shared core insights and strategies distilled from extensive practice, aimed at helping startups accelerate execution and enhance their probability of success.

Summary

• The success of startups is highly correlated with execution speed. In the AI era, new tools make acceleration possible.
• Pursuing "specific ideas" is a key starting point for gaining speed; vague ideas may receive praise but are difficult to execute and verify.
• AI programming assistants are increasing the speed of engineering development (especially prototype construction) tenfold or more, changing cost structures and decision-making patterns.
• The acceleration in engineering speed has made product management and user feedback new bottlenecks, requiring a set of combined strategies to quickly gather feedback and adjust directions.
• A deep understanding of emerging building blocks such as the AI technology stack and agent workflows allows teams to avoid technical dead ends and seize more opportunities for innovation combinations.

Opportunities Lie at the Application Layer, But Tools Are Evolving

Andrew Ng first shared his views on the current opportunity landscape in AI. He believes that, although semiconductor, cloud computing, and foundational model companies have attracted significant attention and hype, it is almost certain that the biggest opportunities lie at the application layer. This is because applications ultimately need to generate more revenue to support the costs of the underlying technology stack. He reminded entrepreneurs that, while there are opportunities at every layer of the technology stack, the application layer has the greatest potential due to its direct creation of value.

In the past year and a half, one of the most important technological trends in AI is the rise of agentic AI. Andrew Ng explained that previously we would typically use large language models (LLMs) in a "one-off prompt to generate complete output" manner, akin to asking a person to write an article from start to finish without drafting it first. However, the agent workflow allows AI systems to work iteratively: starting with outlines, conducting web research, drafting, self-critique, and revising. Although it may be slower, it produces work of much higher quality. In various projects involving complex compliance document extraction, medical diagnostics, and legal document reasoning, the agent workflow has become a key differentiator for "whether a project succeeds or not."

As a result, a new level of the AI technology stack has emerged in the past year: the agent orchestration layer, which helps application developers coordinate multiple calls to lower-level technologies (such as foundational models). The appearance of this layer has made building applications easier, but Andrew Ng reiterates that the application layer remains the most valuable part of the tech stack, and entrepreneurs should continue to focus on it.

The Starting Point of Speed: Pursuing "Specific Ideas"

For gaining execution speed critical to startups, Andrew Ng's first core advice is to focus solely on "specific ideas." A specific product idea should be detailed enough for engineers to build it directly.

• Negative example: "Optimize medical assets with AI." This is too vague, and different engineers will produce completely different outputs, unable to build quickly, resulting in a lack of speed.
• Positive example: "Develop a software that allows patients to schedule MRI machine slots online to optimize utilization." Although this idea might already have companies working on it, it is specific enough for engineers to quickly build a prototype to validate its viability.

Andrew Ng pointed out that many entrepreneurs can be misled because vague ideas often receive more acclaim. When you tell your friends, "Optimize medical assets with AI," everyone may agree it is a good idea. However, vague ideas are often "correct," while specifics mean they could be validated as right or wrong. For startups, quick validation (regardless of the outcome) is essential. Specific ideas provide a clear direction, allowing the team to move forward at full speed to either validate its feasibility or quickly prove its ineffectiveness and pivot.

He found that high-quality specific ideas often originate from experts in a certain field contemplating that problem over a long period. As these experts accumulate rich customer communication experiences and intuitions, their "intuition" (snap decisions) can serve as an excellent proxy for decision-making. While Andrew Ng loves data, acquiring data in many startup scenarios can be a slow decision-making process, while an expert with good intuition often serves as a better way to make rapid decisions.

Additionally, successful startups should focus on a very clear hypothesis for construction and validation at all times. Startups have limited resources and cannot hedge multiple directions simultaneously. Therefore, choosing a specific idea, committing fully, and once data indicates a dead end, immediately pivoting to another completely different specific idea is essential. This is the norm at AI Fund: pursuing one thing with steadfast determination until the world proves us wrong, and then pivoting to the next thing with the same level of resolve and dedication.

Andrew Ng also cautioned that if every new piece of data leads you to change direction ("every customer interview completely changes your mind"), it may indicate that your initial knowledge foundation is too weak, and you are not sufficiently knowledgeable about the domain. At this point, bringing in or becoming someone with more long-term thinking on the subject can help you get back on track more quickly.

The Engineering Speed Revolution: From "One-Way Street" to "Bi-Directional Gate"

Once the specific idea is established, the next acceleration step is engineering construction. Andrew Ng highlighted how AI programming assistants fundamentally changed the speed and cost of the engineering build-feedback loop.

He pointed out that for many application-based startups, the biggest risk is "customer acceptance," not technical feasibility. The traditional loop is: build software (engineering task) → gather user feedback (product management task) → adjust build direction → rebuild. The advent of AI programming assistants has dramatically increased engineering speed and significantly decreased costs.

Andrew Ng categorizes the software he writes into two types:

• Fast and rough prototypes: used to test ideas (e.g., new customer service chatbots, AI handling legal documents).
• Production-grade, maintainable software: maintaining legacy systems or large production-ready codebases.

For production-grade code, using AI systems may increase efficiency by 30% to 50%. However, for building quick prototypes, AI brings not a 50% improvement but achieves speedups of ten times or more. The reason lies in that prototypes do not need to integrate with legacy software infrastructure, and the requirements for reliability, scalability, and security are much lower.

He would even tell the team: "Go ahead, write unsafe code." Because if the software only runs on your laptop and you don't intend to maliciously attack your computer, unsafe code is acceptable. Of course, once the software seems viable and is delivered to others, its safety and scalability must be ensured. This shift in mindset allows startups to systematically explore innovative directions by building 20 prototypes, even if many proofs of concept (POCs) ultimately do not enter production, the costs remain low enough to be manageable.

He reflected on the evolution of AI programming assistants: three to four years ago, it was code auto-completion (like GitHub Copilot), followed by AI-enhanced IDEs like Cursor and Windsurf, and about six to seven months ago, we started seeing a new generation of highly intelligent coding assistants (like the ones he heavily uses, o3 and Claude Code). The iteration of tools is rapid, and the efficiency gap between keeping up with the latest tools and lagging behind half a generation is enormous.

A profound cognitive shift is that the value of code as a "valuable output" is declining. Because the cost of software engineering has decreased, rewriting an entire codebase and selecting new data patterns are no longer that hard. Andrew Ng borrowed terms from Jeff Bezos to explain that choosing a tech stack architecture used to be a "one-way street" decision, once determined it was hard to turn back. Now, it is becoming more like a "bi-directional gate." His team sometimes changes its mind after a week, discarding a codebase and rewriting it on a new tech stack from scratch. While there are still costs involved, this possibility has opened up.

Furthermore, Andrew Ng提出了一个颇具争议的观点：现在是让每个岗位的人都学习编程的时候了. In his team, the CFO, talent director, recruiters, and front desk staff all understand programming, and this has led to better performance across all functions. He believes one of the most important skills for the future is "to tell the computer exactly what you want and let it do it for you." Learning programming (even if not directly writing code, but guiding AI programming) seems to be the best way to master this skill long-term. He pointed to a team member with art history knowledge being able to precisely control Midjourney to generate high-quality backgrounds as an example of how a deep understanding of tools leads to better output.

New Bottlenecks and Rapid Feedback Strategies

As engineering speed rapidly increases, product management work — obtaining user feedback and deciding what to build — is increasingly becoming the new bottleneck. Andrew Ng observed that the traditional ratio of product managers to engineers common in Silicon Valley (such as 1:6 or 1:7) is undergoing a significant transformation.

He revealed that, in July 2025, one of his teams proposed an unprecedented ratio during project labor planning: 1 product manager to 0.5 engineers (i.e., the number of PMs is twice that of engineers). This dynamic is interesting, and although he is uncertain if it is a good idea, it undoubtedly marks a trend in the world. He observed that PMs who understand programming and have product intuition, or engineers with product intuition, tend to perform better.

To address this bottleneck, Andrew Ng shared a set of strategies for obtaining rapid product feedback, arranged from fastest to slowest and from low accuracy to high accuracy:

1. Fastest but intuition-dependent: personally experience the product and rely on intuition (this often works surprisingly well if you are a domain expert).
2. Slightly slower: ask three friends or teammates.
3. A bit slower: seek feedback from three to ten strangers. He emphasized one important skill: learning to politely invite strangers to try the product and provide feedback in high-traffic areas like coffee shops and hotel lobbies. He admitted he used to do this often, but now feels slightly awkward as people recognize him; however, it’s a valuable skill for entrepreneurs.
4. Slower: send the prototype to 100 testers (if you have a user base).
5. One of the slowest strategies: A/B testing. Although he conducts a lot of A/B tests, it has become one of the slowest tactics in his toolbox, as it requires time and a sufficient user base.

Andrew Ng particularly pointed out that when using strategies other than the first one (intuition), the key is not just making decisions based on the data. His team will sit down to carefully examine the data, with the goal of "honing intuition" — updating the psychological model about the users, thereby enhancing future abilities to make high-quality product decisions quickly using "intuition." This is crucial.

Understanding AI Technology for Faster and More Accurate Outcomes

Andrew Ng lastly emphasized the importance of deeply understanding how AI technology itself grants startups a speed advantage. For mature technologies (like mobile applications) or established roles (like sales and marketing), relevant knowledge has been widely disseminated, and finding experts is relatively easy. But AI is an emerging technology, and knowledge on how to effectively leverage AI has yet to be mainstreamed.

Therefore, teams that truly understand AI hold an advantage over those that do not. For instance, technical decisions like “What accuracy can customer service chatbots achieve?” “Should we use prompt engineering or fine-tuning, or a certain workflow?” “How can we achieve low-latency voice output?” — when chosen correctly, may resolve issues in just a few days; when chosen incorrectly, may waste three months stuck in a dead end. Theoretically, a binary (right/wrong) technical decision can yield at most double the acceleration, but in practice, choosing the wrong path may lead to spending ten times the time chasing the wrong direction. Thus, having the right technological judgment is crucial.

Another reason is that in the past two years, a plethora of excellent generative AI tools or building blocks have emerged: prompt engineering, workflows, evaluation metrics, guardrails, retrieval-augmented generation, voice interaction, asynchronous programming, ETL, embeddings, fine-tuning, graph databases, model integration, and more. Understanding these building blocks is akin to having more colors of LEGO bricks. Each time you master a new building block, the number of things you can create in combination grows exponentially, allowing for the creation of software that no one could build just a year ago, generating a wealth of new opportunities for startups.

Andrew Ng used the courses from DeepLearning.AI as an example, stating that learning these courses is like acquiring new "bricks," which allows for building unprecedented applications in much richer combinations.

Highlights from the Q&A Session: AGI, Hype, Education, and Responsibility

In the subsequent Q&A session, Andrew Ng elaborated on his views.

On AGI and Human Positioning: He believes AGI has been overhyped; for a long time, there will still be many things that humans can do that AI cannot. The most powerful people in the future will be those who can instruct computers to execute their commands precisely. Those who master and utilize AI tools will be far more powerful than those who do not.

On AI Hype and Dangerous Narratives: He bluntly stated that over the past two years, some companies have hyped certain narratives for PR, fundraising, etc., without factual verification. For example, “AI could lead to human extinction,” “AI is so powerful that soon no one will have jobs,” “training new models will easily wipe out thousands of startups,” “AI consumes enormous energy and can only be powered by nuclear energy,” in his view, are all exaggerated dangerous narratives. He emphasized that AI itself is not inherently safe or unsafe; safety depends on how we apply it. He prefers to discuss "responsible AI" rather than the generic notion of "AI safety." He also criticized the use of these hyped narratives to attack open-source software, arguing that it stifles innovation.

On Startup Moats: When asked how to think about business in an age where AI can quickly replicate products, he stressed that the early focus for startups should be on "whether you are building a product that users truly love." It is premature to discuss moats and channels before addressing this question. Moats often evolve gradually with product success. Currently, the number of opportunities that "can be built but have not yet been built" seems to far exceed the number of people with building skills, especially at the application level, where there are many blank areas.

On Agent Token Costs: He suggested that developers should largely not worry about token costs at first. Only a few sufficiently lucky startups might have usage large enough to make token costs a concern. For teams encountering such situations, there are usually engineering solutions (like prompt engineering, fine-tuning, optimization) to lower costs. He advised software architecture should facilitate switching between different building module providers to maintain flexibility.

On the Future of EdTech: He believes education will trend towards high personalization, but many experiments are currently underway and the final form is not yet clear. The application of AI in education still mirrors the process of agent workflows and is not yet mature.

On Balancing Development and Responsibility: He offered simple yet powerful advice: examine your conscience; if you fundamentally do not believe what you are building will improve people's lives, then do not do it. The AI Fund has terminated several projects based on ethical considerations rather than financial reasons. Moreover, he calls for more people to master AI skills to prevent increasing inequality due to skills gaps. He remains cautious of certain regulatory proposals (like California's SB-1047) that could be used to stifle open-source software, creating a few "gatekeeper" companies that hinder innovation and knowledge dissemination. He believes protecting open-source and ensuring the freedom of innovation is crucial for spreading AI knowledge and skills.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。