Andrew Ng, a pioneer in AI education and entrepreneurship, shared invaluable lessons at AI Startup School in San Francisco on June 17, 2025. His talk centered around the theme of speed in building AI startups, emphasizing how new AI technologies are enabling unprecedented execution velocity. However, he also underscored the importance of responsible development and ethical considerations, providing a balanced perspective for aspiring founders.
Ng's presentation, based on his experience at AI Fund, a venture studio that co-founds startups, offered practical advice on navigating the evolving AI landscape. His insights are particularly relevant for entrepreneurs looking to capitalize on the current wave of AI innovation while remaining mindful of its potential societal impact.
The AI Stack and Agentic AI
Ng began by outlining the AI stack, comprising semiconductor companies, cloud providers, foundation model companies, and the application layer. Despite the hype surrounding the technology layers, he asserted that the most significant opportunities lie within the application layer. This is because applications are essential for generating revenue, which, in turn, supports the entire AI infrastructure.
A key trend driving innovation at the application layer is the rise of agentic AI. Ng explained that traditional LLMs are often used in a linear, "type-out-an-essay-without-backspace" manner, which limits their potential. Agentic workflows, on the other hand, enable AI systems to perform iterative tasks like outlining, researching, drafting, critiquing, and revising. This iterative process, while slower, yields significantly better results. An agentic orchestration layer has emerged to simplify the coordination of these calls to the technology layers underneath, thus making application development easier.
"Agentic workflows are really a huge difference between [AI] working versus not working."
Ng emphasized that many valuable businesses can be built by integrating existing or new workflows into these agentic systems. However, the application layer has to be the most valuable layer of the AI stack.
Best Practices for Moving Faster
The core of Ng's talk focused on how startups can achieve greater execution speed. He highlighted several best practices:
- Concrete Ideas: A concrete idea is specified in enough detail for an engineer to start building. Avoid vague notions like "using AI to optimize healthcare assets." Instead, opt for specific concepts such as "software to let hospitals let patients book MR machine slots online to optimize usage." Concreteness buys speed, and while vague ideas may garner initial praise, concrete ones allow for rapid validation or falsification.
- Subject Matter Expertise and Gut Instincts: Finding good, concrete ideas often requires deep understanding and experience. Ng shared that, before co-founding Coursera, he spent years thinking about online education. Subject matter experts who have thoroughly explored a problem can often make quick, informed decisions based on their "gut" instincts. While data is valuable, it can be slower to acquire than leveraging the intuition of an expert.
- Focused Hypotheses and Agile Pivoting: Startups should pursue one clear hypothesis at a time. Avoid hedging and trying multiple things simultaneously. Be prepared to pivot quickly if data indicates a need to change direction. Ng described this as doggedly pursuing an idea until proven wrong, then switching to a different idea with equal determination. However, frequent pivoting based on every new piece of data may suggest an insufficient initial base of knowledge.
The Built-Feedback Loop and AI Coding Assistance
Ng discussed the importance of the "built-feedback loop," involving rapid engineering and user feedback. AI coding assistants are drastically changing the speed of engineering, particularly for building prototypes. While production-quality code development may be 30-50% faster with AI, prototype creation can be 10x faster or more. Ng even advised his team to "write insecure code" for prototypes to accelerate development, emphasizing that security should be addressed before shipping to users.
"Move fast and be responsible."
He highlighted the evolution of AI coding tools, from code autocomplete to AI-enabled IDEs and highly agentic coding assistants. The rapidly evolving landscape requires continuous adaptation to stay ahead.
The Empowered Coder and Shifting Bottlenecks
Ng challenged the notion that AI automation diminishes the need for coding skills. Instead, he argued that AI makes coding easier, which means *more* people should learn to code. He even suggested that everyone, regardless of job role, should learn to code to improve productivity.
With faster engineering, the bottleneck shifts to product management. He notes that the traditional ratio of PM to engineers used to be 1:6-7. One of his teams is proposing a 2:1 ratio! This indicates the importance of focusing on rapidly getting quality user feedback. The fastest way is to consult a subject matter expert and act on the expert's gut feeling. Other sources are soliciting feedback from friends or strangers, or AB testing.
Conclusion: The Importance of AI Understanding
Ng concluded by emphasizing that understanding AI itself is a key advantage. Mature technologies like mobile development and jobs like HR and sales roles require less specialized knowledge; however, AI is still a rapidly-developing skillset. His final points underscored this importance:
- AI is not static; it is always changing.
- AI adoption makes teams quicker at execution and development
In summary, Andrew Ng's insights at AI Startup School provided a roadmap for building AI startups with speed and responsibility. His emphasis on concrete ideas, agentic workflows, rapid feedback loops, AI coding assistance, and the importance of understanding AI offers valuable guidance for entrepreneurs navigating the dynamic landscape of artificial intelligence.
Youtube Building Faster with AI
Building AI Startups with Speed and Responsibility