Verbal Content Input Structuring - EN

Key Logic

The core message of this speech is that in the era of rapid AI development, the strongest predictor of startup success is speed of execution, and new AI technologies are dramatically accelerating startup building. Andrew Ng emphasizes that the biggest startup opportunities are at the application layer of the AI stack. He suggests boosting speed by promoting Agentic AI workflows, encouraging the building of concrete ideas, leveraging expert intuition, focusing on a single hypothesis, and adopting strategies like AI-assisted programming and rapid product feedback. He also notes that understanding AI technology itself helps entrepreneurs make faster decisions due to the exponential combinatorial effect of AI building blocks. Finally, he calls for empowering everyone to learn how to use AI tools, warns against hype and threats to the open ecosystem, and stresses the need to be socially responsible while pursuing speed.

Startup Opportunities in the AI Era & The AI Stack

• Andrew Ng points out that AI Fund, a venture studio that incubates an average of one startup per month, has firsthand experience in understanding how to use AI to accelerate development.
• Andrew Ng explains the hierarchical structure of the AI stack and emphasizes that the application layer holds the largest business opportunities.
  • The bottom layer consists of semiconductor companies.
  • Built on top are the cloud service providers.
  • Above that are the AI foundation model companies.
  • The Application Layer has the biggest opportunity:
    • Although media and social media tend to focus less on the application layer, by definition, the biggest opportunities must be there.
    • Because applications need to generate enough revenue to pay the foundation models, the cloud, and the semiconductor layers to sustain the entire ecosystem.
    • Of course, opportunities exist at every layer of the AI stack.

Boosting Efficiency with Agentic AI and its Workflows

• Andrew Ng explains the rise of Agentic AI and why it is technically exciting and creates more startup opportunities.
  • Past usage of Large Language Models: Typically involved providing a prompt to generate a continuous output.
    • This is like asking a human to write an entire article in one go without revision, which is not the best way for humans or AI to write.
  • Agentic Workflows:
    • Allow the AI system to complete tasks iteratively, for example:
      • First, write a paper outline.
      • Then, research the web for relevant pages.
      • Next, write a first draft.
      • Finally, read and critically revise the first draft.
    • "Through multiple cycles, it's slower, but it produces a much better work product."
    • In many projects at AI Fund, from processing complex compliance documents to medical diagnosis and reasoning about complex legal documents, Agentic workflows are the "key differentiator for whether or not the work works."
  • Updates to the AI Stack:
    • In the past year, a new Agentic orchestration layer has emerged, helping application developers coordinate calls to underlying technology layers (like foundation models and cloud services).
    • This orchestration layer makes building applications easier.
    • The conclusion remains that the application layer is the most valuable place in the stack, and he prefers to focus on it.

Best Practices for Startup Acceleration: Concrete Ideas, Expert Gut, Single Focus & Fast Feedback Loops

• Andrew Ng shares best practices for how startups can accelerate their speed.
  • Focus on "Concrete Ideas":
    • A concrete product idea is detailed enough for an engineer to build directly.
      • "If an idea is vague, like 'use AI to optimize healthcare assets,' it's too ambiguous. Different engineers will build completely different things, making it impossible to build quickly."
      • "Conversely, if the idea is specific, like 'develop software for hospital patients to book MR machine slots online to optimize utilization,' an engineer can build it quickly."
    • "Concreteness brings speed," and even if it turns out to be a bad idea, you'll find out quickly and can pivot.
    • Vague ideas easily garner praise ("That's a great idea") but are hard to implement. Concrete ideas can be right or wrong, but you get the result quickly.
    • Benefit: Provides a clear direction, allowing the team to move fast to validate or invalidate it.
  • Leverage "Subject Matter Expert Gut":
    • Find subject matter experts who have thought about a problem for a long time; their "gut feeling" often leads to surprisingly good and fast decisions.
    • "Although I work in AI and love data, in many startups, getting data itself is a slow decision-making mechanism."
    • "A subject matter expert with a great gut is often a better mechanism for making fast decisions."
    • For example, before founding Coursera, Andrew Ng spent years thinking about online education and talking to users, which formed his intuition for building the platform.
  • Pursue a "Single Clear Hypothesis":
    • Most successful startups pursue only one very clear hypothesis at any given time.
    • "A startup doesn't have the resources to hedge its bets and try 10 things at once."
    • If data shows an idea isn't working, you can quickly pivot to a completely different concrete idea and pursue it with the same determination.
    • "If every new piece of data causes you to pivot, it probably means your initial knowledge base was too weak."

Impact of AI on Software Engineering and Developer Productivity

• Andrew Ng discusses how AI is changing the feedback loop in software engineering.
  • Customer acceptance is the biggest risk: Many startups fail not because they can't build what they want, but because no one cares about what they build.
  • AI code assistance is enabling rapid engineering.
    • Engineering speed is increasing rapidly, and engineering costs are dropping quickly.
  • Two types of code:
    1. Quick and Dirty Prototypes: Used to test ideas.
       • He believes AI has made building prototypes at least 10 times faster, if not more.
       • The reason is that prototypes require less integration with existing software, infrastructure, and data.
       • Lower requirements for reliability, scalability, and even security: "I often tell my team to 'feel free to write insecure code.' If the software only runs on a personal laptop with no malicious attacks, that's fine." (But it must be secured and scalable before release.)
       • Startups can now systematically build 20 prototypes to discover which one works.
       • "If the cost of a proof-of-concept is low enough, it's perfectly fine if many of them don't end up in production."
    2. Maintain Production Software:
       • When writing production-quality code, AI might increase speed by 30% to 50%.
  • "Move fast and be responsible": The old "move fast and break things" has a bad reputation, but he believes we shouldn't give up on "moving fast," but rather "move fast and be responsible."
  • Evolution of AI Coding Assistants:
    • 3-4 years ago it was code auto-completion (GitHub Copilot).
    • Then came AI-enabled IDEs.
    • About 6-7 months ago, a new generation of highly Agentic coding assistants emerged (e.g., Claude CodeX using Claude 3).
    • These tools are continuously improving developer productivity; staying updated with tools is crucial.
  • The value of code is decreasing:
    • In the past, code was considered a high-value product because it was difficult to create.
    • Now, as software engineering costs decrease, the value of code also decreases.
    • "Our team once completely rebuilt a codebase three times in one month because rebuilding a codebase is not that hard anymore."
  • "Two-way Door vs. One-way Door":
    • A term by Jeff Bezos: Two-way door decisions are easily reversible with low cost, while one-way door decisions are costly or difficult to reverse.
    • Previously, choosing a tech stack or database schema was a one-way door.
    • Now, with the significant reduction in software engineering costs, many decisions once considered one-way doors are now closer to two-way doors.
    • "My team now more frequently builds on a tech stack for a week, then decides to throw away the codebase and start over with a new stack." While not always the case, the cost has indeed decreased, changing decision-making patterns.

Empowering Everyone to Learn Programming and Use AI Tools

• Andrew Ng believes now is a great time to empower everyone to build with AI.
  • In the past year, some have advised against learning to code, arguing that AI will automate it, which he considers "one of the worst pieces of career advice."
  • When better tools make software engineering easier, more people should learn it, not fewer.
  • Historical review: From punch cards to keyboards, from assembly language to high-level languages, each technological advance made programming easier and attracted more people to learn.
  • "I have a controversial view, which is I think this is the time for everyone, regardless of job function, to learn to code."
    • In his team, everyone including the CFO, head of talent, recruiters, and the front desk knows how to code, and they all perform better in their respective jobs.
  • "In the future, the most powerful people will be those who can 'tell a computer exactly what you want it to do'."
  • Learning to code (not necessarily writing all the code yourself, but guiding an AI to write it for you) is still the best way to do this.
  • Example of Midjourney:
    • A team member knowledgeable in art history can give Midjourney precise prompts (genre, color palette, artistic inspiration) to generate high-quality images.
    • This is in contrast to Andrew Ng himself, who can only say "please draw me a picture of a nice robot" and cannot achieve the same level of control.

Product Management Bottleneck and Strategies for Getting Feedback

• Andrew Ng observes that after the increase in software engineering speed, Product Management has become the bottleneck.
  • For the past year, many teams have complained that the bottleneck has shifted to product engineering and design because of the significant increase in engineering speed.
  • A past Silicon Valley rule of thumb: 1 PM for every 4-7 engineers.
  • This ratio is now changing: Some have even proposed 0.5 engineers per PM (i.e., twice as many PMs as engineers). He sees this as an indicator of future trends.
  • Rapid feedback is crucial for PMs:
    • A PM who can get feedback quickly to shape the product's direction can help the team maintain speed in the context of increased engineering velocity.
    • A portfolio of tactics for getting product feedback (from fast to slow, inaccurate to accurate):
      1. Fastest: Rely on your own gut (if you're a subject matter expert, this works surprisingly well).
      2. Slightly slower: Ask 3 friends or teammates for feedback.
      3. A bit slower: Ask 3 to 10 strangers for feedback.
         • Andrew Ng has learned to politely ask strangers in high-traffic locations like coffee shops or hotel lobbies for feedback on his products.
      4. Slower: Send a prototype to 100 testers.
      5. Slowest: A/B testing.
         • Andrew Ng notes that although Silicon Valley highly values A/B testing, it is now one of the slowest tactics on his menu.
    • The true value of A/B Testing:
      • It's not just about choosing product A or product B, but more importantly, "studying the data carefully to hone your intuition" to make higher-quality decisions faster.
      • "Updating our mental models through all this data to increase the speed at which we make product decisions."

Q&A Session

• Relationship between humans and AI tools: Develop tools or learn to use tools?
  • Andrew Ng believes AGI is overhyped and that humans will continue to do many things AI cannot for a long time.
  • The most powerful people in the future will be those "who can make a computer do exactly what you want it to do."
  • Learning how to use AI to make a computer work for you will make you more powerful than those who don't.
• The future direction of compute?
  • Some mention sending GPUs to space, nuclear-powered data centers, etc.
  • Andrew Ng considers this all overhyped.
  • "These hype narratives make certain companies look more powerful and help them with fundraising."
  • He lists several "false hype narratives" related to AI:
    • "'AI is so powerful we might accidentally cause human extinction'—ridiculous."
    • "'AI is so powerful that soon no one will have a job'—untrue."
    • "'By training one new model, we will easily wipe out thousands of startups'—untrue." (Only a few companies are affected.)
    • "'AI needs so much electricity that only nuclear power will be enough'—untrue."
  • "GPUs in space? I think our terrestrial GPUs still have a lot of room to grow."
• How should startups think about moats in an era where "anything can be disrupted"?
  • Andrew Ng believes the most important thing is "Are you building a product that users love?"
  • Other factors (like go-to-market channels, pricing, moats) are also important, but are a lower priority than building a beloved product.
  • "I find that moats are often overhyped."
  • Moats usually evolve as a product develops, rather than existing from the beginning.
  • Facing a large amount of unmet demand ("a lot of white space"), startups should focus on product needs, and the moat issue can be addressed later.
• Educating the public about AI: Is it necessary for non-technical people to understand Deep Learning?
  • Andrew Ng states that DeepLearning.AI is dedicated to empowering everyone to use AI and believes that knowledge will spread.
  • He sees two main dangers:
    1. Failing to "bring everyone along" quickly enough.
    2. The "gatekeeper" effect: Like the mobile ecosystems (Android and iOS), if a few companies control AI foundation models, it will stifle innovation.
       • He criticizes certain companies for exaggerating the dangers of AI to push for regulation (like California's SB 1047 bill), thereby blocking the release of open-source and open-weight software to become "gatekeepers."
       • "If these regulatory proposals succeed and lead to regulatory capture, where everyone needs permission from a few companies to fine-tune a model or prompt it in a certain way, that would stifle innovation."
  • Therefore, it is crucial to protect open-source and open-weight models to prevent this line of attack from succeeding.