This lecture features Professor Andrew Ng and Google veteran AI expert Laurence Moroney discussing the pace of AI advancement and the evolving job market as of 2025. Andrew Ng emphasizes that the ability to "decide what to build" (PM skills) has become crucial as AI coding tools accelerate software development, while Laurence Moroney offers in-depth insights on strategies for surviving the post-hype job market, managing technical debt in the era of "Vibe Coding," and the rise of "Small AI" (on-device/small models) as a future trend.

1. Andrew Ng: Why Now Is the Best Time to Start an AI Career

The lecture opens with Professor Andrew Ng sharing his perspective on the current pace of AI advancement. Despite recent concerns on social media about AI progress slowing down, Ng disagrees. According to research from the MER meter institute, the complexity of tasks AI can perform is increasing at a remarkable rate.

When measuring the complexity of tasks AI can handle based on the time it takes humans to perform them, that capability doubles every 7 months. Notably, for AI coding ability specifically, the doubling time is even shorter -- around 70 days.

Thanks to this progress, he emphasizes that we are in a "golden age" -- the best time in history to build software. Just one year ago, even the world's top experts would have struggled to build the powerful software that anyone can now create using AI building blocks (LLMs, RAG, voice AI, etc.). The advancement of AI coding tools in particular is dramatically boosting productivity.

The pace of advancement in AI coding tools is truly incredible. There's a huge productivity difference between using the latest generation of tools versus tools that are even half a generation old. This means you might need to switch coding tools every 6 months, or possibly even every 3 months.

The Product Management Bottleneck and the Changing Role of Engineers

As coding speed increases, an interesting phenomenon has emerged. While implementing software specs into code used to be the bottleneck, now "deciding what to build" or "writing clear specs" has become the bottleneck.

Professor Ng explains that the ratio of engineers to product managers (PMs) in Silicon Valley is changing. In the past, one PM covered 4-8 engineers, but now the ratio is approaching 1:1. Furthermore, the fastest execution speed comes when engineers themselves communicate with users and decide the product direction.

Engineers who can listen to user feedback, form deep empathy, and decide "what to build" -- they are the fastest-moving people in Silicon Valley today. (...) If you can write code while simultaneously talking to users and envisioning the product, your execution speed will be much faster.

Advice for Success: Peers and Effort

The biggest predictor of career success is "who you surround yourself with." The greatest benefit of an environment like Stanford is this "connective tissue." Sharing cutting-edge information with exceptional peers and stimulating each other is crucial. Professor Ng also adds some candid, potentially controversial advice.

Some people think it's politically incorrect to encourage working hard. But I want to encourage you to work hard. (...) If you're in an environment where you can work hard, I hope you'll choose to fire up an agentic coder and build something on the weekend rather than mindlessly watching TV.

2. Laurence Moroney: The Reality of the 2025 Job Market and Interview Tips

Taking the mic next, Laurence Moroney -- a veteran who has worked at Google, Microsoft, and other companies, and a bestselling author -- shares realistic insights about the current job market. He first conveys an important lesson about interview attitude, sharing the case of a highly skilled coder who kept getting rejected because he took advice to "stand your ground" too far and came across as aggressive.

Just as companies evaluate you, you should also evaluate the people you'd be working with. But conversely, companies are also evaluating whether you're someone they'd want to work alongside. (...) It's fine to advocate for yourself, but you shouldn't become a jerk in the process.

The Changing AI Job Market (2021-2025)

Laurence summarizes the market changes over the past few years:

1. 2021-2022 (Pandemic): General industry slowdown.
2. 2022-2023 (AI Boom & Over-hiring): Companies hired indiscriminately amid the AI frenzy. It was a time when simply having "AI" on a resume was enough to get hired without verification.
3. 2024-2025 (The Great Awakening & Correction): Companies realized that many of those over-hired were unqualified and became much more cautious.

While junior hiring has declined and competition appears fierce, he emphasizes that there are still plenty of opportunities for those with the "right mindset."

3. Three Pillars for Success and the Age of "Vibe Coding"

Laurence presents three core elements for success in the current market:

1. Understanding in Depth: Not just academic depth, but the insight to distinguish "noise" from "signal" amid trends.
2. Business Focus: You must prove yourself through output, not just hard work.

   Don't produce output aligned with the job you currently have -- produce output aligned with the job you want.

3. Bias towards Delivery: Ideas are cheap. What matters is execution.

Four Workplace Realities

In today's AI industry, "useful" and "production-ready" matter more than "cool."

1. Business focus is non-negotiable: The atmosphere has shifted back to prioritizing business outcomes over internal activism.
2. Risk mitigation: Understanding and reducing the risks associated with AI adoption is a core competency.
3. Evolved responsibility: Beyond simple social slogans, practical responsibility that doesn't damage the company's reputation is required.
4. Learning from mistakes: Mistakes are inevitable.

Vibe Coding and Technical Debt

Generating code with generative AI (known as "Vibe Coding") is powerful, but it simultaneously creates technical debt. Laurence compares this to a mortgage (good debt) versus high-interest credit card debt (bad debt).

A good financier manages debt to become wealthy, and a good coder manages technical debt to become wealthy. (...) When you generate code with AI, you need to think about whether it's manageable debt.

How to avoid bad technical debt:

• Set clear goals and verify whether they've been achieved.
• Evaluate whether it has business value. (e.g., simply building a cool website is meaningless)
• Human Understanding: Documenting your code and cleaning up variable names so others can understand it is the most important thing.

4. Becoming a "Trusted Advisor" Amid the Hype

Social media rewards "engagement" over accuracy, so it's flooded with hype. Laurence says you need to become a "Trusted Advisor" who can find signal amid the noise.

The Right Approach to Agentic AI

When a company said "we want to adopt agents," Laurence asked "Why?" instead of diving into technical details. The company's real need turned out to be "improving sales team efficiency." He applied an agentic workflow as follows:

1. Understand Intent: Use an LLM to understand what the user is trying to accomplish.
2. Planning: Define available tools (web search, etc.) and create a plan.
3. Use Tools: Execute the plan using the tools to get results.
4. Reflect: Verify whether the results match the intent.

Don't get swept up in technical hype -- make the problem mundane and grasp its essence. Text-to-video generation technology is ultimately just "predicting changes in consecutive frames." When you understand the essence like this, you can offer better solutions as an expert.

5. Future Trends: "Small AI" and the Bifurcation of Technology

Laurence predicts that the AI market will split into two major tracks within the next 5 years:

• Big AI: Large models pursuing AGI (OpenAI, Google, Anthropic, etc.). Externally hosted.
• Small AI: Small models self-hosted by companies or individuals (Open Weights).

Laurence strongly argues that Small AI in particular will be the next major wave. Organizations like Hollywood studios and law firms won't send their data to external LLMs due to IP (intellectual property) leakage concerns. Instead, they'll build improved small models internally (self-hosted) and fine-tune them.

Today's 7B (7 billion parameter) model is as smart as yesterday's 50B model. (...) To solve privacy, latency, and cost issues, AI will increasingly move toward running on edge devices (smartphones, etc.) and CPUs.

Beyond Artificial Intelligence to Artificial Understanding

Finally, Laurence explains that AI is moving beyond simply mimicking intelligence to a stage of "artificial understanding" -- deeply grasping the user's intent. When converting a photo of his son playing ice hockey into video, a simple prompt produced irrelevant results (packed stadiums, etc.), but when an agentic workflow was used to understand the "intent" and "context" (practice rink, emotional arc, etc.), the results were far superior.

6. Conclusion and Q&A Summary

The lecture concluded with a Q&A session with students.

• Technical diversity: While some companies like NVIDIA require deep expertise in specific technologies, going all-in on a single technology is risky. Since technology changes, you need to diversify your skills.
• AI in research: Actively leverage AI tools in research, but never blindly trust the results -- always verify. Laurence shared a case where he introduced Google Colab to a researcher who lacked GPUs, dramatically accelerating their research.
• Social impact: AI is just a tool. Always assume good intent, but maintain a posture prepared for the possibility of malicious use.

In closing This lecture went beyond purely technical content to emphasize the attitude and business mindset that engineers need in 2025. Along with a warning that the AI bubble could burst, it conveys the message that "talent who sticks to fundamentals, creates business value, and can distinguish signal from noise" will survive in any situation. The advice to watch the new frontier of "Small AI" and "on-device AI" in particular serves as an important compass for those preparing their careers.