Stanford CS230 | Autumn 2025 | Lecture 9: Career Advice in AI

Key Concepts

• Rapid AI Advancement: AI capabilities, particularly in coding, are progressing exponentially (doubling every 70 days for coding tasks, 7 months for general task complexity).
• Shifting Skillsets: The value is moving from pure coding ability to product management, defining what to build, and effectively leveraging AI tools. The rise of the “Engineer-PM” is crucial.
• Responsible AI Evolution: Responsible AI is maturing beyond abstract ethics to address concrete business risks and learn from failures (e.g., Gemini image generation bias).
• Technical Debt Management: Understanding and managing technical debt is critical when utilizing AI-powered code generation (“vibe coding”).
• Agentic AI Framework: A four-step process (Understand Intent, Planning, Action, Reflection) provides a structured approach to utilizing agentic AI.
• AI Landscape Bifurcation: The AI industry is diverging into “Big AI” (LLMs like Gemini) and “Small AI” (open-weight, self-hostable models).
• AI’s Dual Potential: AI can be a powerful force for both social good and inequality, depending on its implementation.

The Current AI Landscape & Career Shifts

The current moment is considered a “golden age” for building with and in AI. Despite recent skepticism, AI task complexity is doubling approximately every seven months, with AI coding progressing even faster (roughly every 70 days). This rapid advancement is fueled by powerful building blocks like Large Language Models (LLMs), Retrieval Augmented Generation (RAG), Voice AI, and Deep Learning. LLMs are surprisingly capable of assisting with deep learning implementation, such as implementing a Transformer network. Staying current with these tools is paramount; being even half a generation behind significantly impacts productivity.

However, as AI accelerates software creation, the bottleneck is shifting from coding to defining what to build – effective product management. The traditional engineer-to-product manager ratio (4:1, 7:1, 8:1) is being disrupted, potentially moving towards 1:1 or even 1 engineer per PM. This is driving the rise of the “Engineer-PM” – individuals who can both code and effectively define product requirements, gather user feedback, and iterate quickly. While acknowledging the risk of discouraging engineers from product work, the speakers emphasize the increased velocity achieved by combining these skills.

Responsible AI & Navigating Bias

Responsible AI is evolving beyond “fluffy social issues” to address tangible business risks and reputational damage. The recent issues with Google’s Gemini model generating historically inaccurate and biased images (specifically, depicting diverse samurai) serve as a cautionary tale. Attempts to enforce fairness through prompt interception backfired, demonstrating the complexities of bias mitigation and the importance of nuanced understanding. The key is to approach Responsible AI as an ongoing process of learning from mistakes and acknowledging that errors will occur. “Grace” is needed when addressing mistakes made by both the AI and the people working with it.

Technical Debt & the Rise of “Vibe Coding”

AI-powered code generation tools (“vibe coding”) are not replacing engineers but empowering them to become more effective “trusted advisors.” A critical consideration when using these tools is technical debt – the implied cost of rework caused by choosing an easy solution now instead of a better approach. Technical debt, like financial debt, can be “good” (investing in a valuable asset) or “bad” (impulse purchases). The core question is whether the speed of development is worth the accruing technical debt. Classifications of technical debt include soft debt and hard debt.

Agentic AI & Filtering the Noise

Agentic AI represents a more sophisticated approach than simple prompt-based code generation. It operates through a four-step process: Understand Intent, Planning (declaring available tools like web browsing and search), Action (utilizing tools to execute the plan), and Reflection (evaluating results and iterating). This framework promotes more robust and reliable solutions.

The speakers caution against the pervasive hype surrounding AI, particularly on social media, advocating for filtering “signal from noise” and becoming a trusted advisor by understanding the fundamentals. The importance of objective evaluation and focusing on real-world value is emphasized.

The Bifurcated AI Landscape & Future Skills

The AI industry is predicted to bifurcate into two main branches: “Big AI” (LLMs like Gemini, Claude, and GPT, driven by the pursuit of Artificial General Intelligence (AGI)) and “Small AI” (open-weight, self-hostable models, particularly gaining traction in China, offering privacy and customization benefits). Skills in fine-tuning these smaller models will be highly valuable.

Real-World Applications & Social Impact

Case studies illustrate the potential of AI. Alipe (Alipay) improved its photo search feature by leveraging on-device AI powered by scalable matrix extensions (SME) on CPUs, addressing privacy, latency, and cost issues. Google’s TensorFlow Professional Certificate program, despite its positive impact (including helping a Syrian man escape poverty), was cancelled due to a lack of direct revenue generation. Unexpected success stories include a former ice hockey player automating reports for his nonprofit with ChatGPT and a researcher in Wales gaining access to crucial GPU resources through Google Colab.

Conclusion

The speakers emphasize that AI is a powerful tool with the potential for both significant good and harm. Success in the AI landscape requires a pragmatic approach, a focus on fundamentals, a willingness to learn continuously, and a commitment to responsible implementation. AI professionals have a crucial role to play as “trusted advisors,” helping others navigate the hype and harness the power of AI for positive impact. Ultimately, AI is neutral; its impact is determined by human agency and the choices we make in its development and deployment.