Lecture, Fireside Chat, Keynote
Andrew Ng: Building Faster with AI
- Execution speed is identified as a primary predictor of startup success, with AI enabling engineering velocities to increase 10-fold for prototypes and 30% to 50% for production code.
- The application layer remains the most valuable and opportunity-rich segment of the AI stack, expected to see exponential growth in the number of buildable applications as founders combine AI building blocks.
- A new agentic orchestration layer is predicted to emerge, facilitating iterative workflows involving thinking, research, and critique, while an "agentic" approach becomes a key differentiator for complex tasks like compliance and medical diagnosis.
- Software architecture is shifting from "one-way doors" to "two-way doors" due to plummeting engineering costs, allowing teams to rebuild codebases multiple times within a month and easily switch between foundation model providers.
- Engineering speed is outpacing product definition, making product management the new bottleneck and potentially shifting team ratios to include more product managers than engineers.
- The ability to code is becoming a universal requirement for all job roles, as AI tools empower non-engineers, including CFOs and recruiters, to build software and improve performance.
- Decision-making strategies are evolving to prioritize concrete product ideas and direct user observation over abstract concepts or A-B testing, which is described as increasingly slow.
- Regulatory environments face risks from narratives regarding AI dangers and AI safety, which proponents argue are distorted hype used to push restrictive measures like SB 1047 that could stifle open-source software.
- The AI Fund plans to kill projects on ethical grounds even when the financial case is solid, and the organization aims to co-found approximately one startup per month over the coming years.
- Best practices in AI are projected to change every two to three months, necessitating continuous learning and updates to maintain competitive speed.
- Token costs are expected to remain non-critical for most startups, becoming a constraint only for entities with extremely high user usage volumes.
- The long-term outcome of AI in education remains undefined, with the next decade expected to involve mapping complex educational workflows to agentic systems and achieving hyper-personalization.
- The management team's execution capability is highly correlated with startup success, particularly regarding the ability to pursue a single clear hypothesis and pivot rapidly based on data.