Jul 10, 2025, 10:02 AM ETLecture, Fireside Chat, Keynote

Andrew Ng: Building Faster with AI

Execution speed is identified as a strong predictor of startup odds of success.
AI technology is accelerating startup velocity, with best practices evolving every 2–3 months.
AI Fund builds approximately one startup per month, co-founding by writing code and defining features directly.
The application layer is posited as the layer with the highest opportunity, as these applications generate the revenue required to fund the underlying semiconductor, cloud, and foundation model infrastructure.

Agentic AI is identified as the most critical technology trend, moving beyond simple linear prompt-and-output to iterative workflows involving research, drafting, critique, and revision.
Agentic workflows are essential for complex tasks such as medical diagnosis, legal document reasoning, and compliance extraction.
A new agentic orchestration layer has emerged to coordinate calls between application builders and underlying technology layers.
The "AI Stack" is defined as semiconductors at the base, followed by hyperscalers, foundation model companies, and finally the application layer.

Concrete Ideas: Vague concepts (e.g., "optimize healthcare assets") hinder speed; concrete specifications (e.g., "book MRI slots online") allow engineers to build immediately.
Expert Intuition: Subject matter experts with long-term domain intuition can make faster, higher-quality decisions than data-driven approaches in the early stages.
Single Hypothesis: Startups should pursue one clear hypothesis doggedly and pivot rapidly to a new concrete idea if data falsifies the current one.
Rapid Prototyping: AI coding assistance allows for 10x faster prototype construction compared to production code; prototypes may intentionally use lower security standards if isolated from public data.
Architecture Flexibility: The plummeting cost of engineering has shifted software architecture decisions from "one-way doors" (high cost to reverse) to "two-way doors" (low cost to reverse).
Codebase Rebuilding: Teams are now more willing to completely rebuild codebases to change data schemas or tech stacks due to reduced engineering costs.

Engineering Bottleneck Shift: As AI coding tools increase engineering speed, the bottleneck is shifting to product management and design.
PM-to-Engineer Ratios: Ratios are shifting from historical norms of 1:4 or 1:7 toward 0.5:1 (twice as many Product Managers as engineers) to match the new velocity.
Universal Coding Literacy: There is a trend toward empowering non-engineering roles (CFO, recruiters, front desk) to learn coding to increase overall organizational productivity.
Prompting as a Skill: The ability to precisely command computers (via prompting or coding) is becoming a primary differentiator for individual productivity.

Feedback Speed Hierarchy: The fastest feedback comes from the founder's gut (if an expert), followed by friends/teammates, strangers in public spaces, controlled tester groups, and finally A/B testing (identified as the slowest tactic).
Mental Model Iteration: Teams must use data not just to validate decisions, but to update their mental models, improving the speed and accuracy of future intuition-based decisions.
AI Domain Knowledge: Understanding specific AI capabilities (latency, fine-tuning vs. prompting, agentic workflows) provides a massive speed advantage, preventing weeks of wasted effort on incorrect technical approaches.
Combinatorial Opportunities: Mastery of multiple AI "building blocks" (e.g., RAG, fine-tuning, evals, embeddings) allows for exponentially more software combinations than using a single tool.

Ethical Filters: AI Fund has terminated multiple projects with solid financial cases based on ethical grounds.
Avoiding Hype Narratives: Leaders should dismiss narratives regarding AGI-driven jobless futures, existential risks used for fundraising, or the necessity of nuclear power for all AI to avoid distraction.
Responsible vs. Safe AI: Safety is framed as a function of application and responsibility rather than the technology itself; "Responsible AI" is preferred terminology over "AI Safety."
Regulatory Risks: Hyped dangers are sometimes weaponized to justify restrictive regulations (e.g., SB 1047) that would stifle open-source innovation and create gatekeeper monopolies.

Compute & Infrastructure: Hype regarding "shipping GPUs to space" or exclusive reliance on nuclear power is dismissed; terrestrial GPU capacity remains sufficient for the near future.
Token Economics: Token costs rarely become a problem until a startup has significant user volume; architecture should prioritize flexibility to switch models based on evaluation results rather than cost.
Education Future: The convergence of AI in education remains experimental; hyper-personalization is likely, but specific workflows (avatars vs. chatbots) are not yet mature.
Moats & Defense: Product-market fit (building something users love) is the primary defense against AI replication; moats often evolve from momentum and channel access rather than being pre-engineered.
Inequality & Open Source: The primary threat to equitable diffusion of AI is not the technology itself, but regulatory efforts by gatekeepers to restrict open-source models and fine-tuning access.