---

This is a rough transcript of a talk I gave at a Bits in Bio Lightning Talk in June 2024.

The Shift

We're entering a new phase in AI development. Through my recent experience building a DNA foundation model playground for researchers at Stanford and the Arc Institute, I've observed a fundamental shift in how AI products are built and deployed.

The Old World

Building machine learning applications used to follow a predictable pattern:

• Backend-first development, exclusively in Python
• Small, specialized models
• Heavy deployment requirements needing dedicated MLOps teams
• Development cycles measured in quarters or years
• Output focused on academic papers rather than products

The New Reality

Almost every aspect of AI development has been transformed:

• Foundation models have become the new standard
• Deployment has simplified through Models as a Service
• Frontend development has taken center stage
• Product delivery cycles shortened to days or weeks
• Focus shifted from papers to products

The Capabilities Overhang

This transformation didn't happen overnight. We've been sitting on untapped potential:

• 2018: Google Research introduces the transformer architecture
• 2020: GPT-3 completion
• 2022: ChatGPT launch

What's fascinating is that ChatGPT's massive success caught everyone—including OpenAI—off guard. Why? Because we were still thinking in the backend-first paradigm. The spark that ignited this new AI era wasn't a technological breakthrough—it was a user interface breakthrough.

The Power of Good UI

GitHub Copilot stands out as another product that truly understands the importance of UI in AI applications. For non-developers, Copilot might seem like just a sophisticated autocomplete tool. But it embodies crucial principles for AI interfaces:

• Low-cost failures: If suggestions aren't useful, users simply keep typing
• Fuzzy interaction model: Embraces AI's non-deterministic nature
• Immediate feedback: Users see suggestions in real-time
• ~80% success rate with minimal downside risk

Looking Forward

The implications for life sciences and other critical domains are significant. As we build AI tools for these fields, we need to balance powerful capabilities with intuitive interfaces while keeping the cost of AI errors low. The future belongs to teams that can combine deep technical capabilities with thoughtful user experience design.