Beyond the Breakthrough – Peter Norvig

What Comes After the Breakthrough?

This document serves as a reference guide to the keynote, summarizing the core ideas, structure, and actionable insights presented by the speaker.

1. Main Theme: From Scaling to Collaborative Problem Solving

The central theme of the keynote, “What Comes After the Breakthrough?”, explores the evolution of AI from the initial “unreasonable effectiveness of data” to the current era of powerful LLMs and agents. The primary message is that while scaling data and computation has brought us incredible tools, the next true breakthrough lies beyond just coding. It requires us to build systems that understand the “theory of the evolving program”—the context, intent, and evolution of software over time. The ultimate goal is to move from isolated coding tasks towards collaborative problem solving within a global, self-improving ecosystem, where AI helps us build more “livable” technological environments that serve all stakeholders.

2. Sub-themes and Chapters: The Evolution of AI

The speaker structured the keynote by tracing the history of AI to understand its future trajectory, focusing on the progress in mathematics and coding as key examples.

How We Got Here: The Unreasonable Effectiveness of Data

The speaker, an “AI hipster” from the 1980s, contrasted the old approach of building specialized, rule-based systems (like the “Unix Consultant”) with today’s reality. The field once believed that different domains required entirely separate models and handcrafted knowledge. The key shift, as highlighted in the 2009 paper on “the unreasonable effectiveness of data,” was realizing that massive amounts of data (“chugging in lots of text into that data ingestion funnel”) combined with computation and search were more powerful than designing better algorithms alone. The continuous success of scaling laws, as shown in log-log graphs from 2020 to 2022, proves this point, with models still not falling “off that curve.”

What’s Happening Now: Breakthroughs in Math and Coding

Progress is not just linear; it’s “jagged,” with systems solving complex problems while failing at simple ones.

• Mathematics: AI has made astonishing leaps, from failing basic arithmetic in 2022 to winning a gold medal at the International Math Olympiad (IMO) in 2025. This success is attributed to several factors:
  • Verification: Using formal languages like Lean to check proofs.
  • Unmatched Breadth: LLMs have read all the math papers, allowing them to connect disparate fields in ways human specialists rarely can.
  • Depth through Search: As OpenAI demonstrated by setting “10,000 agents” on the Navier-Stokes problem, AI can apply massive search capabilities to find solutions.
  • Finding Counterexamples: AI excels at finding specific counterexamples (e.g., “set a = 17”), which is often harder for humans who are trying to prove a general theory.
• Coding: AI has become an indispensable tool, leading even skeptics like Linus Torvalds to become “dogmatic advocates.” This is because coding benefits from a “perfect simulator” (you can run the code to see if it works) and vast training data from platforms like GitHub, which include not just successes but also failures and bug fixes. However, the speaker noted that while models can generate code, it is often verbose and inefficient compared to a human’s concise solution, as they don’t feel the “cost” of writing extra lines.

What Comes Next: From Coding to a Global Ecosystem

The true challenge lies in moving beyond isolated tasks. The speaker outlined a clear progression for the future:

1. Coding: This is a largely solved problem, with continuous improvements expected.
2. Software Engineering: This is the next frontier. It’s “coding + time + other people.” The core challenge is capturing the “theory of the evolving program”—the original design intent and its evolutionary path. Without this, systems decay when maintained by new teams. The anecdote of the speaker accidentally trying to check in 6,000 temporary files illustrates how our current “best practices” will need to evolve.
3. Problem Solving: Inspired by architect Christopher Alexander’s work on “design patterns,” the focus must shift from just code to creating “livable environments.” This means embedding AI solutions in the real world and solving problems holistically.
4. Collaborative Problem Solving: The next step is to build systems that incorporate the viewpoints of all stakeholders from the beginning, avoiding the common pitfall of building a product in a vacuum and then reacting to user complaints.
5. Global Self-Improving Ecosystem: The ultimate vision is a system where knowledge from papers, code from GitHub, and other sources are integrated into a faster, “recursive self-improving loop” that benefits everyone.

3. Inspiring Quotes

• “Proof that if you work in a field for 40 years, it can become an overnight sensation.”
• “We thought the world had changed… it’s more important to say if I can have better data, that’s more important than having a better algorithm.”
• “It’s still jagged, and they still make mistakes on things that should have been much easier.”
• “The core skill of coding plus time. Plus, other people. This idea that you have to interact with the rest of the world—that’s what makes it hard.”
• “I think that’s what we should really be focusing on: how do we make livable environments that are easy to work with and solve problems that way, and move away from just the code.”

4. Key Action Points

1. Capture the “Theory of the Program”: When developing software, don’t just document the “what” (the code). Focus on documenting the “why” and “how it should evolve.” This “theory of the evolving program” is crucial for long-term maintenance and success, especially as teams and AI agents change.
2. Embrace AI as a Broadness Engine: Use AI’s “unmatched breadth” to your advantage. Encourage it to connect ideas from different fields or explore unconventional solutions. Pair this breadth with rigorous verification and search to achieve depth, as demonstrated in the mathematical breakthroughs.
3. Rethink Your Processes for Collaboration: Move beyond seeing AI as just a tool for individual productivity. Start redesigning your workflows—from specifications and code reviews to stakeholder engagement—to leverage AI for collaborative problem-solving. Consider all stakeholders upfront to build more robust and “livable” solutions.