The Problem Solvers: Scott Wu at Cognition

1. From Math Competitions to Building the AI Software Engineer

Scott Wu’s origin story compresses a remarkable arc into a few sentences. He grew up immersed in math and programming competitions — winning the world championship in competitive programming in 2014, at age 17. The video opens with an even earlier scene: a much younger Wu, at a competition, correctly answering a math buzzer question in seventh grade. The juxtaposition is deliberate. This is someone who has spent his entire life at the frontier of what humans can do with logic and computation.

What makes Wu’s trajectory interesting is not just the achievements but the question they led him to. He recalls talking with friends about whether AI would ever be able to do the things competitive programmers do. “And to say the obvious, it’s crazy,” he says — reflecting both on how wild the question seemed at the time and on how dramatically circumstances have changed. The person who was asking “will AI ever match humans at competitive programming” is now the person building AI that writes production software for banks, governments, and private equity firms.

Competitive programming is not arbitrary preparation for building Devin. The skill set required — modeling complex problems precisely, reasoning about edge cases, optimizing solutions under constraints, debugging incorrect logic systematically — is exactly what an AI software engineer needs to do well. Wu’s fluency in that domain gave him an unusually precise intuition for what “good” looks like in automated software engineering. He knew the hardest cases because he had solved them himself.

2. The Art and Craft of Software Engineering at Scale

What kept Wu committed to Devin through the period when “people didn’t really believe it was possible” was a grounding in craft. “We just care so much more about the art and craft of software engineering on the ground,” he says, “and how do we figure out how to go and make that much faster and solve all the practical problems and unblock that.”

This is a revealing statement about what it actually takes to build autonomous software engineering agents. The hard part is not the impressive demos — it is the long tail of practical problems that show up in real codebases: legacy dependencies that do not behave as documented, error messages that are ambiguous or misleading, edge cases that require understanding the business context not just the code, test suites that are slow or flaky. These problems require software engineering judgment, not just coding ability.

Cognition’s mission is quantitatively ambitious: help build software around the world ten times faster. They are pursuing that mission with enterprises — banks, health insurers, governments, private equity firms — where the codebases are complex, the stakes are high, and the “practical problems” Wu mentions are the default rather than the exception. Getting to 10x in those environments requires not just a model that can write code, but an agent that can navigate ambiguity, maintain context across a long task, and recover gracefully from errors.

3. No Single Product for All of Software Engineering

One of the most practically important things Wu says is about market structure, not technology. When asked about Anthropic’s launch of Claude Code — a product that competes in some ways with Devin — Wu’s response is thoughtful: “I really just don’t think there’s gonna be a single product that does all of code and all of software engineering. There are so many experiences to serve. And I think from that perspective, there’s so much more for us to do together than there is separate.”

This is a specific claim about the shape of the software engineering market that is worth examining carefully. Software engineering spans an enormous range of contexts: solo developers building personal projects, startups shipping product fast, enterprise teams maintaining decades-old codebases, regulated industries with specific compliance requirements, specialized domains like embedded systems or scientific computing. The needs in each context differ significantly — the tools that help a solo indie developer are not the same as those that help a Fortune 500 engineering team.

Wu is saying that the market is large enough and diverse enough that multiple specialized products can succeed without needing to displace each other. An AI software engineer optimized for enterprise banking codebases is a different product from a coding assistant optimized for individual developers. Both can be excellent at their respective use cases. The total opportunity is so large that competing for a piece of it is a more productive frame than assuming one winner takes all.

4. A World Where Building Software Is Telling Your Computer What to Do

Wu closes with a vision statement that connects his competitive programmer past to the broader future: “From the beginning, I always loved that you could take any idea that you had and turn it into reality. Everyone will be able to turn those ideas into reality.” And then the headline claim: “This is a world where building software is just as easy as telling your computer what to do.”

This is the endpoint of a journey that started with Wu being one of a small elite who could wield computation at the highest level. The progression is: first, machines make computation possible; then, programming languages make computation accessible to those who invest in learning them; then, AI makes software creation accessible to anyone who can describe what they want. Each step removes a layer of required expertise without reducing the power available.

The depth of Wu’s commitment to this vision — and his credibility in pursuing it — comes from the fact that he has lived on both ends of the spectrum. He was the person who could do the thing that seemed impossible. He is now building the system that makes the impossible thing unnecessary for everyone else. The competitive programmer who could solve problems in milliseconds is building the AI that will let everyone else solve their problems without needing to know how.

🧠 Check your understanding

5 questions · your answers are saved in this browser only

1. 1. What was Scott Wu's competitive programming achievement that forms the backdrop of his story?

   He won a Google Code Jam competition in college He was world champion of competitive programming in 2014 He placed first in the International Mathematical Olympiad He won a hackathon that led to Cognition's founding

   Wu was world champion of competitive programming in 2014 — this achievement, and his lifelong immersion in math and programming competitions, directly shaped his intuition for what autonomous software engineering requires.

2. 2. What does Cognition's stated mission mean in concrete terms?

   Building the most capable AI coding assistant for individual developers Helping build software around the world 10 times faster, primarily for large enterprises Creating open-source AI tools for competitive programming Replacing human software engineers with fully autonomous AI systems

   Cognition's mission is to help build software 10x faster. They pursue this at institutional scale — banks, governments, health insurers — where the codebases are complex and the stakes are high.

3. 3. How does Wu respond to Anthropic launching Claude Code, a potential competitor?

   He expresses concern about increased competition in the AI coding market He says no single product will serve all of software engineering — the market is too large and diverse for winner-takes-all He announces plans to differentiate Devin by targeting smaller developers He credits Claude Code with setting the direction for autonomous coding agents

   Wu says he 'really just don't think there's gonna be a single product that does all of code and all of software engineering' because 'there are so many experiences to serve' — framing the market as collaborative rather than zero-sum.

4. 4. What kept Wu committed to the Devin vision during the period when 'people didn't believe it was possible'?

   Early customer revenue that validated the product direction A deep care for the art and craft of software engineering on the ground Research results from academic benchmarks showing AI progress Encouragement from investors who believed in the mission

   Wu says the grounding was caring 'so much more about the art and craft of software engineering on the ground' — the practical, real-world problems of making software faster, not the impressive demos.

5. 5. What is the ultimate vision Wu articulates for where software creation is heading?

   AI will write all software autonomously with no human involvement Competitive programmers will use AI to win programming competitions Building software will be as easy as telling your computer what to do — turning ideas into reality for everyone Software development will be consolidated into a small number of AI-native platforms

   Wu's closing vision is 'a world where building software is just as easy as telling your computer what to do' — connecting his love of turning ideas into reality to a future where everyone has that power.

Reset quiz

🛠️ Build it yourself

Follow these exact steps to reproduce it yourself · estimated time: ~60 min

Prerequisites

• Access to an autonomous coding agent (Devin, Claude Code, or equivalent)
• A clearly specified engineering task with defined success criteria

Step 1 — Specify a real engineering task precisely

Write a one-paragraph task description for a real software problem: what system to modify, what behavior to add or fix, what constitutes a correct solution. Precision matters more than complexity — a vague task is harder for an agent than a clear but ambitious one.

Step 2 — Evaluate planning behavior

Before the agent writes code, observe how it breaks down the task. Does it identify ambiguities and ask clarifying questions? Does its decomposition match how you would approach the same problem? The quality of planning is often more predictive of success than the quality of the eventual code.

Step 3 — Observe edge case handling

Watch for how the agent handles the first failure: a test that does not pass, an error it was not expecting, a dependency that behaves differently than documented. Does it recover systematically or get stuck? This is where the “art and craft” that Wu describes shows up most clearly.

Step 4 — Measure the autonomy-accuracy tradeoff

Track how many times you intervened to correct the agent versus let it proceed. In retrospect, were your interventions necessary? Could you have trusted the agent further, or did your corrections prevent real errors? This tradeoff is the core design challenge in autonomous software engineering.

Step 5 — Consider the market structure question

After the exercise: would you use a general-purpose coding agent or a specialized one for this specific task type? What would a specialized agent optimized for your domain look like? This is Wu’s “many experiences to serve” insight applied to your own context.