Cheeky Pint: Cognition CEO Scott Wu on AI Agents and the Moneyball-ification of Everything

Podcast: Cheeky Pint
Host: John Collison (Stripe co-founder)
Guest: Scott Wu — CEO of Cognition, three-time IOI gold medalist
Duration: ~59 minutes
Listen: Apple Podcasts | Transcript on Substack | YouTube

Scott Wu is the CEO of Cognition, the company behind Devin, an AI agent that handles software engineering tasks. Before that, he won three gold medals at the International Olympiad in Informatics, left high school early to work at Addepal, briefly attended Harvard, and dropped out to start a company. This conversation covers the state of AI coding, the Windsurf acquisition, and a broader theory about how every field gets more mathematical over time.

The Moneyball-ification of Everything

Wu’s central framework: domains evolve from being intuition-driven to mathematical and data-driven. Poker, chess, and professional Super Smash Brothers all went through this transition. Startups are going through it now.

In the early days of any field, first principles thinking dominates because there is no data. As the field matures, playbooks develop, experience becomes more valuable, and gut feel gets replaced by analytics. This explains why younger founders are less common now than they were a decade ago — the startup game has become more sophisticated.

Devin: How AI Agents Change Software Engineering

Devin is different from tools like Copilot. Copilot helps a human write code faster. Devin works asynchronously — you give it a task through Slack or Jira, and it goes off and does the work. In the companies that use it well, Devin handles roughly 30-40% of merge pull requests.

Wu distinguishes between essential complexity (the actual decisions and logic in software) and accidental complexity (routine implementation, boilerplate, standard processes). AI is good at the latter. Humans should focus on the former.

“We Have AGI”

Wu makes a provocative claim: we already have AGI. He clarifies that he does not mean a sudden singularity is coming. He means that current systems demonstrate general intelligence capabilities, even if they do not match the science fiction version. He expects continuous improvement rather than a sharp break.

His prediction: within 2-4 years, software engineers will no longer look at code directly. They will focus on high-level architectural decisions, the same way modern programmers do not work in assembly language.

Jevons Paradox and Engineering Jobs

AI will not reduce the number of software engineering jobs. Wu cites Jevons paradox: making something more efficient increases demand for it. The more code AI can write, the more software people will want built. The bottleneck becomes human judgement and architectural thinking, not implementation speed.

Acquiring Windsurf Over a Weekend

The Windsurf acquisition is a good story. On a Friday, Wu heard that Google might acquire Windsurf, an IDE company. Cognition reached out, negotiated over the weekend, and announced the deal on Monday morning. Wu explains that Windsurf filled gaps in Cognition’s go-to-market capabilities and brought a complementary product. Speed matters in AI.

The AI Industry Structure

Wu argues that all layers of the AI stack — hardware, foundation models, applications — will thrive. Each layer solves different problems and requires different expertise. Hardware companies solve physics problems. Model labs solve research problems. Application companies solve product problems. There is meaningful differentiation at each level, so vertical integration is limited.

Final Thoughts

Wu’s information diet is mostly Twitter, and he cultivates relationships with other founders for honest feedback. Cognition’s culture is intense — the team works out of a house, focused on shipping. He is clear-eyed about the polarized nature of AI investing: the space trends toward hyperscaler or bust. But he believes the companies that survive will change how software gets built entirely.

Crepi il lupo! 🐺