No PriorsNo Priors Ep. 62 | With Cognition CEO and Co-Founder Scott Wu
CHAPTERS
- 0:00 – 1:12
Scott Wu’s early coding journey and love of math
Sarah introduces Scott Wu, CEO/co-founder of Cognition, and frames the conversation around Devin—an AI software engineer. Scott shares how he started programming at age nine and why turning ideas into reality hooked him early.
- •Cognition positions itself as an AI lab focused on reasoning
- •Devin demo and SWE-Bench results set the context for the episode
- •Scott’s early exposure to programming via his older brother
- •Math and competitions shaped his identity and career direction
- 1:12 – 1:53
What IOI is and what it tests
Scott explains the International Olympiad of Informatics (IOI) as the ‘Olympiad of code.’ He describes its country-team format and its emphasis on algorithmic problem solving plus implementation under constraints.
- •IOI as a global competitive programming championship
- •Teams represent countries; medal-based outcomes
- •Problems require optimal algorithm selection
- •Implementation skill matters alongside theoretical insight
- 1:53 – 3:35
How competitive programmers train: fundamentals + creative reduction
The discussion digs into what practice looks like for elite competitive programming. Scott contrasts learning standard algorithms with the real contest skill: creatively mapping novel problems onto known techniques.
- •Training is intense repetition and analytical self-critique
- •Standard tools (shortest paths, data structures) are foundational
- •Each contest problem is unique by design
- •Core skill is reduction and adaptation of known methods
- 3:35 – 4:40
The online IOI community and its influence on careers
Elad asks about the role of online communities in shaping Scott’s network and development. Scott describes how remote friendships and periodic camps created a tight-knit cohort that later fed into startups and AI.
- •Geographic isolation made online community essential
- •Competitions and camps formed long-term friendships
- •Ongoing online collaboration for most of the year
- •Many peers became founders and AI practitioners
- 4:40 – 7:03
Why competitive programming maps to entrepreneurship
Sarah highlights Scott’s view that competitive programming has surprising overlap with startup building. Scott explains the shared habit of challenging assumptions, thinking independently, and finding non-obvious solution paths.
- •Startups and contests both reward first-principles thinking
- •Great solutions often come from questioning hidden assumptions
- •Creative leaps matter more than rote application
- •Continuous improvement and self-push are central to both
- 7:03 – 8:20
Cognition’s origin story and founding team composition
Scott explains how Cognition formed around November, drawing heavily from his competitive programming network. He outlines the founders’ prior experience across AI, infrastructure, and developer tooling, motivating a focus on accelerating software creation.
- •Cognition formed with friends from contest communities
- •Founders bring AI/infra/tooling experience (e.g., Lunchclub, Cursor, Scale)
- •Belief that software drives much of modern progress
- •Thesis: demand for engineering is massive; accelerate coding throughput
- 8:20 – 9:12
Meet Devin: an autonomous AI software engineer (end-to-end workflow)
Scott defines Devin as an agent that can make its own engineering decisions and carry tasks from prompt to completion. He emphasizes tool use—shell, browser, docs, tests, debugging—and an interactive workflow where humans can watch and steer.
- •Devin can plan, code, run commands, read docs, test, debug, deploy
- •Designed for end-to-end task completion, not just code snippets
- •Human can observe what Devin is doing and give feedback
- •Goal: mirror working with another engineer ‘over the shoulder’
- 9:12 – 12:08
Devin discourse: skepticism, job fears, and why engineers won’t disappear
The hosts ask about the online reaction and employment implications. Scott argues AI will expand engineering output and unlock latent demand, shifting human engineers toward higher-level problem definition and creative decision-making.
- •Mixed reactions: disbelief vs. ‘jobs are doomed’ narratives
- •Demand for software is huge; productivity gains increase total building
- •AI doesn’t decide what to build—humans still define goals and constraints
- •Role shift: less time typing code, more time on problem solving
- 12:08 – 14:20
Designing Devin’s UI: steering beats ‘fire-and-forget’ agents
Elad focuses on Devin’s interface paradigm—plan/checklist, shell, code, browser—and why it matters. Scott describes building the UI from their own daily usage, optimizing for frequent lightweight feedback rather than one-shot autonomy.
- •Many agents fail because they return too late with wrong output
- •Devin UI supports transparency across planning, execution, and browsing
- •Inspired by managing interns/junior engineers via quick check-ins
- •Built iteratively from internal use while building Devin itself
- 14:20 – 16:54
Where Devin shines vs. struggles: DevOps, setup, and execution vs. deciding goals
Scott details Devin’s strongest areas (encyclopedic knowledge, DevOps/setup, iterative debugging, data analysis pipelines). He also clarifies limitations: Devin executes well when the task is precisely specified, but it’s not the decider of what to build.
- •Strong at environment setup, debugging, and tool-driven iteration
- •Notable win: unblocking Kubernetes/database setup when humans were stuck
- •Effective for end-to-end data analysis: sourcing, cleaning, analysis, viz
- •Weakness: doesn’t originate product/business goals; needs precise specs
- 16:54 – 18:39
How Devin works (at a high level): planning, evaluation, and tool-based problem solving
Pressed on implementation details, Scott stays high-level: the key is optimizing the ‘interface’ of the engineering problem and getting the agent to behave like a pragmatic engineer. He contrasts perfect one-shot diffs with the real-world loop of reproducing bugs, running code, inspecting logs, and iterating.
- •Cannot share deep internals, but emphasizes system-level optimization
- •Agent must operate via iterative execution: reproduce, inspect, modify, rerun
- •Planning and evaluation are core investments
- •Focus on practical debugging workflows vs. idealized omniscient patching
- 18:39 – 21:13
The evolution of coding agents and software engineering over 1–10 years
Scott forecasts big shifts: in 5–10 years, agents become a new human–computer interface, and ‘software’ itself changes. In the near term, AI-native engineers multiply output as models, hardware, and agent tooling rapidly improve.
- •Longer-term: today’s languages and workflows may look archaic
- •Engineers shift to architecture, edge cases, and solution specification
- •Short-term: AI-native workflows deliver significant personal leverage
- •Multiple accelerants: better hardware, better base models, better agents
- 21:13 – 22:35
What will most improve agents: models, reasoning, memory, planning, tool use, speed
Elad asks which breakthroughs matter most for agents broadly. Scott answers that progress across all dimensions compounds, describing a ‘race’ between improvements in core intelligence and scaffolding like tooling and planning systems.
- •All factors matter: reasoning, memory, planning, tool use, inference speed
- •Agents are composites of several interacting capabilities
- •More capable base models reduce need for scaffolding—but scaffolding still helps
- •Impact depends on where progress is fastest across the stack
- 22:35 – 24:06
Advice for humans: fundamentals remain; engineers become architect-PM hybrids
Sarah asks what skills remain important for humans as English becomes a ‘programming language.’ Scott argues fundamentals (systems, algorithms, reasoning) still matter, analogous to how network experts benefit from knowing TCP despite abstractions.
- •English interfaces rise, but technical fundamentals stay valuable
- •Deep understanding helps when abstractions break or need shaping
- •Future engineer role blends technical architecture and product thinking
- •Core competency: break down problems into precise solution requirements
- 24:06 – 29:28
Long-range uncertainty and Cognition’s hiring philosophy
The conversation touches on distant futures (singularity uncertainty) but focuses on near-term concrete impact. Scott then explains Cognition’s hiring: high-ownership, outcome-driven builders—often former founders—who can span research, engineering, product, and customer work.
- •Hard to predict 20-year outcomes; nearer-term impact is clearer
- •AI revolution is early; lots left on research and product fronts
- •Hiring emphasizes ownership, creativity, communication, and execution
- •Seeking strong engineers/researchers who thrive across functions
