← Back to list
Colossus (Invest Like the Best / Business Breakdowns)Podcast17 Dec 2024Source: joincolossus.comHost: Patrick O'Shaughnessy

Scott Wu - Building Cognition - [Invest Like the Best, EP.402]

In plain words

This interview covers the current state and future of AI coding assistant Devin. Scott Wu, CEO of Cognition, believes AI won't replace programmers but will free them from tedious coding to focus on creativity. He predicts AI will beat the world's top programmer within 1-2 years. Key holdings mentioned: Devin (priced at $500/month, 10x more efficient than humans), GitHub Copilot (10-20% efficiency boost, a complementary tool), and Datadog (benefiting from AI trends, market cap ~$50-60 billion).

AI SummaryAI-generated · may contain errors · verify against the original

At a Glance Scott Wu, co-founder and CEO of Cognition, discussed the company's AI software engineer Devin on the Invest Like the Best podcast. Devin has already reached the level of a junior engineer, capable of handling the complete workflow from bug fixes to submitting pull requests. Wu predicts t

~11 min full read · 8 sections
Deep Analysis

At a Glance

Scott Wu, co-founder and CEO of Cognition, former International Olympiad in Informatics (IOI) champion. This episode explores the current capabilities, technical architecture, and industry impact of his AI software engineer Devin. Core thesis: Scott Wu predicts that AI will defeat the world's best competitive programmer, Gennady Korotkevich, within 1-2 years, but AI will not replace programmers. Instead, it will shift software engineering from "survival mode" to "creation mode"—freeing humans from 90% of implementation work to focus on the 10% of creativity and problem definition.


Theme 1: Devin's Capability Positioning — "Junior Engineer" Rather Than "Super Tool"

Scott Wu believes that Devin's current capability level is equivalent to that of a junior software engineer, but its value lies not in replacing humans, but in changing workflows.

  • Capability Evolution: One year ago, it was equivalent to a "high school CS student"; six months ago, it was an "intern"; and currently, it is an "entry-level junior engineer." Devin can handle a complete engineering workflow: reproducing bugs, reviewing logs, writing code, adding unit tests, and submitting pull requests.
  • Usage Pattern: Users treat Devin as an asynchronous collaborator — @Devin on Slack, assign a task, and then continue with their own work. Devin autonomously completes the task and submits code for review.
  • Quantified Impact: Internal client research shows that each hour of Devin usage saves 8–12 hours of manual engineering time. For tasks Devin can handle, the efficiency gain is 10x rather than 10%.

> "It's very much a junior engineer today... you're working asynchronously with your team of Devins while you're doing your own stuff."


Theme 2: From "Synchronous Completion" to "Asynchronous Agent" — The Fundamental Paradigm Shift

Scott Wu distinguishes between IDE assistant tools (e.g., Copilot) and the Agent paradigm represented by Devin, arguing that the core difference lies in "synchronous vs. asynchronous" automation and "task-level vs. line-level" automation.

  • IDE Assistant Tools: Provide code completion, but users still need to think and confirm each line. Overall efficiency improvement is roughly 10-20%.
  • Devin's Agent Paradigm: Users delegate the entire task, and Devin autonomously completes the full workflow from environment setup to code testing. For tasks it can handle, efficiency improves by roughly 10x.
  • Technical Differences: Devin needs the ability to autonomously test code, consult documentation, browse websites, and interact with tools like GitHub/Slack. It has its own machine environment, enabling multi-step decision-making and state rollback.

Scott Wu believes that Agents will become the primary consumption scenario for GPUs and language models — because behind each Agent task lie hundreds or even thousands of model calls, rather than a single query-response.

> "With Devin... for the tasks that Devin can do, it's more like a 10x rather than a 10%." (Meaning: for tasks Devin can handle, the efficiency improvement is 10x, not 10%.)


Theme 3: The Essence of Software Engineering and the Future of AI—"Tell the Computer What You Want"

Scott Wu defines the essence of programming as "telling the computer what you want it to do," arguing that AI represents the next-generation human-computer interface, which will fundamentally transform the cost structure of software development.

  • Historical trajectory: From punch cards to assembly, C, and Python, each layer of abstraction has made "telling the computer" easier. Python represents the best compromise to date, but it remains a trade-off between "humans must learn code" and "computers must tolerate inefficiency."
  • Future vision: Any human can precisely describe requirements in natural language, and the computer automatically generates efficient, complete code. Scott Wu estimates this goal can be achieved within 5 to 10 years.
  • Cost structure transformation: Current software development is extremely expensive, making it worthwhile only for applications that can reach millions of users. In the future, it will become feasible to build custom software for niche needs serving just a few thousand people or even a single individual.

Scott Wu cites a metaphor from his co-founder Walden: We have been playing Minecraft in "survival mode" (constrained by resources), and now we are about to enter "creative mode" (focusing solely on ideas and imagination).

> "All of programming... it really is just about telling your computer what to do."


Theme 4: Business Impact — Winners and Losers After Supply Bottlenecks Are Removed

Scott Wu argues that the software industry has long been constrained by supply (the number of engineers), not demand. When AI significantly lowers the cost of software creation, business models that rely on "switching costs" will be disrupted, while network effects and data personalization will become more important.

  • Jevons Paradox: The easier software is to build, the more software people will build. Over the past 40 years, software development efficiency has improved roughly tenfold, yet both the number of programmers and the total volume of software have grown.
  • Business models under pressure: SaaS companies that depend on switching costs and lock-in effects — when AI can easily handle migration and reimplementation, competition will revert to "better products" themselves.
  • Business models that benefit: Network effects, data personalization, and infrastructure owners. Companies that can deliver truly personalized products for each user will gain a greater advantage.
  • Pricing logic: Devin adopts usage-based value pricing (Agent Compute Units), with the goal of making each task ten times cheaper than human labor. The cost of one hour of Devin work is approximately $8–12.

Scott Wu points out that enterprise tools (such as Slack, GitHub, Datadog, and Atlassian) will see "Agent customers" — these tools will serve not only humans but also AI Agents, and need to be prepared accordingly.


Theme 5: Competitive Landscape and Construction Strategy — "Software Engineering Is Very Messy"

Scott Wu believes the current AI software engineering space is still in a "green field" stage, but builders need to focus on "engineering details beyond model IQ."

  • Competitive Landscape: The software engineering field is extremely vast, and each sub-direction (data observability, incident response, testing, migration) could give rise to multi-billion-dollar companies. The space is still early, with product forms and use cases far from settled.
  • Relationship with Foundation Models: Cognition is 100% focused on solving the "messy" problems in software engineering — tool integration, codebase learning, multi-step decision-making, state management, user interfaces, etc. Improvements in foundation model IQ are incremental, but translating that IQ into real-world value requires a great deal of engineering detail.
  • On Scaling Laws: Scott Wu believes "scaling laws are somewhat mythologized" — progress comes not only from piling on compute but also from introducing new techniques (e.g., post-training optimization, reinforcement learning).
  • Internal Usage: The Cognition team itself uses Devin extensively — @Devin in Slack's crash, bug, and feature channels. In about 50% of cases, Devin's output can be merged directly.

> "Software engineering is just very messy... there's this entire thread between of how do you take that IQ and turn that into something that is actually meaningfully useful."


Mentioned Positions

Position Guest Sentiment Key Data
Devin (Cognition) Bullish (core product) Priced at $500/month; hourly cost approx. $8–12; client research shows 1 hour of Devin = 8–12 hours of manual labor
GitHub Copilot Neutral (complementary tool) Efficiency improvement of approx. 10–20%
OpenAI (O1 Pro) Neutral (different positioning) Suitable for single-file code generation and Q&A, not suitable for multi-step iterative workflows
Datadog Bullish (benefiting from Agent trend) Revenue in the billions of dollars, market cap approx. $50–60 billion
PagerDuty Bullish (benefiting from AI incident response) No specific data provided
Slack / GitHub / Atlassian Bullish (will see Agent clients) No specific data provided

Judgments Worth Remembering

1. AI will defeat the world's best competitive programmer within 1-2 years (Scott Wu) — Gennady Korotkevich is currently the world's top programmer, and models from Google and OpenAI are already approaching the level of elite competitors; reaching the top is only a matter of time. This will be an "AlphaGo moment."

2. Software engineering has long been constrained by supply, not demand (Scott Wu) — Every engineering team has 20 projects they want to pursue but can only choose 3. When AI lowers the cost of creation, we will see Jevons paradox: the easier software is to build, the more people will build it.

3. "Minecraft survival mode vs. creative mode" (Walden, co-founder of Cognition) — In the past, humans were limited by resources (time, skills, teams) and could only realize a few ideas; AI will usher humans into "creative mode," where they can focus solely on imagination and creativity.

4. Business models reliant on switching costs will face disruption (Scott Wu) — When AI can easily handle migration and reimplementation, competition will revert to "better products" themselves. Network effects and data personalization will become more important.

5. Agents will become the primary consumption scenario for GPUs (Scott Wu) — Each agent task involves hundreds of model calls, not a single query-response. Foundation model companies will need to optimize for agents (long context, multi-turn tracking, etc.).

6. "The scaling law is somewhat mythological" (Scott Wu) — Progress comes not only from piling on compute but also from the introduction of new technologies (post-training optimization, reinforcement learning, etc.). It cannot be simplistically assumed that "just adding compute will make things better."

7. Enterprise tools will see "agent customers" (Scott Wu) — Platforms like Slack, GitHub, and Datadog will serve not only humans but also AI agents. These companies need to optimize their product experience for agents.

8. Sometimes solving a bigger problem is actually easier (Scott Wu) — Of the 20 people on the Cognition team, 14 were former founders. A truly grand vision can attract top talent, making "difficult" things "easy."