What this covers

This is a conversation between host Patrick and Scott Wu, co-founder and CEO of Cognition, the lab behind the Devin software agent. Wu, a former competitive programmer, defines programming as telling a computer what you want, with only the form factor changing since punch cards. The first stretch covers what Devin does today as a junior engineer, and how asynchronous work separates it from IDE assistants. Wu cites customer studies in which an hour with Devin matched eight to twelve hours without it. Early uses were legacy migrations and version upgrades at banks and healthcare companies.

The middle of the conversation turns to how agents get built and sold. Wu says the hard part is sustaining good decisions across roughly 100 steps, along with the infrastructure for rollback and parallel sessions. He expects multi-turn agent workloads to dominate GPU usage and push pricing toward usage-based models. He contrasts self-driving, which needs extreme reliability, with software agents that pay off at 90 percent because a human reviews the output. He also suggests Python is a compromise between humans and machines, and that natural language could be translated into code that runs 10 to 100 times faster.

The more contested material comes later. Wu predicts AI will beat Gennady Korotkevich, the strongest competitive programmer ever, within one to two years, and cites near-top results at the IOI. He argues AGI already exists in practical terms because the definition is recursive, and he calls scaling laws partly a myth. The talk closes on founder recruiting and a card trick.

Sharpest takeaway

Wu argues that autonomous, asynchronous AI agents like Devin will shift software engineering from hands-on implementation toward specifying problems, and that because software has always been supply-constrained, cheaper software will expand what gets built rather than eliminate programmers.

  • Agents handle end-to-end workflows asynchronously, so for delegable tasks the gain is closer to 10x than the 10-20% gains from autocomplete-style assistants
  • Software has always been bottlenecked by engineering supply rather than demand, so reducing the cost of building should unlock latent demand (Jevons paradox)
  • Foundation model progress is incremental, so durable value lies in the product, tooling, and infrastructure built around models rather than in the models alone

The claims · ranked66 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.78

Defining AGI as doing 80 percent of human labor is recursive, because if AI did that the remaining 20 percent would be human work, and people keep moving the goalposts as AI masters each remaining human-only skill.

definitionhigh valuecontestednovelty 3/4durability 4/4· Scott Wu

“oh, it's doing 80% of human labor. It's like, well, if you could do that, then the humans would be doing the 20%”

0.75

Network effects will always be strong, and data plus personalization will become even more powerful, so businesses owning infrastructure or personal relationships and data can deliver individualized products that generic apps cannot.

normativehigh valueestablishednovelty 2/4durability 3/4· Scott Wu

“Network effects I think will always be golden. I think data and personalization is going to be even more powerful.”

0.73

Software has always been bottlenecked by engineering supply rather than demand, since every team has about 20 projects but picks only three per week, so making engineers 10x more productive should unlock far more software built.

causalhigh valuecontestednovelty 3/4durability 3/4· Scott Wu

“We've always been bottlenecked by supply.”

0.73

Cheaper software creation will follow a Jevons paradox pattern, where building software is about 10 times easier than in the 1980s and the number of programmers and software built grew over 40 years, so similar expansion should occur faster in the coming years.

forecasthigh valuecontestednovelty 3/4durability 3/4· Scott Wu

“I honestly think we're going to see a lot of this Jevons paradox where we're just going to be building more and more and more software as it gets easier.”

0.69

Wu expects AI to beat Gennady Korotkevich, the best competitive programmer in history, within one to two years, and expects the moment to be like AlphaGo, since Google and OpenAI already reach near-top-competitor levels on the IMO and IOI.

forecasthigh valuecontestednovelty 3/4durability 2/4· Scott Wu

“I would imagine that we have AI that can beat Gennady in the next one or two years”

0.69

Wu views scaling laws as partly a myth, arguing that progress comes from new techniques such as post-training and reinforcement learning, along with efficiency innovations, rather than from simply adding hardware to a fixed recipe.

causalhigh valuecontestednovelty 3/4durability 2/4· Scott Wu

“I think scaling laws are a little bit of a myth.”

0.69

Wu holds that AGI already exists in practical terms, since the definition is recursive and what matters is whether AI is good enough, built into products, and widely distributed so that humans are doing far more work because of it.

normativehigh valuecontestednovelty 3/4durability 2/4· Scott Wu

“I'm in camp number two on that. We already have AGI.”

0.59

Customer internal studies found that one hour of engineering time spent using Devin was equivalent to 8 to 12 hours of work without it, because engineers review and ask questions instead of writing code.

factualhigh valuecontestednovelty 2/4durability 1/4· Scott Wu

“every hour of an engineering time using Devin was equivalent to 8 to 12 hours actually without Devin”

0.57

Devin learns a codebase over time, internalizing what files do, how schema migrations are performed and the decision-making behind the code, so its value compounds like a long-tenured engineer's rather than resetting with each task.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“Devin knows the entire codebase in and out. Devin wrote half the files in the codebase.”

0.57

Devin turns each engineer into an engineering manager with a team of Devins, delegating many small tasks and working on them asynchronously while the Devins run in parallel and ask questions or submit code for review.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“it really does function turning every engineer into an engineering manager, where you get to work with your own team of Devins and hand off tasks”

0.57

The soul of programming, and what separates 1x from 10x and 100x engineers, is deep understanding of abstractions and the ability to design the right representation, architecture, and schema, while the rest of an engineer's day is implementation, debugging, and DevOps.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“it's probably only about 10% of what the average engineer does, because the other 90% of their time obviously is debugging, implementing, setting up, testing”

0.57

A single model query costs a tiny fraction of a cent, but Devin may make hundreds or thousands of decisions over half an hour of work, so agents will dominate GPU usage and push foundation labs to optimize for multi-turn traces and long context.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“the dominant usage of these GPUs, of these language models and everything will be for agents”

0.57

Self-driving cars need roughly 99.9999 percent reliability to be useful, while software agents like Devin can deliver value at one nine or 90 percent reliability because human plus machine review still saves about 90 percent of the time.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“if you're doing one nine instead of six nines, it's still super, super useful”

0.57

The bigger gain from cheap AI software creation is better ideas, because seeing a working prototype reveals why an idea is bad, and turning the 10 percent of time spent on ideas into 100 percent should accelerate iteration.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Patrick

“the thing that excites me is not that I can make my ideas, but that I can get better ideas”

0.57

With agents doing tasks rather than answering questions, pricing will move toward usage-based models because agents cost more to run on both the input side and the output side, and usage is the best available proxy for value delivered.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“with agents it does come down to usage for better or for worse.”

0.57

Earlier technology shifts such as smartphones were step functions that created new businesses like Uber, Airbnb, and DoorDash, whereas AI improves incrementally from GPT-3 to 3.5 to 4, so builders should focus on what models will not do by themselves.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“AI is a lot more of a very incremental thing.”

0.57

The biggest step-function difference between IDE code assistants and Devin is asynchronous versus synchronous work: assistants give roughly 10 to 20 percent speedups by autocompleting while the human confirms each suggestion, whereas Devin takes a whole task off the engineer's hands, making it closer to 10x for delegable tasks.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“I think the biggest step function change is about asynchronous versus synchronous”

0.57

Python is an awkward compromise forced by human-computer communication: humans must learn years of syntax, and Python is slow relative to bare-metal code, so AI could translate natural language into efficient, feature-complete software and make it 10 to 100 times faster.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“99% of people in the world don't know how to code and it takes years to learn all the little details”

0.57

Humans repeatedly find one more task that AI cannot do, then AI solves it and humans move the goalposts, a pattern driven by human pride, so the relevant test is not percentage of human labor automated but whether the intelligence is actually used and creates value.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott Wu

“there's going to be lots of little things that are like, well, actually humans are still better at this one thing.”

0.56

Programming has always been the art of telling a computer what you want it to do, and only the form factor has changed from punch cards and assembly to BASIC, C, cloud, and mobile.

definitionhigh valuespeaker onlynovelty 2/4durability 4/4· Scott Wu

“software engineering, this whole time, programming, or whatever we want to call this, has always been basically just the art of telling the computer what you want it to do”

0.56

In the long term, the soul of software engineering becomes each person deciding which problems to solve and specifying exactly the solution to build, while AI supplies the execution power but does not solve problems on its own.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Scott Wu

“the soul of software engineering, again, is each person getting to decide what are the problems that I want to solve”

0.54

The hardest part of building an agent is not single-model intelligence but sustaining step-by-step decisions to maximize success over around 100 steps, plus the infrastructure for rollback of machine state, fast spin-up, and parallel sessions.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Scott

“after 100 steps that I'll have solved the issue, then it's obviously a pretty different thing”

0.53

Many SaaS products like Slack, GitHub, Datadog, and Atlassian will serve AI agents as well as humans, so businesses are preparing their products for a new wave of agent customers because logs, documents, and structured context make agent tasks easier.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Scott Wu

“a lot of these companies are actually going to be serving agents in addition to humans”

0.53

Going after an enormous, exciting problem can make it easier to attract a passionate, talented team, because ambitious missions draw founders who would otherwise raise money for their own projects.

causalhigh valuespeaker onlynovelty 3/4durability 2/4· Scott Wu

“sometimes it's easier to solve a bigger problem than it is to solve a smaller problem”

0.53

One of the first emergent use cases for Devin was large legacy migrations, replatforming, and version upgrades for banks and healthcare companies with decades-old code, which is tedious but high-volume work.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Scott Wu

“One of the first things that companies that we work with have been using Devin for actually is a lot of these migrations, replatforms, modernization and version upgrades”

0.53

Roughly 10 percent of the average engineer's time is spent on the creative work of deciding what to build and how to architect it, while the other 90 percent goes to debugging, implementation, setup, testing, DevOps and similar execution work.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Scott Wu

“it's probably only about 10% of what the average engineer does, because the other 90% of their time obviously is debugging, implementing, setting up, testing”

0.52

If AI absorbs the routine implementation, debugging, and iteration work, engineers can spend roughly ten times more of their time on the creative 10 percent of programming, which is the core value proposition of agentic software tools.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Scott

“what happens then is, honestly, you just get to do 10x more of the 10%.”

0.52

Foundation models will keep increasing in IQ, but the tool integration, decision-making scaffolding, user interface for checking in on the agent, and communication flow are where software-specific value lies, and these complement rising model capability rather than being subsumed by it.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“But a lot of the detail of all of the tools that you work with, how you make decisions with all of that, even how you communicate with the user”

0.52

Software is so expensive to engineer that only applications reaching millions of people, like Facebook, YouTube, or DoorDash, have been worth building, leaving niche communities and personal needs underserved.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“the only software that you can actually build that's worth building is basically applications that reach millions of people”

0.52

The new agent paradigm differs from text completion products because models take actions in the real world, looking up websites, running code, reading logs, and iterating on feedback, the way humans do rather than producing a perfect paragraph in one pass.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“models can actually act and make decisions in the real world”

0.52

Building Devin requires integration with all of a team's tools, including GitHub, Jira, and Confluence, and continual learning of the specific codebase, because a new engineer, however smart, is unproductive compared with someone who has been on the codebase for years.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“Solving problems in a vacuum is great, but Devin has to work in your GitHub.”

0.52

Fixing a bug is a multi-step decision sequence: reproduce the bug locally, inspect logs and variables, read relevant files, edit, retest, and add a unit test, and an agent must choose actions that maximize the chance of success across many steps rather than produce a single answer.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“How do you use the tools available to you and make a sequence of decisions that maximizes your chance of success?”

0.52

The toughest category of software businesses is those reliant on switching costs and lock-in, because if agents make migration and implementation nearly free, competition will shift to who has the better product.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“I think the toughest category to be in is business that's heavily reliant on switching costs and lock-in.”

0.52

The practical question is not whether AI reaches 95 or 97 percent of a human's workflow but whether humans are actually doing about 25 times more because of it, so distribution and product integration drive GDP and quality-of-life effects.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“It's more a question of are humans actually out there doing 25 times more because of the AI technology.”

0.52

Beyond model capability, the hard infrastructure for Devin includes its own persistent machine, rolling back machine state, fast spin-up of new instances, navigating past decisions, and running many Devins in parallel, which are messy problems that come from building for real environments.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“there's so much with the infra and with the whole experience that gets quite involved”

0.52

Wu's first company Lunch Club, an AI-driven professional networking service, taught him that common startup advice such as going fast, staying close to customers, never compromising on recruiting, and building what people love is almost tautological, yet there is always another gear to push with those fundamentals.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“there's always another gear that you can kick into with your fundamentals”

0.49

Many computing technologies were initially interesting but only became valuable once a use case was found: ENIAC's ballistics calculations, then the US census as the first real valuable application, the Macintosh's early graphic design use, and the spreadsheet as the killer app.

factualestablishednovelty 2/4durability 3/4· Scott Wu

“The first real big thing that really caused that to be valuable was the U.S. census.”

0.49

Devin uses usage-based pricing through Agent Compute Units, roughly 8 to 12 dollars per hour of work, because agent tasks consume far more inference, and the intended value-based proxy is that Devin should make projects about 10 times cheaper than doing them manually.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Scott Wu

“we want the cost of Devin to run on those to be 10x less”

0.49

Within five to ten years, a person with no computer science background should be able to specify a precise personal app in natural language and get it built with high fidelity without relying on human programmers, because software is currently too expensive to build for individual needs.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Scott Wu

“I think five to 10 years is the reasonable realm for that”

0.47

GitHub Copilot-style autocomplete saves an engineer the time of typing a predicted line, but it does not save the time of deciding what the correct line should be, because the human still judges correctness.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Scott Wu

“It doesn't save me the time of thinking of the line because obviously it's still me that's deciding what's correct”

0.46

Roughly half the time, Devin's output on a reported bug, crash, or feature request is a pull request ready to merge; the other half requires coaching Devin, touching it up, or starting over, but tagging Devin still gives a free first pass.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Scott Wu

“probably about 50% of the time, Devin's output on one of these is just like a pull request code that's straight out of the box ready to merge”

0.46

Setting up MongoDB locally took iterative cycles of commands, errors, and searches, and Devin resolved it by checking ports, schema, and socket connections, then iterating on each new error, which Wu says is the core difference from a Stack Overflow style answer.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Scott Wu

“Devin sees the error and says, Okay, let me see what's going on. Let's check what ports are open.”

0.42

Scott's fast solutions to the card game come from memorized patterns and intuition built over time, such as knowing that 11 times 15 equals 165, so the solver searches for combinations that hit the target.

causalestablishednovelty 2/4durability 3/4· Scott

“A lot of it is an intuition, I guess, that you build over time.”

0.35

Competitive programming at the IOI is fundamentally algorithmic problem solving, where code is only the medium and success depends on understanding the right approaches to problems.

definitionestablishednovelty 1/4durability 3/4· Scott

“at the core of it, it's really just problem solving.”

0.34

AI should free humans from execution so they can focus on ideas and imagination, just as Minecraft creative mode removes survival constraints like resource scarcity and lets players build what they imagine.

normativespeaker onlynovelty 2/4durability 3/4· Scott Wu

“it's basically just focusing on the ideas and the imagination rather than the execution”

0.31

The pace of technical progress will run far ahead of real-world adoption in fields such as medicine and law, creating a gap between what is possible and what has reached everyday life.

forecastspeaker onlynovelty 2/4durability 2/4· Scott

“the pace of progress is going to be at such that technology in some place is going to be way ahead of reality.”

0.31

Large platform companies such as Microsoft, Amazon and Google spend billions of dollars each year migrating customers onto their platforms, so automating migration and modernization represents an enormous market for AI software businesses.

factualspeaker onlynovelty 2/4durability 2/4· Scott Wu

“It's billions of billions of dollars every year.”

0.31

Living and working together in a house creates shared late-night energy and tight coordination that a conventional office would not replicate, which Scott sees as part of why the team is effective.

causalspeaker onlynovelty 2/4durability 2/4· Scott Wu

“I feel like it would wouldn't be quite the same at an office.”

0.30

Wu spent his first 17 years focused on the International Olympiad in Informatics after his older brother introduced him to it in Baton Rouge, and the competition is fundamentally about problem solving with code as the medium.

factualestablishednovelty 1/4durability 2/4· Scott Wu

“it was the number one focus of my life for the first 17 years of my life”

0.30

Self-driving cars were among the first real-world AI agents, making decisions and iterating on feedback, and their long timeline from the 2014 Waymo era shows how slowly agents can reach full deployment when reliability is demanding.

factualestablishednovelty 1/4durability 2/4· Scott Wu

“I would almost argue one of the first agents, if you want to call it that, actually was self-driving cars.”

0.30

Wu believes he is largely the sum of the mentorship he received, and that the kindest acts often build more together through demanding coaching rather than comfort, in both competitive math and startups.

normativespeaker onlynovelty 1/4durability 3/4· Scott Wu

“I feel like I'm really only the sum of all the mentorship that I've received”

0.27

Scott's achievements are the sum of the mentorship he received, and the greatest kindness is building more together through demanding, honest coaching rather than sacrifice.

normativespeaker onlynovelty 1/4durability 3/4· Scott

“I feel like I'm really only the sum of all the mentorship that I've received”

0.26

Beyond a legal Q&A bot, the logical progression in each vertical is a full agent that acts like a lawyer or teacher, and in software it is an engineer you work with rather than a text autocomplete.

forecastspeaker onlynovelty 1/4durability 2/4· Scott Wu

“the next step is to have like a full teacher”

0.26

Pre-existing friendships among Cognition's leaders make tough conversations easier, because everyone knows the intent is to take care of each other and build something great together.

normativespeaker onlynovelty 1/4durability 2/4· Scott Wu

“it's honestly the other way around where it's easier to have tough conversations because, you know, obviously we were friends before all this”

0.26

Senior executives at large technology companies are asking how to manage engineering headcount, based on the view that AI will not replace workers but someone using AI better than them might, and that the shift will come faster than most people expect.

factualspeaker onlynovelty 1/4durability 2/4· Scott Wu

“people like to say, this AI won't replace you, but someone using AI better than you might.”

0.26

The AI market is still greenfield, with each use case such as data observability, incident response, testing, modernization, and migration supporting enormous businesses, and AI will make these categories orders of magnitude bigger, as Datadog and PagerDuty already show.

forecastspeaker onlynovelty 1/4durability 2/4· Scott Wu

“I think it's still very greenfield.”

0.26

Because bug reports must be filed anyway, tagging Devin at the end of the report adds almost no cost and yields a free first-pass attempt, which is why asynchronous workflows are easy to adopt.

normativespeaker onlynovelty 2/4durability 2/4· Scott

“you had to report the bug anyway. And so you might as well just tag on a Devin at the end and then you get a first pass for free.”

0.26

Teams would rather use autonomous agents for work they would otherwise delay, and Devin's value is less about flawless output and more about moving work forward while humans do other things.

normativespeaker onlynovelty 2/4durability 2/4· Scott

“a lot of these little things which you would want to delegate and spread off”

0.26

Base AI models' math problem-solving is already shockingly good and would place among the upper echelons of humans, though not yet number one in the world, and the important shift is whether this intelligence is used by everyone to create value.

factualspeaker onlynovelty 2/4durability 2/4· Scott Wu

“The math problems that these base AIs could solve already are honestly shockingly good”

0.25

During a last-minute launch, Devin built much of the onboarding flow for Slack Connect because the team needed extra bandwidth, showing agents being used to ship their own launch.

factualspeaker onlynovelty 2/4durability 1/4· Scott Wu

“Devin built a lot of that because obviously we needed the extra bandwidth from Devin in the moment”

0.24

Pre-launch load tests can fail the night before a launch, forcing last-minute changes, and startups must accept some embarrassment to ship.

factualspeaker onlynovelty 1/4durability 2/4· Scott

“we did our load testing, this shit didn't work at all.”

0.23

Devin is currently the only product in the market doing this kind of autonomous software engineering, but many competitors are expected soon, which Scott views as good for consumers.

forecastspeaker onlynovelty 1/4durability 1/4· Scott Wu

“I think right now Devin is really the only one out there in the market that's doing that. But as you're saying, I'm sure there will be lots and lots of competitors pretty soon even.”

0.23

Devin functions today at the level of an entry-level junior software engineer, having been roughly a high school CS student level at launch and an intern about six months earlier, so it is not suited to re-architecting an entire codebase but to delegated, well-scoped tasks.

factualspeaker onlynovelty 1/4durability 1/4· Scott Wu

“It's very much a junior engineer today. That's how we think of it.”

0.21

Engineers at Cognition kick off small refactors and type-safety fixes to Devin from the IDE while continuing their own project, so small distractions get handled asynchronously rather than breaking focus.

factualspeaker onlynovelty 1/4durability 2/4· Scott Wu

“you just kick off Devin from the IDE for each one of these”

0.21

The 24 card game builds number sense because it rewards creative reasoning about numbers rather than mechanical arithmetic, and experienced players memorize about 1300 combinations to solve harder variants.

factualspeaker onlynovelty 1/4durability 2/4· Scott Wu

“There's only about 1300 total combinations that you have to know.”

0.18

Devin's default plan for engineering teams is 500 dollars per month, with usage billed on top.

factualspeaker onlynovelty 1/4durability 1/4· Scott Wu

“The plan is 500 bucks a month for engineering teams by default.”