Mar 13, 2025 · 50m · no-priors
No Priors Ep 106 | With GitHub CEO Thomas Dohmke
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of No Priors, GitHub CEO Thomas Dohmke joins Sarah Guo and Elad Gil to discuss the evolution of GitHub Copilot into agentic peer programming, the convergence of software engineering disciplines, AI compute economics, and the enduring vitality of open-source ecosystems.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 19.1% of the talking time here. How this is scored →
The hosts as informed peer 4.9 Guest teaching 4.9 Guest disagreement 1.2 The hosts pushing back 1.4
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 33:05 Rejecting the labor replacement framing with the dishwasher analogy
Thomas firmly dismisses Elad's suggestion that coding agents will command human-labor-equivalent value pricing, insisting customers will not pay labor rates for compute-driven automation.
Hardest push from the hosts ▶ 25:00 Elad challenges tool diversity in the face of base model generalizabilityElad refuses Thomas's premise that developers will always want diverse specialized tool stacks, citing the historical subsumption of Codex into general frontier models.
Biggest teaching moment ▶ 39:29 Thomas corrects Sarah on open source model availability in GitHubThomas educates Sarah after she assumes Copilot only supports closed models, detailing how the GitHub Models catalog and extension ecosystem integrate open weights models.
The host holds their own ▶ 25:05 Elad traces the lineage of Codex and base model generalizabilityElad demonstrates strong industry knowledge by detailing the historical evolution from GPT-3 to Codex to generalized frontier models to substantiate his argument.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Advancing GitHub Copilot to Agent Mode and Project Padawan | 5 | 4 | 1 | 1 | Sarah and Elad kick off asking about Copilot agent mode and Project Padawan. Elad probes the exact technical bottlenecks across frontier models versus UI flow. Thomas outlines reasoning model benchmarks like SWE-bench and key UX criteria like steerability and predictability. | |
| Evaluating Median Programmer Parity and Systems Architecture | 5 | 5 | 1 | 1 | Elad inquires when coding agents reach median developer parity and superhuman performance. Thomas distinguishes high-level systems architecture from localized bug fixes, explaining why human architects remain necessary. Sarah follows up with a question on Copilot eval cycles and internal AI engineering. | |
| Developer-First Culture and Market Competition in Developer Tools | 4 | 3 | 0 | 1 | Sarah asks how GitHub plans to win developer loyalty amid rapid competitive entries in the SWE agent ecosystem. Thomas emphasizes GitHub's internal dogfooding culture and uses a Formula One analogy to praise healthy market competition. | |
| Telemetry Insights and Current Limitations of Code Generation | 4 | 5 | 1 | 0 | Sarah inquires about recent telemetry surprises from Copilot usage. Thomas recounts historical telemetry showing Copilot writing 25% to 50% of code, while noting current agent modes alternate between writing entire applications and getting stuck on trivial UI tasks. | |
| Expanding AI Beyond Writing Code to Review and Remediation | 4 | 5 | 0 | 0 | Elad asks about strategic priorities beyond agentic coding. Thomas details the expansion into AI code review, automated vulnerability remediation, and burning down legacy security debt in cloud environments. | |
| Deterministic Machine Code Versus Natural Language Abstractions | 4 | 6 | 3 | 1 | Sarah asks whether shifting to majority AI-generated code breaks traditional testing and tech debt models. Thomas explicitly pushes back against the premise that all code will be written by AI, detailing the division between deterministic machine languages and non-deterministic natural language. | |
| Converging Disciplines Across Engineering, Product, and Design | 4 | 4 | 0 | 0 | Sarah asks how organizational engineering talent profiles are shifting with AI. Thomas describes the convergence of product managers, designers, and engineers through natural language specifications and tools like Copilot Workspace. | |
| Developer Choice, Pluralistic Stacks, and Multi-Agent Ecosystems | 5 | 4 | 1 | 1 | Elad asks whether a single unified platform will supply all developer agents or if the market will remain fragmented. Thomas advocates GitHub's core philosophy of developer choice, arguing teams will always assemble pluralistic stacks of specialized tools and models. | |
| Generalizability Versus Specialization Across Five-Year Horizons | 7 | 5 | 3 | 6 | Elad challenges the developer choice thesis, arguing that rapidly advancing generalist base models might render specialized developer tooling redundant over a five-year horizon. Thomas counters by comparing general model progress to the long tail of autonomous driving, arguing differentiation will move up the stack. | |
| Eliminating Mimetic Trends and the Rise of Personalized Software | 6 | 4 | 1 | 2 | Elad explores whether removing human mimetic trendiness from tooling decisions leads to rationalized stack choices. Thomas responds that human oversight remains essential for intent verification and predicts a shift toward hyper-personalized bespoke software generated on demand. | |
| Enterprise Adoption Velocity and Return on Investment Metrics | 5 | 4 | 0 | 1 | Elad asks about enterprise traction and Copilot financial metrics. Thomas recaps public milestones including 77k organizations and 1.8M paid seats, explaining how a $20 per month price point delivers undeniable ROI against developer salaries. | |
| AI Pricing Dynamics, Compute Economics, and Labor Economics | 6 | 6 | 4 | 5 | Elad questions whether AI coding tools will transition to value-based labor replacement pricing rather than cheap seat-based subscriptions. Thomas rejects the labor replacement comparison with a dishwasher analogy, arguing pricing will track compute metrics while higher-tier specialized capabilities command premium pricing. | |
| The Trabant Paradox and Software Price Deflation | 6 | 7 | 2 | 2 | Sarah draws an analogy to East Germany's Trabant car scarcity and asks if AI supply abundance causes software value collapse. Thomas corrects the historical waitlist detail to 17 years and explains that while software creation faces deflation, high-value productivity tools can capture significant margin. | |
| Open Source AI Innovation and the GitHub Models Catalog | 4 | 7 | 1 | 0 | Sarah asks about open source versus proprietary models in Copilot, noting open source models seem absent. Thomas educates her on the GitHub Models extension architecture allowing developers to run models like Llama, Mistral, and DeepSeek directly. | |
| The Infinite Game of Software and Preserving Engineering Depth | 6 | 6 | 3 | 3 | When Sarah frames proprietary versus open source as a win-or-lose race, Thomas reframes software development as an infinite game like Minecraft. Sarah then presses on whether abstracted AI development threatens the depth and architectural taste of junior engineers. | |
| Democratizing Coding Education and the Conductor of Agents | 4 | 5 | 0 | 0 | Thomas describes AI coding companions as patient, democratizing tutors for upcoming generations compared to his own resource-constrained upbringing in East Germany, introducing the concept of the developer as an orchestra conductor of agents. | |
| Historical Paradigm Shifts and Optimism for the AI Era | 4 | 3 | 0 | 0 | Sarah asks how experiencing the fall of the Berlin Wall shapes Thomas's perspective on rapid technological disruption. Thomas shares personal reflections on irreversible paradigm shifts across computing history and expresses unyielding optimism for the AI transition. |
Statements from this episode (24)
Dohmke: Copilot will resolve GitHub issues via draft PRs in 2025
“But we think, you know in 20, 25, we get into a place where you can assign a GitHub issue, a well-defined GitHub issue to a co-pilot and then it starts creating a draft pull request and it outlines the plan and then it works through its plan and you can, simil…”
Dohmke: Improved model reasoning will push SWE-bench scores near 100%
“As the models get better in reasoning we're going to get closer to a hundred percent of this VBench, which is that benchmark out of 12 repos open source Python repos a team in Princeton identified 2200 or so issue pull request pairs. Effectively, all the model…”
Dohmke: AI agents remain far from replacing median developers in systems thinking
“I think that's the part where we're far away from agents actually being good enough to take a very rough idea and break it down to small pieces without you as developer or as architect, or even when planning your travel, constantly getting questions back of wh…”
Dohmke: Non-coding GitHub PM used AI agent to merge real pull request
“We actually, you know, when we recorded the demo for the Padawan project we actually had one of our product managers use an issue and the agent create the progress themselves. Right. A PM that usually doesn't code and doesn't write code in the code base was ab…”
Dohmke: AI roadmaps cannot be planned more than two months ahead
“Like the market is moving so fast. We're literally sitting on an exponential curve of innovation where it's hard to keep up and you can't really plan more than a month or two ahead of time.”
Dohmke: GitHub manages all internal functions in GitHub repositories
“Everything that, that we do in the company, including, you know, the, our legal terms and our HR policies and Or product management, sales sales enablement, all these functions are in GitHub issues and GitHub discussions and GitHub repos.”
Dohmke: Open-source AI innovation equals proprietary model innovation
“As much as there's innovation on proprietary models and software, there is as equal amount of innovation in open source.”
Dohmke: GitHub Copilot preview generated 25% of code early on
“Soon after we launched Copilot Preview it already wrote like, 25% of the code.”
Dohmke: Developers spend only 2-3 hours daily writing code
“But if you actually look in the developer life the day to day, you know, in, in, in most companies that's maybe two or three hours of your day that you're actually writing code. And then you're spending an equal amount of time of reviewing code of your coworke…”
Dohmke: Software teams will always require human review before production merges
“And while we don't believe that goes away from a pure security and trust perspective, you always want to have, you know, Another human in the loop before you merge code into production.”
Dohmke: Programming languages remain the essential deterministic layer below AI prompts
“We have the machine language layer, you know, which is Python or Ruby or Rust, right? Those are effectively abstractions of the chipset and the machine instruction set. And that's the last layer that's deterministic, right? Like programming language inherently…”
Dohmke: Developers will maintain legacy PHP and COBOL for a decade
“And so we are still I think for like a decade or so at least going to have software developers that work in, in lots of old school, you know, PHP code and COBOL code and all that stuff.”
Dohmke: AI will converge product management, design, and engineering roles
“I think tomorrow you're going to, as a designer, type effectively the same specification as a product manager and you have, you know, an AI to render the code for the wireframes and then apply, you know, grounding out of your design system to make it look like…”
Dohmke: AI dev tools will remain a multi-vendor ecosystem
“In AI, I think we're going to see the same thing. We're going to see a stack or universe of companies that offer different parts of the software development life cycle, and developers pick the one that you know, they like the most, that they have experience wi…”
Dohmke: Specialized code models were replaced by general base models
“And then we got this model that then eventually became Codex which was this code-specific, you know, version of the model. And today that, you know No longer really exists, right? Like today, everybody sits on top of one of these more powerful base models.”
Dohmke: Developer tool differentiation will come from UX, not models
“And so I think the differentiation is going to come from both, you know, where the developer gets the most The best experience and doing that day to day, right? Like where can I, you know, start my morning, pick up, you know, something I want to work on, explo…”
Dohmke: Completely personalized, natural-language software will arrive in five years
“I think that future that will happen
in the next five years, for sure.
It's just the question is how, how good this job is going to be, and can I just tell it, Springbank is coming up, same hotel, same family, you know, say, and it books me the trip, and the …”
Dohmke: Copilot reached 77,000 organizations and 1.8M paid users
“I think the last number we shared was a few quarters ago, 77,000 organizations using Copilot and back then the number of paid users was 1.8 million paid users.”
Dohmke: Copilot yields 25-28% end-to-end developer productivity gains
“And then we're talking about, you know, 25, 28% productivity gains on, on the end to end 55% or higher on, on the coding task.”
Dohmke: AI software pricing will be compute-based rather than human-salary equivalent
“I think it's going to be compute based or some unit that's, you know, a derivative of,
Compute as a metric.”
Dohmke: AI agent supply is infinite and constrained only by GPU capacity
“Human developers are expensive because there's limited supply agents will have infinite supply that, that will only be limited by the amount of compute capacity GPUs available in data centers.”
Dohmke: Software prices will deflate as certain categories become free
“I think we're going to see deflation of software prices. And so I think it's a mix of both, you know, some things we won't pay for it anymore.”
Dohmke: Future programming languages will be closer to human natural language
“I think there's going to be a next programming language after Python and TypeScript and Rust and Rust in itself, you know wasn't really a thing five years ago. And then, so there's going to be more languages that are probably closer to human language and to be…”
Dohmke: Coding competitions must assume developers are using AI
“As these AI models get better, these competitions of who's the best hacker or coder Are going to have to move to a whole different level where you assume that the developer is using AI to solve the challenges because otherwise it's going to be way too easy.”