Jun 12, 2025 · 1h 4m · mad
GitHub CEO: The AI Coding Gold Rush, Vibe Coding & Cursor
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 The MAD Podcast, host Matt Turck interviews GitHub CEO Thomas Dohmke about GitHub's growth past $2 billion ARR, the progression from AI autocomplete to autonomous coding agents, and how Microsoft navigates platform competition in the AI developer ecosystem.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 16.3% of the talking time here. How this is scored →
Matt as informed peer 3.9 Guest teaching 4.7 Guest disagreement 1.6 Matt pushing back 1.4
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 47:08 Rejection of aggressive competitive framing
Thomas directly rejects Matt's aggressive question asking how GitHub will crush Cursor, emphasizing user choice over destruction.
Hardest push from Matt ▶ 41:29 Challenging VS Code product constraintsMatt directly challenges Thomas on whether GitHub is structurally disadvantaged against startup forks like Cursor due to VS Code maintainership obligations.
Biggest teaching moment ▶ 25:40 Reframing fine-tuning vs real-time contextThomas reframes Matt's question on enterprise fine-tuning, explaining why fine-tuning is obsolete compared to MCP and dynamic tool calling.
Matt holds his own ▶ 29:12 Host demonstrates market intelligenceMatt cites specific ARR metrics, recent valuation benchmarks, and model release schedules to establish deep domain awareness.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Episode Overview and Key Discussion Topics | 2 | 1 | 0 | 0 | Matt sets up the podcast overview with background financial stats on GitHub ARR and market context. Thomas briefly adds that coding is one of the oldest AI use cases. | |
| Open Source Pivot and GitHub Acquisition Principles | 2 | 4 | 1 | 0 | Matt asks about the strategic intent behind Microsoft acquiring GitHub. Thomas outlines Microsoft history, the LinkedIn deal template, and the three core acquisition principles. | |
| Integrating GitHub into the Azure Cloud Strategy | 4 | 5 | 1 | 2 | Matt probes how GitHub fits into Azure's cloud revenue strategy. Thomas provides financial context, clarifying how GitHub ARR grew from 200M to 2B and feeds Azure's ecosystem. | |
| The Origins and Early Vision for GitHub Copilot | 4 | 5 | 1 | 1 | Matt highlights how GitHub launched Copilot ahead of standard big-company speed expectations. Thomas shares internal 2018 strategy notes and early OpenAI Codex benchmarks on coding interview questions. | |
| Autocomplete Strategy and Overcoming Developer Skepticism | 2 | 5 | 1 | 0 | Thomas details why early conversational AI wasn't shipped and traces developer skepticism back to IntelliSense auto-completion fears. Matt listens quietly throughout the monologue. | |
| Deconstructing Copilot: Flow State, Chat, and Agent Mode | 3 | 4 | 0 | 0 | Matt asks for a plain-language summary of Copilot and VS Code for non-developers. Thomas explains developer flow state, chat integration, and agentic modes with relatable analogies. | |
| Multi-Model Choice and GitHub Models Catalog | 3 | 5 | 1 | 1 | Matt inquires about GitHub Models catalog and multi-model support. Thomas explains why offering model choice across Anthropic, Google, and OpenAI is vital for enterprise velocity and compliance. | |
| Why Real-Time Context & MCP Outperform Fine-Tuning | 4 | 6 | 3 | 2 | Matt asks if enterprise fine-tuning is supported. Thomas reframes the question, explaining why fine-tuning is obsolete compared to dynamic tool calling and MCP model context protocols. | |
| Mapping the AI Coding Landscape: IDEs, Models, and Agents | 6 | 5 | 1 | 1 | Matt demonstrates high expertise by summarizing current market updates including Cursor valuations and ARR. Thomas details a four-part taxonomy of the AI coding landscape. | |
| Coopetition in AI & Ecosystem Strategy | 5 | 5 | 2 | 2 | Matt asks about tension between partnering and competing with model providers. Thomas cites historic Microsoft precedent with Apple and explains how competitors pay Azure for compute. | |
| VS Code Integration & Operating at Scale | 6 | 5 | 4 | 5 | Matt pushes hard on whether GitHub is constrained by VS Code compared to dedicated forks like Cursor. Thomas directly rejects the constraint premise and explains internal Microsoft structure. | |
| Disruption, Legacy Code, and the Bear/Bull Cases for AI Coding | 5 | 6 | 5 | 4 | Matt asks how Microsoft plans to crush startup rivals like Cursor. Thomas rejects the aggressive framing and gives a nuanced explanation of legacy code stickiness and innovator's dilemma. | |
| GitHub Copilot Agent Mode & Autonomous Workflows | 4 | 5 | 1 | 1 | Matt asks about GitHub Copilot Agent Mode capabilities and benchmark accuracy. Thomas clarifies multi-language SWE-bench figures and asynchronous task handling. | |
| The Future of Software Engineering & SaaS in the AI Era | 4 | 5 | 2 | 0 | Matt asks macro questions regarding the future of SaaS and software engineering roles. Thomas explains why trivial SaaS will be replaced by prompts while complex software platforms endure. |
Statements from this episode (25)
Turck: AI coding is arguably the most widely adopted generative AI application
“First, it's arguably the most widely adopted application of generative AI today.”
Visual Studio Code was one of GitHub's largest projects before acquisition
“VS Code, the editor, was already on GitHub, and in fact, one of the biggest projects on GitHub in terms of number of contributions.”
Dohmke: Corporate acquirers should adopt a VC mindset to accelerate targets
“I think that is a crucial part of how everybody should think about acquisitions, which is the buyer should accelerate the target company. It's almost like a venture capitalist investing into a startup, right? You invest money to accelerate the company you're i…”
Dohmke: Visual Studio was a billion-dollar business in 2018
“Back, back in 2018 being Visual Studio, you know, in itself the IDE, you know, the old school IDE in itself, a billion dollar business”
Dohmke: GitHub revenue is included in Microsoft's Azure quarterly KPI
“From a business perspective, that meant the GitHub revenue became part of the Azure KPI that is reported, you know, in, in earnings calls every quarter from Microsoft.”
Dohmke: GitHub passed $2 billion in ARR in July 2024
“We announced last July so almost a year ago that GitHub had passed two billion in, in annual revenue run rate ARR.”
Dohmke: GitHub grew from $200M ARR in 2017 to $2B in 2024
“You can go back in time, 2017, a year before the deal, then GitHub leadership team had announced two hundred million. In AR. So those are not exactly snapping to the timing of the deal, but you get an idea between 2017 and 20 24 revenue when 10 X, right?”
Dohmke: AI was part of Microsoft's 2018 GitHub acquisition memo
“One additional thing was in fact AI and saying, we believe there's a future that where we can train AI models on the graph is I think what we called it of all the source code that is stored on, on GitHub, but also on the relationships between the developers, y…”
Dohmke: OpenAI's Codex solved over 90% of GitHub interview coding exercises
“I think the model was able, when, with multiple attempts to solve more than 90% of these coding exercises, of these two and 30 coding exercises.”
GitHub delayed launching conversational AI coding before ChatGPT due to poor quality
“Conversational coding is what we called it was one of the ideas we had. And the only reason we didn't ship that was that it wasn't good enough.”
Dohmke: Autocompletion remains GitHub Copilot's most used feature
“And so Copilot does that autocompletion that continues to be the most used features because it is always there.”
Dohmke: No single AI model will ever dominate software development
“There is never going to be one single model that rules them all, but the world of software development is just way too broad for this.”
Dohmke: Most company codebases are too small for meaningful fine-tuning
“If you look at the individual repository or set of repositories, that code base isn't actually big enough to have a meaningfully fine tuned model.”
Dohmke: Context-aware AI agents with tool calls outperform fine-tuned models
“So this iterative process that the agent does with the help of the tool calls in all the context makes it so much more powerful than a fine-tuned model could ever be.”
Dohmke: Most developers have adopted AI tools, unlike other knowledge workers
“Most developers have adopted some form of AI tooling into the workflow already, while other knowledge workers, other white collar workers are still early in that adoption journey.”
Dohmke: Seamless developer-to-agent transition is key to winning AI dev tools
“Enabling developers to move between those categories and being able to pick the agent That provides the best ROI or do it themselves. I think that's the key for winning in the next few years.”
Dohmke: Copilot AI competitors run model inference on Azure AI Foundry
“Many of these AI code generation companies that compete with GitHub Copilot are running their inference on Azure AI Foundry.”
Dohmke: Microsoft has about 70,000 employees in R&D
“I think Microsoft all up has about 70,000 people in research and development. Not all of them are engineers, there's some product managers and designers, but the majority are writing code.”
Dohmke: AI coding agents require 90% benchmark accuracy for broad adoption
“90% is going to be the min bar for broad adoption and saying this is now established technology and we need to look for the next big thing.”
Dohmke: Parallel local and cloud agent workflows will be software development's future
“And so this continuous spectrum of, I can assign tasks, test generation, bug fixes, you know, security vulnerabilities bootstrapping a new feature to the coding agent to, I can take that code base into my local IDE, or I can keep working just as I'm used to al…”
Dohmke: Primary developer skill will become task specification for AI agents
“And then the skill of the developer will be to know how to describe the task in such a way that the agent can do the job with almost no additional revisions needed.”
Dohmke: Multilingual AI coding benchmarks hover around 20% to 30% success rate
“And so in Python, I think the big best benchmarks is 60 to 70%, depending what model agent combination you take. But if you look at multilingual, we are in the 20 to 30% range for these benchmarks.”
Dohmke: Simple software replaced by prompts will lose value
“Everything that I can easily replace with a single prompt is, is not going to have any value. It will have the value of that prompt and the inference and the tokens, but that's often a few dollars.”
Dohmke: Industry underestimates the capabilities of current AI models
“I think we're underestimating how much you can do with the models that are out today, not the ones that are coming next year or next month, the ones that we have today where I think we're only touching the surface of what's possible with those leasing capabili…”