Thomas Dunk

CEO of GitHub and advocate for AI-driven software development

Mentions

Appeared in

Discussed in

Key positions and views

The future of software development is a two-layer system where humans specify intent in natural language and AI generates, verifies, and maintains the underlying deterministic code.

AI agents will be priced based on compute consumption, not value-based models equivalent to human salaries, because their supply is theoretically infinite and limited only by GPU availability.

The competitive interplay between open-source and proprietary closed-source AI models is a critical catalyst that accelerates innovation for the entire industry.

While AI is rapidly advancing, its ability to perform complex, high-level reasoning—such as breaking down a vague idea into actionable steps or modernizing legacy mainframes—remains a significant long-term challenge, likely a decade or more away.

AI agents are a crucial tool for enhancing software security at scale by automating the remediation of known vulnerabilities, thereby reducing the persistent backlog that plagues many organizations.

Podcast consensus on Dunk

Points of consensus

▶Thomas Dunk consistently views AI as a transformative force in software development, predicting a future where natural language interfaces and AI agents fundamentally change how software is created, verified, and maintained.Mar 2026

▶He repeatedly emphasizes that AI's primary immediate value is in augmenting developer productivity by automating tedious and time-consuming tasks, such as fixing linter errors, applying formatting, and resolving security vulnerabilities.Mar 2026

▶Dunk consistently frames the evolution of AI as a rapid but incremental process, highlighting both near-term product releases like 'agent mode' and 'Project Padawan' while also noting that more complex tasks like legacy code transformation are at least a decade away.Mar 2026

▶Across multiple claims, Dunk underscores GitHub's strategy of model choice and integration, mentioning support for models from OpenAI, Google, and Anthropic, as well as open-weight models, positioning Copilot as a versatile platform.Mar 2026

Points of debate

▶Dunk presents a vision of users building personalized software via natural language within five years, which contrasts with his more cautious assessment that AI agents are currently 'far from' being able to autonomously break down high-level ideas without significant human guidance.Mar 2026

▶He predicts a future two-layer system where humans use non-deterministic language and AI generates deterministic code, yet the current state he describes involves AI taking specific terminal commands and working on well-defined issues, indicating a significant gap between the present and the predicted future.Mar 2026

▶Dunk describes AI agents as having 'infinite supply' limited only by GPUs, suggesting a commodity, but also details GitHub's extensive A/B testing and fine-tuning process, which implies that effective agent performance is a scarce and carefully engineered resource, not just a matter of raw compute.Mar 2026

▶He highlights massive productivity gains (up to 55%) and adoption (1.8 million paid users), while also noting that even advanced models need improved reasoning to master benchmarks like VBench, suggesting a tension between current market success and the underlying technological maturity required for full autonomy.Mar 2026

Key themes

▶The Agent-Driven Future of Software DevelopmentMar 2026

Dunk outlines a clear trajectory for software development moving from AI-powered code completion to autonomous AI agents. He details GitHub's strategy with features like 'agent mode' and the upcoming 'Project Padawan,' which will allow developers to assign entire tasks to an AI, transforming the development workflow from writing code to specifying and verifying AI-generated solutions.

This focus on agent-based development signals a strategic shift from tools that assist developers to tools that replace entire workflows, suggesting that the value chain will move towards high-level problem decomposition and architectural skills.

▶Pragmatic AI Adoption and LimitationsMar 2026

Despite his bullish long-term outlook, Dunk maintains a pragmatic view of current AI capabilities. He asserts that agents are still far from true autonomy in breaking down complex problems and predicts it will take at least a decade to reliably modernize legacy systems, grounding his futuristic vision in present-day technical realities.

Investors should note this realistic assessment of limitations, as it implies a long and incremental path to full AI autonomy in software engineering, with significant human-in-the-loop roles persisting for the foreseeable future.

▶Economics and Market Dynamics of AI DevelopmentMar 2026

Dunk discusses the business model for AI agents, predicting pricing will be based on compute consumption rather than value-based salaries due to their 'infinite supply.' He also highlights the competitive dynamic between open-source and proprietary models as a key accelerant for innovation across the entire AI space.

The compute-based pricing model suggests that access to and efficiency of GPU resources will be a primary competitive battleground, potentially favoring large-scale cloud providers and hardware manufacturers.

▶Quantifying AI's Impact on Productivity and SecurityMar 2026

Dunk grounds his arguments in specific metrics, citing official GitHub data on significant developer productivity gains (25-55%) and high user adoption (1.8 million paid users). He also frames AI agents as a critical tool for improving software security by automating the process of fixing vulnerabilities, thereby reducing security backlogs.

The ability to provide hard metrics on ROI is a key driver for enterprise adoption of AI tools like Copilot, and the security angle provides a compelling risk-mitigation argument that expands the tool's value proposition beyond pure productivity.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

Post-Preview Launch

Dunk notes that shortly after its preview launch, GitHub Copilot was already writing 25% of user code, a figure that quickly rose to around 50%.

A few quarters ago

Cites official metrics that GitHub Copilot had reached 1.8 million paid users across 77,000 organizations, indicating significant commercial traction.

Recent Announcement

Discusses the recent announcement of an 'agent mode' for Copilot in VS Code, a feature that allows the AI to execute terminal commands and take direct actions to implement features.

2025 (Projected)

Predicts the arrival of 'Project Padawan,' which will enable developers to assign a well-defined GitHub issue directly to Copilot for autonomous resolution via a draft pull request.

Next 5 Years (Projected)

Forecasts that users will be able to build completely personalized software on their devices using natural language interfaces, akin to J.A.R.V.I.S. from Iron Man.

Next Decade+ (Projected)

Believes it will be at least a decade before AI agents can reliably transform complex legacy code (e.g., COBOL on mainframes) into modern cloud applications.

Suggested prompts

How does Thomas Dunk's vision of a two-layer development system (human language and AI-generated code) reconcile with his admission that agents are far from breaking down high-level ideas autonomously? &nearr;What are the strategic risks and opportunities for GitHub in offering models from competitors like Google and Anthropic within its own flagship Copilot product? &nearr;How will the 'infinite supply' and compute-based pricing model for AI agents, as described by Dunk, impact the software development labor market and the valuation of developer skills over the next decade? &nearr;To what extent does GitHub's internal use of Copilot in 'agent mode' for its own feature delivery create a competitive advantage in product development and AI model refinement? &nearr;

Key concepts

AI Agents 1 ep GitHub Copilot 1 ep Future of Software Development 1 ep Developer Productivity 1 ep AI Economics & Pricing 1 ep Software Security 1 ep AI Model Competition (Open vs. Closed) 1 ep Legacy System Modernization 1 ep

Notable quotes

“i no longer install an app from the app store i just you know use a natural language interface um to build all the apps myself and so half as a completely personal software on my personal computer... And so I think that future, that will happen in the next five years for sure.”

Thomas Dunk · No Priors Ep 106 | With GitHub CEO Thomas Dohmke

“we're talking about 25%, 28% productivity gains on the end-to-end, 55% or higher on the coding task.”

Thomas Dunk · No Priors Ep 106 | With GitHub CEO Thomas Dohmke

“we think, you know, in 2025, we get into a place where you can assign a GitHub issue, a well-defined GitHub issue to Copilot, and then it starts creating a draft pull request and it outlines the plan and then it works through its plan.”

Thomas Dunk · No Priors Ep 106 | With GitHub CEO Thomas Dohmke

“I think it's going to be compute based or some unit that's a derivative of compute as a metric... Agents will have infinite supply that will only be limited by the amount of compute capacity GPUs available in data centers.”

Thomas Dunk · No Priors Ep 106 | With GitHub CEO Thomas Dohmke

Report last updated: Apr 28, 2026

Get started free

Back to Entities Intelligence Report