Summary

Microsoft’s GitHub acquisition worked because it invested in the acquired platform before extracting synergies from it. The three principles were: keep developers first, ask how Microsoft could accelerate GitHub, and only later ask how GitHub could accelerate Microsoft. GitHub announced in July 2024 that it had surpassed a $2 billion run rate, leaving its revenue and growth scorecards “more than green.”
Copilot’s advantage is becoming workflow-wide distribution, not merely better autocomplete. It began in mid-2020, previewed in June 2021, reached general availability in June 2022 before ChatGPT, and now has 15 million users; its code-review agent has reviewed more than 8 million pull requests. Dohmke’s platform call is a continuum spanning completion, chat, agent mode, MCP, pull-request review, vulnerability fixes, and Project Padawan’s issue-to-PR workflow.
The headline 55% productivity gain is real but clinically narrow, while AI simultaneously makes production software more complex. GitHub assigned the same task to 50 developers with Copilot and 50 without, producing the 55% result; real teams never build the same thing twice, so the baseline continually moves. AI simplifies code generation but adds models, evaluations, post-training, monitoring, and failure modes—the objective is helping teams manage ever-growing codebases and complexity, not simply generating more lines.
Dohmke expects agents to automate heavily without removing the need for engineers to understand systems and code. He put the best SWE-bench result at 62% or 63%—and argued that even a 65% score across 2,000 issue–pull-request pairs from about a dozen Python repositories leaves a large gap. He thought the new multilingual version pushed scores into the 20s and 30s. At scale, agents can still change two requested lines while damaging 5,000 others; “that back and forth between the agent and its output and the developer is going to stay crucial for many years to come.”
Programming languages remain the deterministic boundary beneath inherently ambiguous natural-language instructions. Two people can describe the same interface and build completely different products, while code ultimately maps to processor instructions. The scarce skill therefore moves upward into “systems thinking”: choosing architectures, decomposing problems, validating generated work, and knowing when one line of hand-written code beats an expensive prompt.
AI coding adoption has crossed the chasm, but latency and workflow design will determine who captures durable usage. GitHub ships a new completion model roughly monthly and tracks “accepted and retained characters”; it also targets accuracy, lower latency, efficiency, and reduced GPU use. Dohmke calls latency “the core metric for every developer tool.” Watching an agent work for 15 minutes breaks flow, so the winning interface may let developers keep building while asynchronous agents quietly write tests, review code, or repair failures.
AI will reshuffle developer performance and blur the boundaries among engineer, product manager, designer, operator, and marketer. GitHub now treats appropriate Copilot and AI usage as a cultural expectation, while recognizing that forcing a disliked workflow does not create adoption. Dohmke expects “full-stack builders” to conduct orchestras of agents and predicts very small companies with very high valuations—a chance to “build a business in your garage all over again,” provided people still understand what the machines create.

Deep dive

Not yet available upstream; scheduled sync will retry.

Episode 144 - June 6, 2025 Bill Gurley - The Gift and The Curse of Staying Private - [Invest Like the Best, EP.427]