GitHub CEO Thomas Dohmke has stated in a personal blog that the most advanced developers have “moved from writing code to architecting and verifying the implementation work that is carried out by AI agents.”
The post follows one from March when Dohmke repeated Anthropic CEO Dario Amodei’s assertion that “in a short time, 90-100 percent of all code will be written by AI.”
Referencing recent interviews with 22 developers using AI tools, Dohmke identified four stages to AI developer maturity, from skeptic, to explorer, to collaborator, and finally strategist. In this last stage, developers no longer write code, having delegated that task to AI agents, but focus on refining prompts and reviewing and validating the generated implementation.
Dohmke acknowledges that AI output must be scrutinized and suggests that developers need to understand programming basics, algorithms, and data structures, and that they continue to perform “manual code review.” How they will acquire these skills in future is less certain, since he sees computer science education changing to be about understanding systems, debugging AI-generated code, and expressing ideas clearly both to people and to large language models (LLMs). Abstraction, decomposition, and specification must be taught not as pre-coding steps, but as “the new coding,” he wrote.
It is a rose-tinted view of the future of AI-driven coding, and one that Dohmke no doubt hopes will persuade businesses of the value of GitHub Copilot subscriptions.
More nuanced views of AI coding include issues such as research showing that it can erode code quality. There are also new security issues such as those listed by OWASP (open source foundation for software security), which includes prompt injection, sensitive information disclosure, supply chain compromise, and data or model poisoning.
There is also the open question of what happens when AI models are trained on the output from AI models. In 2023, research from Rice University and Stanford University suggested that “without enough fresh real data in each generation of an autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease.”
That downsides exist is not a reason to reject AI coding, but does suggest that a rapid move towards AI dominance in this field may be a mistake.
In the meantime, the Microsoft-owned GitHub platform appears to be getting worse, according to a discussion on Hacker News. “GitHub’s performance has been rapidly degrading ever since they started rewriting everything in React,” said one developer.
Another claimed: “I worked at GitHub for three years and they are very aware that slowness is a big issue throughout the whole product. There was a year-long cross-team effort to improve things but the main goals were not achieved.”
Developers will hope that the drive towards Copilot and AI features is not detrimental to the resource dedicated to maintaining the core GitHub platform. ®