Essay: 'AI sharpens craft' argument
An essay announcing the rebrand of "Beyond the Build" to "Primitives" argued that AI sharpens the craft of building software—helping teams iterate faster—rather than replacing human judgment. The piece frames AI as a tool for producing more options and increasing iteration, not as a substitute for strategic decision‑making. (primitives.blog)
A new essay launching the blog name Primitives argues that artificial intelligence changes software work by increasing iteration, not by replacing human judgment. (primitives.blog) The post appears at Primitives.blog under the URL slug “01-a-fresh-start-building-with-ai,” and it presents the rename from Beyond the Build as part of a broader editorial shift toward building software with artificial intelligence in the loop. (primitives.blog) Its core claim is simple: if models can produce drafts, variants, and working code faster, the scarce part of software work becomes choosing what to build, which version to keep, and what tradeoffs a team will accept. (primitives.blog) That argument lands as coding tools are being sold less as autocomplete and more as collaborators that can edit files, run tests, and work across a codebase. OpenAI said in August 2024 that SWE-bench Verified is a human-validated subset of 500 real software issues for measuring how well models resolve engineering tasks. (openai.com) Model makers now use those coding tests to market practical software help. OpenAI said GPT-5 scored 74.9% on SWE-bench Verified, while Anthropic said Claude 3.7 Sonnet reached state-of-the-art performance on the same benchmark when it introduced the model on February 24, 2025. (openai.com; anthropic.com) The productivity pitch is older than the benchmark race. In a Microsoft study on GitHub Copilot, developers with the tool completed a JavaScript server task 55.8% faster than a control group, and GitHub later reported survey and activity data linking Copilot use to higher perceived productivity. (microsoft.com; github.blog) The essay’s position is narrower than “artificial intelligence writes the software for you.” It treats the model as a machine for generating more possible moves, while people still decide product direction, architecture, quality bars, and whether an output is safe to ship. (primitives.blog) That view also lines up with how benchmark creators describe the limits of the tests. SWE-bench measures whether a model can resolve repository issues under controlled conditions, not whether it can set company strategy, negotiate requirements, or own the consequences of a bad release. (openai.com; swebench.com) The rebrand to Primitives puts that bet in the name: software teams may get more raw building blocks from models, but the craft the essay defends is still the human work of selecting, combining, and rejecting them. (primitives.blog)
