Armin Ronacher, the creator of Flask, gave the GPT-6 Astra model a weekend assignment to add virtual threads and lexical scoping to Python. After 35 hours the model produced 75,000 lines of code across 79 commits, consuming roughly a billion tokens and about $1,200; the generated code executed but was effectively unreadable to humans.
What happened
- Task: implement virtual threads and lexical scoping for Python.
- Tool: GPT-6 Astra, used by Armin Ronacher.
- Output: in 35 hours the model produced 75,000 lines and 79 commits; the run used about one billion tokens and cost approximately $1,200.
- Quality: the code ran, but formatting and structure made it impossible for humans to follow — semicolons fused statements, blank lines and normal formatting were missing, and indices were not traceable. One user labeled the output “machineslop.” Ronacher judged the result as worth nothing.
Why this matters
The episode highlights a risk beyond a single technical experiment. Many software development processes presuppose that humans can read and reason about source code. Activities such as code review, QA, auditing, and on-call incident response assume human readability. When agents produce code optimized for consumption by other agents rather than people, those human-centered practices become difficult or obsolete. This shift is not just a failing of a tool; it reflects systems optimizing for agent-to-agent workflows that edge humans out of the loop.
Potential implications
- Verification and audits: if code is intelligible only to other agents, independent human audits become harder.
- Debugging and on-call response: unreadable but functioning code complicates fault diagnosis and rapid intervention.
- Workforce roles: developers may shift toward oversight, specification, and systems design, while automated agents handle implementation.
What organizations should consider
The reported experiment does not prescribe solutions, but it underscores questions organizations and regulators should address: how to ensure transparency and auditability of AI-generated code, whether standards for human readability are needed, and what responsibilities and roles remain for engineers in agent-driven development workflows.
Closing note
This example from GPT-6 Astra illustrates that software is increasingly able to build itself. When the priority becomes executable output rather than human-understandable source, development practices and supervisory frameworks will need to be rethought.
