Piero Bosio Social Web Site Personale

You can’t add quality into a system through inspection, and you can’t add learning through specification.
9/9

Over time, the codebase becomes Ptolemaic: as the agent adds epicycles, guards, exceptions, and compensations to keep the system passing its proofs.

The result is a Pyrrhic victory: The time saved writing code is instead consumed by specification and verification.
8/9

Since specification and verification can never cover everything, the agent always satisfies the known checks while new issues appear, leading to a specification and regeneration whac-a-mole.
7/9

Debugging moves from fixing code to adjusting specs and tests.

You must anticipate correctness conditions in advance, and creating the specifications and verification is also technical, sometimes even more so than coding.
6/9

With spec-driven workflows, agile tight inner loops become slow outer loops that depend on high-fidelity specifications.

Work shifts from coding to making specs, prompts, lemmas, tests, etc. The agent generates code really fast, but feedback becomes hidden.
5/9

You know how and why things break, start to get a clear picture of the problem domain, and learn to identify patterns, anti-patterns, and code "smells."
4/9

What we've learned from Agile, XP, etc., is that software development productivity comes from small iterations and tight inner loops. Write code, run it, observe behaviour, learn, adjust. This keeps feedback immediate and drives learning.
3/9

The term is inspired by Pyrrhus of Epirus, who, after defeating the Romans, is said to have remarked, "Another such victory and we are lost."
2/9

I want to coin another term, "Pyrrhic Productivity." When AI-coding agents promise speed by removing human effort but end up slowing work by breaking the feedback loops that produce understanding and shifting skilled work from doing to specifying and controlling.
1/9

Tho few would believe it, writing code is often easier than arguing with a chatbot.

Although it might have been hard for sailors to use the new model at first, the fact that it was based on a consistent and logical way of understanding meant that people could improve and develop it over time, like Kepler did.
3/4

Having a more accurate understanding of the system doesn’t automatically make applications based on on that understanding easier to use.
2/4

I received a lot of responses to the Ptolemaic Code discussion, which brought up a good point. The early versions of the Copernican model were actually less practical than the Ptolemaic one.
1/4

It’s absolutely true that AI is reducing jobs, but not like you think. AI is not *doing* those jobs instead of people, AI speculation is redirecting *investment* away from job-creating businesses to capital-intensive businesses.