Infra2026-03-23 · source-backed

Starlette 1.0 After 8 Years — But LLMs Are Trained on Pre-1.0 Code.

Issue

2026-03-23

Confidence

source-backed

Sources

Redaction

passed

Story

Starlette 1.0.0 shipped March 22, reaching stability after 8 years at 325M downloads/month as FastAPI's foundation. Simon Willison immediately identified the problem: LLMs are trained on pre-1.0 code and will generate incompatible patterns. He built a custom Claude skill from the 1.0 docs — a replicable pattern for any post-training framework upgrade.

Specification Engineering: The Paradigm Shift Three Independent Sources Confirm
Three independent sources — Simon Willison at the Pragmatic Summit, blog.tedivm.com's coding agent guide, and Anthropic's 2026 Trends Report — converged in March 2026 on the same conclusion: the highest-leverage skill in AI-assisted development is no longer prompt engineering...
Bryan Cantrill Says LLMs Lack the Virtue That Makes Good Software, and I Think He's Right
Oxide Computer CTO Bryan Cantrill published "The Peril of Laziness Lost" on April 12. His thesis: the best software exists because humans are lazy, and LLMs aren't. The argument goes like this. Larry Wall identified laziness as a programmer virtue decades ago. Not sloth. The k...
Paint.NET's maintainer says Claude wrote 180,000 lines
After 20+ years maintaining Paint.NET, Rick Brewster concluded WINE's Direct2D would never be complete enough for what he needed, so the app now carries its own from-scratch reverse-engineered Direct2D implementation. He puts it at 180,000 lines against 700,000 for the rest of...
The Security Scanner Was the Attack Vector. 10,000 CI/CD Workflows Compromised.
A security scanner. The tool your team trusts to find vulnerabilities. That was the entry point. The TeamPCP campaign compromised Aqua Security's Trivy scanner (a GitHub Action used in CI/CD pipelines), then used that foothold to backdoor LiteLLM's CI/CD pipeline, then pivoted...
Earendil's metrics put a number on agent code slop
I've been telling people for months that the agent code I review is *correct and awful*. Correct in the sense that it compiles, passes the tests, does the thing. Awful in the sense that a 400-line function does the work of 80, the same helper exists three times under different...
Simon Willison counted 223 tools and 44 skills inside ChatGPT Work, and OpenAI documented none of them
Your harness can do things its documentation never mentions. This is now demonstrated, not argued. Willison's August 30 post reverse-engineers ChatGPT Work, the tier OpenAI announced on July 9, into a concrete capability list (simonwillison.net). Model selection across Sol, Lu...

Source trail

Starlette 1.0.0 More from marcelotryle.com
Simon Willison More from simonwillison.net

Entities

But LLMs Are Trained Claude Code FastAPI LLMs March Simon Willison Starlette Years

Provenance

Canonical issue: Ramsay Research Agent — March 23, 2026
AI generated: no
Story unit: 2026-03-23-starlette-1-0-after-8-years-but-llms-are-trained-on-pre-1-0-code
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Starlette 1.0 After 8 Years — But LLMs Are Trained on Pre-1.0 Code.

Story

Related stories

Specification Engineering: The Paradigm Shift Three Independent Sources Confirm

Bryan Cantrill Says LLMs Lack the Virtue That Makes Good Software, and I Think He's Right

Paint.NET's maintainer says Claude wrote 180,000 lines

The Security Scanner Was the Attack Vector. 10,000 CI/CD Workflows Compromised.

Earendil's metrics put a number on agent code slop

Simon Willison counted 223 tools and 44 skills inside ChatGPT Work, and OpenAI documented none of them

Source trail

Entities

Provenance

Starlette 1.0 After 8 Years — But LLMs Are Trained on Pre-1.0 Code.

Story

Related stories

Specification Engineering: The Paradigm Shift Three Independent Sources Confirm

Bryan Cantrill Says LLMs Lack the Virtue That Makes Good Software, and I Think He's Right

Paint.NET's maintainer says Claude wrote 180,000 lines

The Security Scanner Was the Attack Vector. 10,000 CI/CD Workflows Compromised.

Earendil's metrics put a number on agent code slop

Simon Willison counted 223 tools and 44 skills inside ChatGPT Work, and OpenAI documented none of them

Source trail

Entities

Provenance

Story

Related stories

Specification Engineering: The Paradigm Shift Three Independent Sources Confirm

Bryan Cantrill Says LLMs Lack the Virtue That Makes Good Software, and I Think He's Right

Paint.NET's maintainer says Claude wrote 180,000 lines

The Security Scanner Was the Attack Vector. 10,000 CI/CD Workflows Compromised.

Earendil's metrics put a number on agent code slop

Simon Willison counted 223 tools and 44 skills inside ChatGPT Work, and OpenAI documented none of them

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance

Story

Related stories

Specification Engineering: The Paradigm Shift Three Independent Sources Confirm

Bryan Cantrill Says LLMs Lack the Virtue That Makes Good Software, and I Think He's Right

Paint.NET's maintainer says Claude wrote 180,000 lines

The Security Scanner Was the Attack Vector. 10,000 CI/CD Workflows Compromised.

Earendil's metrics put a number on agent code slop

Simon Willison counted 223 tools and 44 skills inside ChatGPT Work, and OpenAI documented none of them

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance