Because the LLM generated code is custom code written to-order to solve a specific problem Similar to if you hired other programmers and gave them a detailed spec for the code you needed them to write
@simonw
-
Library Dependencies in Modern Software Development Practices
By
–
It's entirely different, because libraries had to be written by other people and you needed to get lucky enough that the exact libraries your unique project needed existed already 99.9% of everything I build is imported libraries once you count Django and Python and C and Linux
-
GPT-5-Codex outperforms in SVG bicycle drawing capabilities
By
–
It's SVG of a pelican riding a bicycle wasn't as good as GPT-5-Codex though, Codex is much better at drawing bicycles
-
Claude Sonnet 4.5 Code Interpreter GitHub Integration Features
By
–
Where it really shines is in their new Code Interpreter mode – I had it checkout my GitHub repo, install dependencies, run tests and experiment with a complex new feature, all prompted from the web browser on my iPhone https://
simonwillison.net/2025/Sep/29/cl
aude-sonnet-4-5/#it-really-shines-with-claude-ai-code-interpreter
… -
Claude Sonnet 4.5 Review: Anthropic’s New Best Coding Model
By
–
Wrote up my initial impressions of the brand new Claude Sonnet 4.5 – I think it may live up to Anthropic's claims of being the "best coding model in the world", for the next few weeks at least!
-
AI Impact on Programming Careers Reassurance Note
By
–
More on my blog, including this hopefully reassuring note for anyone afraid of the impact this will have on their career as a programmer https://
simonwillison.net/2025/Sep/29/ar
min-ronacher-90/
… -
AI-Generated Code in Infrastructure: Armin’s Credible 90% Achievement
By
–
Most of the "90% of code written by AI" claims come from vendors selling AI tools and lack credibility as a result Armin (creator of Flask, Jinja, Click) is different – when he says 90% of a new infrastructure project he's building was AI generated it's worth paying attention
-
Quarantined LLM Systems Modeled as Tool Calls
By
–
Yeah, the quarantined LLM stuff can absolutely be modeled as tool calls instead
-
Video Model Limitations and LLM Improvement Potential
By
–
In both cases you can come up with edge-cases that expose their limitations – with LLMs may or those limitations see overcome as the models improve, will be interesting to see if that happens for video models
-
Chain of Thought vs True Reasoning in LLM Next Token Prediction
By
–
Feels about equivalent to chain of thought in next token guessing LLMs being used for "reasoning" to me