Skip to content
Main Site News Console

The Past Six Months of Large Language Model (LLM) Development: Understand It in Five Minutes

· Simon Willison Translated
技巧LLM

I put together these annotated slides from my five-minute lightning talk at PyCon US 2026, using the latest version of my annotated presentations tool.

Simon Willison - simonwillison.net

PyCon US 2026 lightning talk ” style=“max-width: 100%” />

#

I gave this lightning talk at PyCon US 2026, trying to summarize six months of LLM progress in just five minutes.

The November inflection point
#

Six months is a pretty good window because it includes what I call the November 2025 inflection point. November was a pivotal month for LLMs, especially for coding.

The title of 'best' model changed hands 5 times
between Anthropic, OpenAI and Google
#

To start with, the model generally considered the “best” one — largely depending on subjective judgment — changed hands five times between the three major providers.

Generate a
pelican riding a bicycle
#

As usual, I use my “generate an SVG of a pelican riding a bicycle” test to help illustrate the differences between these models.

Why this test? Because pelicans are hard to draw, bicycles are hard to draw, pelicans don’t ride bicycles... and no AI lab is likely to have trained a model specifically for this absurd task.

Five pelicans, one each for the models below. The quality varies!
#

At the start of November, the widely accepted “best” model was Claude Sonnet 4.5, released on September 29. It drew this pelican for me.

By November it had been overtaken in sequence by GPT-5.1, Gemini 3, and GPT-5.1 Codex Max, before Anthropic reclaimed the crown with Claude Opus 4.5.

Of those, I think Gemini 3 drew the best pelican, but pelicans aren’t everything. Most practitioners would agree that Opus 4.5 kept the top spot for the next few months.

Coding agents got good
#

This took a little while to become obvious, but the real November story was that coding agents got good.

OpenAI and Anthropic had spent much of 2025 doing reinforcement learning from verifiable rewards to improve how well their models wrote code, especially when used through their Codex and Claude Code agent frameworks.

By November, the results of that work were becoming obvious. Coding agents crossed a quality threshold from “often useful” to “generally useful” — good enough that you could rely on them as day-to-day tools for real work, instead of spending most of your time cleaning up their stupid mistakes.

This is a copy of the MIT license” style=“max-width: 100%” />

#

Also in November, this happened — the first commit to what was then a fairly obscure repository called “Warelay”, by someone named Pete.

December/January
(a little bit of LLM mania)
#

Over the holidays, from December into January, many of us used the break to try out these new models and coding agents and see what they could do.

They could do a lot! Some of us got a little too excited. As I started kicking off ambitious projects to see how far I could push them, I briefly experienced a bit of LLM mania myself.

var numbers = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]; var doubled = numbers.map(n => n * 2); console.log(‘Doubled: ”’, doubled); var evens = numbers.filter(n => n % 2 === 0); console.log(‘Evens: ’, evens); var sum = numbers.reduce((a, b) => a + b, @); console.log(‘Sum:”, sum);

Output 27 Doubled: [2, 4, 6, 8, 10, 12, 14, 16, 18, 20] Evens: [2, 4, 6, 8, 10] Sum: 55 Execution time: 8.00ms About: micro-javascript is a pure-Python JavaScript interpreter with configurable memory and time limits. This demo…” />