The last six months in LLMs in five minutes (simonwillison.net)

Article summary

The article discusses the recent developments in Large Language Models (LLMs) over the last six months, highlighting the significant improvements in coding agents and the emergence of new models. The author notes that the coding agents have become more capable, allowing for daily use without needing to spend most of the time fixing mistakes. The article also mentions the rise of 'Claws', personal AI assistants, and the increasing capabilities of laptop-available models. The author reflects on the rapid progress in LLMs and the need for new benchmarks to measure their performance.

Main themes

• LLM progress
• Coding agents
• Claws and personal AI assistants
• Model benchmarks
• Rapid advancements

What commenters say

• The recent improvements in LLMs have been significant, with some models showing step-function changes in capability, but others argue that the progress is not as dramatic as claimed.
• The 'inflection point' of November 2025 was a real turning point for LLMs, marking a significant improvement in their capabilities, but some commenters disagree on its impact.
• The use of LLMs for coding has become more practical and efficient, with some users reporting significant improvements in their workflow, but others warn that 'flash' or 'fast' models can be worse than useless for coding.
• The emergence of Claws and personal AI assistants has been a major development, with some users finding them extremely useful, but others are concerned about their limitations and potential drawbacks.
• The need for new benchmarks to measure LLM performance is becoming increasingly important, as existing benchmarks are becoming less relevant, and some commenters suggest that the 'pelican on a bicycle' test has exceeded its limits as a useful benchmark.
• Some users have reported significant improvements in their productivity and workflow using LLMs, but others have found that the models can be prone to errors and require careful prompting to produce useful results.
• The rapid advancements in LLMs are leading to diminishing marginal returns in core capability, but can still unlock new capabilities and applications, and some commenters argue that the field is evolving too quickly for anyone to keep track of the changes.