Add another LLMs post

hugo/content/posts/2025-06-16-llms-redux.md

@@ -0,0 +1,52 @@
+---
+lastmod: "2025-06-16T09:28:00.0000000+01:00"
+author: patrick
+categories:
+- programming
+date: "2025-06-16T09:28:00.0000000+01:00"
+title: LLM effect on my programming (2025 edition) 
+summary: "Since my last post about this in 2024-03, LLMs have become a distinctly positive addition to my ability."
+---
+
+I wrote in a [previous post] that ChatGPT 4 was not clearly having a positive effect on my ability to code.
+
+We're over a year later now, and we have larger models and reasoning models, and those *are* definitely an improvement over not having them.
+
+# As documentation-knower
+
+[Opus 4](https://claude.ai) has memorised how .NET works to a terrifying level of detail.
+In general, I can ask it minutely specific questions about the .NET runtime and it will simply answer them correctly.
+
+I found Gemini 2.5 Pro to be fairly solid as an engine for searching the .NET spec (which I uploaded to its context), although still quite prone to making things up.
+
+# As bug-finder
+
+Opus 4 is also a powerful debugging engine, the first one that I actually trust to tell me the cause of a bug without handholding.
+
+My workflow is:
+
+* Put a whole load of tracing into my code
+* Run the code to obtain the tracing
+* Dump the trace into Opus 4 (using the ordinary chat interface) along with the code (usually from working in a Project, but sometimes also supplying the `git diff` against `main` if I believe that the diff introduced the bug), and ask it what is wrong.
+
+This Just Works alarmingly frequently.
+
+# As sanity-checker
+
+At work, I feed many (non-sensitive) diffs automatically on `git push` through o3-mini-high with a prompt asking for a sanity check, obvious mistakes, etc.
+This catches a bunch of really dumb stuff like typos, and is particularly valuable for the cursed dynamic languages like Groovy where you can't know whether anything is wrong without running it.
+
+I've found o3-mini-high less valuable for "proper" code reviews (the sort that you'd have an engineer do to identify design problems, for example).
+
+# As LeetCode solver
+
+o3 is better than I am at solving well-defined tightly-scoped LeetCode-style problems.
+(Fortunately for me, I basically never have to do this in real life.)
+
+# Wrapup
+
+Many of my points in the previous post are still correct: I still don't trust any of the models to write code for me, and for the same reasons.
+But now the utility of the models in other areas is much higher - I now have what I described in the "What I actually want from such a system" section of the previous post, for example - and I would absolutely recommend a Claude Pro subscription.
+(I'm even considering Claude Max; Opus consumes a *lot* of quota.)
+
+[previous post]: {{< ref "2024-03-27-chatgpt-and-programming" >}}