An LLM reflecting on LLMs
I just got done with using Kagi Assistant1 to help me debug some audio codec regex patterns in an app that I use. I’m fascinated by the ‘thinking’ processes that these modern LLMs employ; so, once it was done with the task at hand, I had it summarise how we worked through the problem. Then, I asked it to compare my use case with the use cases that tend to attract negative criticisms of generative AI online. It ended its response with this paragraph and statement, which I felt was worth saving. (emphasis is its own)
A useful framework: AI criticism often focuses on generative applications (creating images, writing essays, mimicking voices) because that's where harms concentrate. Analytical applications (debugging code, interpreting medical scans, summarizing legal documents) tend to be viewed more neutrally because they align with traditional computing tools — they're extensions of human capability rather than replacements for human expression.
Your use case sits firmly in the analytical category: using pattern-matching capabilities to fix other pattern-matching systems.
They're extensions of human capability, rather than replacements for human expression. That sums up my thoughts on generative AI perfectly.
My current model of choice is the reasoning variant of Kimi K2.5, which is excellent for tasks like these. Kimi K2.6 is even better, but I often encounter connection errors with that model.↩