Big on LLMs Since GPT Started Making Waves

2026-01-31

I’ve been big on LLMs since GPT started making waves. They were obviously good enough to do great things when the tooling, ecosystem, and the users caught up with it, even if intelligence stalled, which of course it didn’t. Thing is, the tools have been slow to catch up (how is Copilot so painfully unaware of its own capabilities?), and by “ecosystem” I largely mean optimized documentation and ways of maintaining it, which I figured would be a lot more automated by now than it is. Well, the latest generation of LLMs are good enough to rapidly build tools to help with that automation. This month I built tools to introspect our company data at an unprecedented scale. Granted, it was far more more possible because I’ve been growing a data warehouse for years spanning dozens of separate systems (timecards and payroll, finance, forecasting, job progress, safety, vehicle tracking, fleet maintenance, invoicing from multiple major vendors, M365 ecosystem, hardware inventory, usage metrics, software change logs, and plenty more) but even if I hadn’t done that, the tools now could rapidly tap me into APIs to get me half way there. The thing is, the tools layer. First, use AI to build a general purpose tool for analyzing databases, pulling usual facts like numeric ranges, null percentages, unique values, etc. Broaden it to search for candidate keys to match on, source comments, select random samples, check for quality issues like inconsistent casing, precision, or whitespace. Then, link AI to the tool it just built, have it scour the warehouse, and build out documentation for itself, explaining what it’s learned. Spend a couple hours reading through the results with voice transcription turned on, remarking on things it misunderstood, or should just double-check, and adding some colour around the why and how of our businesses, common abbreviations, key players, organizational structure, etc. Last step, give it a tool to log undesired outcomes, so if it gets a question it can’t answer, it clarifies what the question was and what it would need to resolve it. Refine the documentation for a couple weeks… and it works. Link this tool into our company’s internal ChatGPT-style tool (permissioned appropriately for each user, which may well be the hardest part!), and it can answer questions, complete with actual database queries you can then verify, modify, and make repeatable. I use it for one-off questions. I use it to make people reports to use long-term. I use it for software development (“replace this hard-coded dropdown with all the values you’ll find in table X”). It’s completely blown past my hopes for it. It’s shocking what it can answer. Everything from “Who sits next to Matt?” to “how do our ticket response times compare to industry norms?” to “Propose where we could cut 5% from the X budget” to “Look for irregularities in our Microsoft licensing” to “Flag common engine fault codes on our heavy equipment” to “Compare incident investigations to daily work plans and point out anything people seem to be failing to plan for”. I’m pretty confident we are WAY ahead of the curve on this stuff internally now, but at the same time the rate of acceleration is picking up, and I could imagine AI seeing patterns of success stories in the news, getting better still at coding, and this stuff being in reach for many others in a year. AI is getting really good. November 2025 will likely be the last month I write large blocks of code by hand. Knowing how to code still makes me more effective, but in another year, I’m not sure. Still, there’s cool stuff to be done, and I’m not too nostalgic - for every opportunity to impress myself for being clever, there was previously 10x the time spent catching myself up on boring documentation for something that wasn’t interesting at all, but couldn’t be intuited. Coding is fun, but so is designing, and so is grand scale thinking.