Comment by raynr
19 hours ago
1. The funnel is even steeper than the linked post suggests.
There's a huge range of users between "casual and infrequent users" and "professional use of frontier-grade LLMs in math & code".
I think I'm in between there somewhere. I'm far from a software dev, my experience is playing with delphi and C when I was a kid. But I've been using LLMs to vibe code personal and work projects that I could never have done even a few years ago. They're genuinely helping work, because there's always some place where a simple little script or other utility can help. (Hell, isn't this partly why so many people start up startups?)
I imagine that many departments in many companies that are not "tech" companies face similar challenges - there's no budget to hire a software dev or even engage a third party to do things, but now because everyone's aboard the AI train, they have access to something (chatgpt, gemini, copilot, claude), and that's helped users implement their own solutions. Somewhere in the world, someone's excel-file-based accounting system has gotten an upgrade with help from an LLM.
2. The funnel is also driven by the the fact that users need products, and an LLM itself isn't a product, it's a capability.
Software devs are at the front lines because they can build their own products that can integrate LLM's capabilities. You already know how to automate testing, orchestrate, work together in a team. You're writing / configuring harnesses and flows that work for you. You're sharing / hoarding skills files.
The rest of us, like me, don't have that skill. Copilot the product is a mostly useless chatbot that takes up space in MS word. I guess its excel cleanup capability is semi useful if your company runs on excel and you've never used a filter before.
There is no work I can put an LLM to other than a "writing aid" or "better search", because the only AI product we have is the chatbot and maybe someone's equivalent of an agent studio, none of which we are trained to use.
Until the products come, there'll be a funnel.
It's honestly shocking how terrible copilot is at generating interesting and relevant metrics from excel.
The local optimum still seems to be: Let LLM of your choice generate a polars script that reads the spreadsheet and explores them for you.
My partner recently used that "pattern" at work to find outliers and interesting metrics in years of excel reports. Data they were working hours for. Now available in a single human language query.
I don't think the HN crowd understands how many million jobs are primarily staring at excel reports and copying values between forms, pdfs and spreadsheets.