Korva is a scroll of the internet’s best ideas.
AI needs an Excel-style expert
The report that can't be built
Ask Claude or ChatGPT to search for one person's investments and email a report every week, and it can't do it directly. It offers to build an agent that gives alerts, then hands back a JSON file or code, so the user has to know how to program. What comes out is slop, and fixing it takes correcting and reiterating.
The Excel parallel
Advanced use feels like figuring out PowerPoint or Excel used to. Every office had its Excel expert or its PowerPoint person, and companies grew around the gap, like one that sold nothing but PowerPoint templates to download and redo. AI is not even that advanced once the second level starts.
Agents drift
Agents also drift. As the underlying large language model changes, the way an agent was originally programmed no longer matches what the model has become. Someone with a basic computer background can walk into a small, medium-size, or large business and fix those failures, and what AI can do once they are fixed is phenomenal.
Why do agents drift when models change?
An agent's prompts and tool setup are tuned to how one version of a model behaves. When the provider updates or replaces that model, it may read the same instructions differently. A 2023 study by Stanford and Berkeley researchers found that the behavior of GPT-3.5 and GPT-4 changed substantially between March and June of that year on the same tasks.
Why did every office need an Excel expert?
Spreadsheets let people without programming training build models, but formulas, pivot tables, and macros carried a real learning curve. The person who mastered them became the office's go-to. Microsoft added the Visual Basic for Applications scripting language to Excel with version 5.0 in 1993, which widened that gap further.
Why do chatbots hand back code instead of results?
A chat window can only produce text, while a weekly emailed report needs a place to run, a schedule, and stored credentials. The model can write the script but cannot run it for the user. Anthropic introduced the Model Context Protocol in November 2024 as a standard way for models to connect to outside tools.
More from All-In with Chamath, Jason, Sacks & Friedberg
