Jan, 2023

Breathless extrapolations like this one are the reason we're about to have a nasty bubble pop.

It's almost halfway through 2025 and AI can generate perfectly fine code from conversational language interactions.

The interactions are where it's at - prompting obviously - but also feedback and clarification after breaking.

I went through the process of trying to figure out what a coding agent that coded for someone who didn't code look like when it first got going and created this: https://github.com/kordless/evolvemcp

It's less than perfect, but works well. I've used it to "auto code" several MCP servers/tools (honestly the MCP language is super confusing) and I can see where an MCP proxy with a search engine attached to it will be super handy later (and if you understand that and build such a thing, you're welcome for the idea if you hadn't had it already).

I see this argument a lot, people are really bad at extrapolating out into the future, but and maybe this is controversial, but I feel like most technologies don’t get _that_ much better as time goes on. I guess it’s like, people assume LLM’s are like the first CPU’s, where they doubled in speed every year, but maybe they’re more like the internet. A little more high fidelity but mostly just used to sell you shit and show you ads.

Written by "Head of AI Systems at Palantir" So I doubt without ulterior motive.