So I'm coding in Claude Code, and to solve my next coding task, Claude actually chooses to use my created MCP tool to solve the problem!
Proud moment as a product creator.
I guess we're now firmly in: Build something AI wants.
Proud to see so many Asians working in frontier labs and kicking ass.
Colleges can play their social engineering games, but when the chips are down, big tech will find what it needs.
Let me know when itโs okay to acknowledge this.
My personal take: GPT 5.5 for planning, Opus 4.8 for actual coding.
I feel Codex is vastly superior at the strategy/planning stage, but I feel the exact opposite with Claude Code on the coding stage.
Great to have them interplay with each other though.
Imagine you have a world-class employee who finishes everything you request in under 10 minutes.
Naturally, you start giving them more tasks. It suddenly becomes a race to keep them busy.
You start multi-threading tasks, giving them task B while they're working on A, and C while they're working on B. You start lining tasks up.
They always keep you on your toes because they need to be fed with a new task every 10 minutes, and every 10 minutes they have made progress, so you can also unlock more progress on your overall goal.
It's addicting and kinda exhausting. That's coding with AI today.
I'm addicted to vibe-coding. Can't stop, won't stop. I'm vibe-coding up a project I will release soon. Coding first thing in the morning, in between coaching sessions, and right up till bedtime. Best time to be a somewhat technical solo founder or an AI SMB.
@a16z We are at the beginning of the wave of what I'm calling SMAIBs: Small and Medium AI Businesses.
People will use AI as solopreneurs to build entire businesses, and they will generally aim to stay small in headcount throughout their lifecycle.
You know you're having fun coding when a Railway outage turns you into a zombie with nothing better to do.
Guess I'll go play the piano instead of refreshing https://t.co/3wpHWgNjHr every few minutes...
I think itโs bullish for Tesla that Ethan is joining as an intern. Polishing these operational hiccups aggressively and assiduously, line item by line item, is key to scaling.
Getting picked up or dropped off haphazardly is super annoying and is common in self-driving scenarios. A generalized AI reasoning solution would be nice, but in the meantime, solving each instance with brute force will do fine.
As a YC founder, one of the more remarkable things I witnessed was another YC founder working on his laptop during a VC CEO summit. I sat right behind him, and I could see him coding as speakers were talking on stage about PMF. He was coding up a prototype that would eventually be a radical pivot for his company. And that company became a unicorn.
Now I find myself with my own laptop open, sitting in hotel lobbies and airport lounges, coding with glee, using Codex, Claude Code and all the good stuff, and I am building a product for fun that I will release out into the world.
Don't make me say it, but it's true: you can just do things.
Someone should develop an AI โcommon senseโ bench/eval. Adapting from Voltaire, common sense is not so common, especially in AI.
Case in point: most humans can easily identify the entrance vs exit path to a parking lot, but Tesla FSD can struggle with this. You can RL the crap out of this situation by feeding the AI tons of these scenarios, but I think it would be useful to abstract out to a higher layer of common sense.
This AI common sense would have a general understanding of the world and be able to pattern match logic rather than situations. Maybe there is no difference between the two, but maybe there is.
In my opinion, the difference is the meta layer between data and decision, and this middle layer is where common sense sits. Itโs not pattern recognition, itโs pattern recognition of pattern recognition.