What's Left to Do When AI Does It All?

I see people struggling with identity when it comes to AI coding. What's left for them if AI does all the work. I'd love to say, and I do say, AI doesn't do all the work, the human adds a lot of value. I sure do. But the reality is - AI can do almost all of SOME kinds of work. There's a word for just telling the AI what you want and pretty much getting it. If not the first request, then after a few more rounds. It's called Vibe Coding. It can be amazing. It can be tragic. But it's NOT all there is to developing software assisted by AI. Here's a video where someone rewriting Postgres in Rust using AI does a WONDERFUL job of displaying just how much human effort goes into developing real world, sophisticated software even when all the coding is done by AI - video linked in comments I know - it's Rust and that gives some of you a rash. I don't care about the Rust part, just listen to this person's perspective. They aren't paying for Claude to convert C to Rust - they have built a software factory tuned to creating database software. The first four attempts at solving this problem with AI didn't work. They saw success the fifth attempt. That's not vibe coding. And listen to this person talk about how they've improved Postgres along the way. This is no mere bug for bug port. They are improving Postgres one susbsystem at a time. At the current state they are 300x more performant than regular Postgres. That's not coming from Rust, it's coming from their design changes and that's coming from human minds who are very experienced in creating database software and who are reading the papers written by other human minds on the topic. They aren't "doing the coding" themselves, but neither is Claude "doing the conversion" itself. I just threw out my work of almost a year because I got frustrated with Astra trying to polish it to perfection. My orchestrator and autonomous AI employees have always been brittle. Astra worked for near a week solid. Several times 24 hours at a stretch. And STILL brittleness remains. I called a halt. Clearly I was insane - doing the same thing and expecting different results. It was me that called for a do over based on lessons learned and the fact that I was solving for problems that today's llm's and harnesses don't have. My orchestrator was all about taking control away from the llm, making them prove every step along the way. The contracts and proofs between each stage became so cumbersome after a year of handling ever-occuring edge cases. So I worked with Astra to strip down the essence of what an AI Employee is in it's simplest form. What an Employee needs that you don't get just by prompting Claude Code or Codex. And built up from there. Astra on xHigh took my input and came up with a plan - spent maybe an hour to do so. And STILL the plan was deficient. It didn't understand autonomy as I had directed it and more. I'm aiming for "more than GrokBot but less than Azure/AWS". I'm the taste maker, the decider, the person who sets what is to be built and oversees the quality. I define WHAT we are aiming at that goes far beyond "hey Claude, build me an AI Employee factory". I'm on my own 4th or 5th gen from when I started. I take lessons between each generation. The ability of agentic AI has been changing and improving as well. Both of these are also examples of AI growing the pie instead of replacing humans. There was no $100M budget with a couple hundred devs and 3 years to rewrite Postgres to improve performance 300%. It's work that wouldn't happen, not work taken away from humans. Same for the AI Employees I have and the platform I'm building. I don't have staff to replace. But with AI, I have staff. Everyone doesn't need to play at this level. I just hope to encourage those who are discouraged by vibe coding, that there's a LOT more role for your amazing human mind than putting money into a vending machine or ordering a burrito. And it's fun and challenging.