Read on Substack: https://bitkarrot.substack.com/p/building-in-the-agentic-era
Two podcast appearances this month — on Open Markets and Fundamentals of Fundamentals — capture what it actually feels like to build this way, with unusual honesty.
For most of software history, the bottleneck was fingers on a keyboard. In 2026, the bottleneck is knowing what to ask for.
Two thousand hours to unlearn old habits#
The number everyone asks about: for an experienced engineer, getting good at working with AI takes about 2,000 hours — “a full year of just trying all these different tools.” Not because the tools are hard to click, but because the whole way you think about building changes. The machine does the typing now; your job is the thinking.
The tell that someone has crossed over, from my Prague talk: someone from a major grant org asked how HiveTalk got non-technical people running their own servers.
My answer: “You just give the prompt to the AI.”
The asker expected a script.
In the Age of AI, There is no code based installation script.
There is just a prompt, that’s it.
And the humbling part:
“You realize just how bad your prompts are when the AI builds something that actually touches other humans, and the problem was not what you said — it was how you defined the problem.”
The moment things tipped#
Two numbers frame the new era:
- Half the internet’s traffic is now bots. As of early 2026, Cloudflare reports more robot traffic than human traffic — the same kind of step-change as March 2020, when the whole world suddenly went online.
- AI usage is doubling every 11 weeks. Computing power used to double every 18 months; AI usage now does it in under three months. In that kind of compounding environment, waiting for things to “stabilize” before jumping in isn’t a strategy — the gap between early adopters and late adopters is widening every week.
And a demo that made the point better than words: I asked for a full Bitcoin-and-Lightning setup on a new hosting service — the kind of thing that used to take a weekend of cursing at config files — and the whole thing was ready in 60 seconds. Non-technical team members now run their own servers this way; one HiveTalk volunteer went from zero to contributing code in months. “If Buttercup can do it, any reasonably intelligent person can.”
The speed story that broke a mathematician’s brain#
The funniest moment of the Fundamentals episode: I pulled up a chart of how fast different AI systems think. The familiar names — GPT, Gemini, Claude — all run at roughly 100 tokens a second. Then I showed a new kind of chip, from a company called Cerebras (the biggest hardware company IPO in history): ten times faster from 1000 tokens a second. And then Taalas: a hundred times faster, running at 14,000 -17,000 tokens per second.
I invited Brian, the host, to try it live. Brian asked it a hard math question — the answer appeared in 0.004 seconds. (It dodged the actual question, but at that speed, who’s counting.) “That broke my brain,” Brian said. When answers arrive instantly, the whole pace of building changes again.
Read the full post on Substack: https://bitkarrot.substack.com/p/building-in-the-agentic-era
Comments
No comments yet. Start the conversation.