The Big-O equivalent of AI productivity complexity.
H = Human interventions to your AI interface (prompts, permission approvals, starting a new session etc.) P = Projects you’re working on T = Tasks within a project, actual atomic work AI + you ship out (code commits, published PPTs etc.)
From H, P, T, the levels of AI productivity then become -
L0 : O(H x P x T) = Manual with AI assistance
You open a new cold Claude chat session for every subtask T, explain context, approve permissions one at a time.
L1 : O(log(H) x P x T) = Context automated.
You’ve learnt to use CLAUDE.md, settings.local.json, Skills effectively. You still open a new session for each T but context is pre-loaded, your workflows (working in new git branch each time, PPT themes etc.) are baked in and only unforeseen permissions require manual intervention H driving it exponentially down but non-zero.
L2 : O(log(H) x P x log(T)) = Systematic workflows.
You have effectively implemented subagents based on your dev patterns and you use them in T and you ship exponentially more output. Project P’s state management is still in your wet human brain.
L3 : O(P) = Autonomous loops.
You have setup routines on claude and you use claude dispatch to trigger subagent teams producing autonomous outputs with your context on P, barely needing H (constant). Either because you’re running this in sandboxes or you’ve gone yolo on permissions. Congrats, you’ve built an AI colleague. Or installed OpenClaw.
L4 : O(log(P)) = Portfolio management.
You have offloaded context management - Andrej Karpathy’s LLM Wiki style. You view your autonomous agents on a dashboard Paperclip style. You are literally boiling the ocean (for your tokens). You either are one of the most productive people or you have not realized yet that OpenClaw texted your ex.
L5 : O(1)
This is when you stop believing in free will because your AI setup anticipates everything you will want to do and just does it. Ted Chiang’s scifi story - What’s Expected of Us - comes true.
Alright L5 was a joke, but L4 is very much real.
The 2 interesting things about AI productivity seem to be -
- it feels like a maslov’s pyramid - one cannot get L4 right and risk free without having done L3 first, and
- it’s a paradox - you need to slow down and invest mindful time in building your setup first to get the super accelerated autonomy later.
Build the tool that builds the tools.
![[malows-ai-productivity-pyramid.png]]