The OODA Hourglass
Squishy Computer OODA-pilled me recently. What a sentence!
I have been liking OODA as a framework for thinking about both my own work and how LLMs might be working through tasks.
This is where my thinking has currently landed:
Observe: LLMs very good at this for text. They can process heaps more text than I can, so I want LLMs to augment my observing capabilities (Deep Research, etc).
Orient: Now things get tricksy. LLMs can absolutely get disoriented in a codebase. And I can get disoriented when the output of an LLM overwhelms my capacity to read it all. I want my LLM to assist me in maintaining orientation.
Decide: I don't want my LLM making decisions.
Act:LLMs are very good at taking digitally-native actions. My computer is better at using a computer than I am. I want to be sandboxed, and I have git to version control everything so that the LLM can take actions
OODA + AA_A!
I want LLMs to:
Augment my Observing capacity.
Assist in maintaining Orientation / staying on the same page.
_ Don't make Decisions
Act
Orientation ends up being where I spend a lot of time. I blab to the LLM and then request a quick ASCII diagram. I give very minimal instructions like "plan a fix for Foo" and then see what happens.
html
html takes longer for the LLM to generate than raw text, so it only makes sense when you're working async, or to create "bookmarks" of persistent work
you're spending more LLM tokens to create an artifact that will be easier on you to take in. You're owning your own limitations in information processing. Using your visual cortex
templates are nice! I am still experimenting, but I am liking how they constrain the LLM-generated html into known forms. A template can "say" a lot in a very strict way. Instead of "EVERY NODE MUST HAVE A DIAGRAM, summaries NEVER longer than a tweet!!!!!" you just say, "copy the template, then add components from the library:":