Thinking in Systems — notes
last updated 2026-05-15
A reading log on Donella Meadows’ Thinking in Systems. Filed here because most of the entries are half-thoughts that I’ll come back to, not finished arguments.
Stocks dominate flows
The earliest mental shift the book asks for is to stop treating change as the unit of analysis. Stocks change slowly; flows change quickly, and most failure modes come from optimising the visible flow while ignoring what’s accumulating in the background.
This shows up in ML evaluation more often than you’d expect. A model’s current accuracy on a benchmark is the flow; the underlying drift in the test distribution is the stock. Watch the stock.
Leverage points
Meadows’ twelve leverage points list (in increasing order of leverage):
- constants, parameters, numbers
- the size of buffers
- structure of stocks and flows
- delays
- balancing feedback loops
- reinforcing feedback loops
- information flows
- rules of the system
- self-organisation
- goals
- paradigms
- transcending paradigms
I keep returning to (7). Most engineering organisations have lots of information; very few have it routed to the people who can act on it. The cheapest interventions in real systems are usually the boring ones: moving a dashboard, changing who’s on a thread, making a metric visible.
Open questions I’m chasing
- Is there a useful analogue of “leverage points” for ML systems specifically?
- The book’s case studies all assume slow-moving systems. What changes when feedback is sub-second?
For a parallel reading on second-order effects, see Why VLMs Fail at Tables.