The ARPA Model is an unbuffered system

An unbuffered system is extremely responsive at the cost of margin for error.

The dependence of DARPA on high quality program managers mirrors the obsession with “talent” in other disciplines means that a bad apple can have a rather large effect.

In general, you can see the tradeoff between responsiveness and error in many domains. In chemistry, a buffered system resists change in pH despite the addition of acids or bases that would otherwise change pH. Artificially stable systems in control theory use active control to turn a naturally unstable system into an incredibly responsive system that needs constant control input. Control Theory.

These analogies might inform strategies for mitigating downsides. 🤔

Ideally DARPA can move quickly because the Director and Program managers are really the only two checkpoints. That means that only two failures can cause a cascade failure. A cascade failure looks like the director hiring a low-quality or megalomaniacal program manager who he trusts for reasons besides competence and provides no checks on. The Total Information Awareness Program in the early 2000s was an example of this. The Director pulled in a buddy as a PM, that PM went meglomaniac and started hiring his own firm to do things, which led to congressional review-worthy misspending.

Over time DARPA has buffered the system more, but that has also slowed it down: using competitive grants, the tech council, and congressional oversight.


I’ve been treating the potential for congressional review as a bad thing but it fits with an ask forgiveness not permission model. You don’t need a lot of checks and balances if everybody knows that they will be judged in the end.

There are a couple of saving graces - programs are relatively uncorrelated once they are approved. The need for uncorrelated programs would suggest that it is important that the director’s role be purely approval-based, injecting as little opinion as possible. DARPAs director is important.

It suggest that we might step back and look at The ARPA model through the lens of checks and balances in a government. Maybe creating a constitution is actually a good idea? PARPA needs a constitution with checks and balances


Web URL for this note

Comment on this note