Revisiting Conway's Law

post by annebrandes (annebrandes1@gmail.com) · 2025-02-25T08:33:52.421Z · LW · GW · 0 comments

Contents

  The Feedback Loop is the Primary Product
  Dry powder ready to explode
  Credit attribution systems
  Command and control
  Revisiting Conway’s Law
None
No comments

This is a post about how running companies will change. It seems safe to say that markets are becoming more competitive, since AI tools are raising the floor for incumbents and new market entrants alike. But does this shift raise the floor symmetrically? 

And what happens if this shift benefits incumbents? It's easy to imagine a world in which tech giants subsume previously difficult to aggregate assets and edge out new competition due to their ability to resource share under their umbrella. This is a scary world. Social mobility in an ultra-consolidated marketplace is not protected. 

It's my belief that we can create tools to empower even individuals to become competitive with incumbents. I'll make that argument here, and I'd be interested to hear what you think. 

The Feedback Loop is the Primary Product

The purpose of a company is to learn about its environment.^[1] [2] Companies are learning machines, and the best ones focus on extracting more information with less noise as fast as possible. This dynamic has always been true, and promises to become even more true because smaller teams increasingly generate outsized returns. This rising competitive pressure means organizations have an ever-shrinking margin for error in their information uptake rates. If you’re planning on participating in the new world, the feedback loop is your primary product.^[3]

Dry powder ready to explode

Companies waste their data, and it’s not for lack of trying. Companies spend billions of dollars extracting and storing information about user interactions. But at the end of this effort they are left with a few colorful graphs designed to persuade, not to inform. And it’s not their fault. Companies are bottlenecked by intelligence. Learning from data has been a hugely expensive effort until recently because of labor costs. “Until recently” because frontier models solve the cost of labor problem. 

In other words, companies are learning machines, and now it’s possible for the first time to learn at scale. 

Credit attribution systems

Imagine you’re playing chess, you don’t know the rules, and you only get one bit of feedback: whether you won or lost at the end of the game. Over 1,000 iterations of the game, you may improve by classifying every sequence of actions that led to a winning strategy as “good” and every sequence of actions that led to a losing strategy as “bad.” This is a black box RL approach, and the simplest version of credit assignment. However, if you have the ability to figure out why you won or lost, you will become a better chess player in a smaller number of iterations.^[4]

Now imagine you’re running a company. You’re operating in an exponentially more complex environment. It’s not a viable strategy to call the sequence of actions that led to the company going bankrupt “bad”. Instead, we can model the environment itself as a white box that can be cracked open and analyzed. In other words, we’re trading off computational requirements for state of the art sample efficiency. But that’s fine, because the cost of intelligence is dropping by an order of magnitude every year. If indeed a company’s purpose is to learn about its environment, then a company’s purpose is to do credit attribution.^[5] 

Companies are learning machines, it's possible for the first time to learn at scale, and the way to learn at scale is to build credit attribution systems. 

Command and control

This new world will move at a fast clip. Every learning will be extracted and operationalized in real time. Company operations will start to resemble something more akin to high frequency trading or military operations rather than the paper-pushing politics of companies today. 

Data-rich regimes with fast and complex decision-making requirements have a common interface pattern: a centralized operating picture that combines observability, reasoning, and action in a single place. Traders have Bloomberg terminals, and military operations have command and control systems (see Anduril's Lattice OS). Conway (my company) is building a command and control system for companies, powered by a credit attribution engine.

Revisiting Conway’s Law

Conway’s law suggests that the outputs of a company inevitably mirror how the company is internally structured. Companies who refuse to learn have always suffered, and now they’ll be saved from their suffering by the hands of sweet death. The converse is also true. Companies that choose to embrace their roles as learning machines may join the first rarified cohort of companies that last. 

If you want to talk more about this, I'd love to chat. You can learn more about me here, email me at anne@conway.ai, or message me directly. 

 

1. ^{^}

   See Patrick McKenzie's discussion here. 

2. ^{^}

   Isn't a company's purpose to make money? Of course. But if you needed to pinpoint the guiding principle of a company, it's to build a machine that makes money, not to opportunistically make money. This is the difference between winning the battle of a favorable quarter and winning the war of conquering a market. How do you build a machine that makes money? You learn about your environment, and you learn about your environment better than your competitors. Companies are learning machines. 

3. ^{^}

   Inspiration for this language comes from Raemon's post a while back Feedback-Loop First Rationality [LW · GW].

4. ^{^}

   Great grounding from Andrei Karpathy's work Deep Reinforcement Learning: Pong from Pixels.

5. ^{^}

   If you want to read more about credit attribution, especially in a company context, read Gwern's Evolution as a Backstop for Reinforcement Learning. 

0 comments

Comments sorted by top scores.