Five New Thinking Styles for Working With Thinking Machines

Was this newsletter forwarded to you? Sign up to get it in your inbox.

Our default thinking style in the West is scientific and rationalist. When was the last time you heard someone talking about a hypothesis or theory in a meeting? When was the last time, when sitting down to solve a problem, you reminded yourself to think from first principles? When was the last time you tried an experiment in your work or personal life? 

Even the frameworks we use to understand business are scientific: It’s unlikely that Harvard Business School professor Michael Porter would have looked for or found five “forces” governing business without physics as inspiration; Clay Christensen’s jobs-to-be-done framework is close to an atomic theory of startup ideas. 

We romanticize science and rationalism because it's been so successful. Since the Enlightenment, when Galileo, Newton, Descartes, and Copernicus began to think in this way, we have used rationalism to generate modernity. It's where we get rockets and vaccines from, and how we get computers and smartphones.

But new technologies demand new thinking styles. As the AI age unfolds, we are shifting away from what former Tesla and OpenAI engineer Andrej Karpathy calls Software 1.0—software that consists of instructions written by humans, and which benefits from a scientific, rationalist thinking style. 

Instead, we're moving into Software 2.0 (a shift that Michael Taylor recently wrote about), where we describe a goal that we want to achieve and train a model to accomplish it. Rather than having a human write instructions for the computer to follow, training works by searching through a space of possible programs until we find one that works. In Software 2.0, problems of science—which is about formal theories and rules—become problems of engineering, which is about accomplishing an outcome.

This shift—from science to engineering—will have a massive impact on how we think about solving problems, and how we understand the world. Here are some of my preliminary notes on how I think this shift will play out.

1. Essences vs. sequences

In a pre-AI world, whether you were building software or teams, or writing books or marketing plans, you needed to strip the problems you were facing down to their bare elements—their essence—and work your way forward from there. In building software, you need to define your core user and the problem you want to solve; in writing books, you need a thesis and an outline.

In a post-AI world, we are less concerned with essence and more concerned with sequence: the underlying chain of events that leads to a certain thing to happen. Language models do this when they predict what word comes next in a string of characters; self-driving cars also do this when they predict where to drive next from a sequence of video, depth, and GPS data. 

To understand this better, consider the case of a churn prevention feature for a SaaS business in a pre-AI world. In order to automatically prevent a customer from churning, you needed to define what a customer who might churn looked like with explicit rules—for example, if they hadn’t logged into your app in a certain number of months, or if their credit card was expiring soon. This is a search for essences.

In a post-AI world, by contrast, you don’t need to explicitly define what a customer who is about to churn looks like, or which interventions you might use in which circumstances. 

All you have to do is identify sequences that lead to churn. For every customer who churns, you can feed their last 100 days of user data into a classifier model that categorizes inputs. Then you can do the same for customers who haven't churned. You'll create a model that can identify who is likely to churn, in all of their many thousands of permutations, without any rules. This is what it means to search for sequences. 

2. Rules vs. patterns

Another way to think of essences versus sequences, especially when it comes to intellectual and creative tasks, is as a shift from the search for rules to a search for patterns.

In a pre-AI world, you needed to define the rules of the game you were playing—to think from first principles and apply them to your circumstances. In a post-AI world, you need to build and use models that recognize underlying patterns—patterns that can’t be reduced to simple rules.

Consider building software. Pre-AI, you needed to define exactly which tasks you wanted a user to be able to accomplish and exactly how you wanted the system to behave, so you could encode those definitions into rules written by programmers that were readable by a computer. 

In a post-AI world, however, instead of defining rules, you need to look for examples. You don't need to explicitly lay out what users can and can't do with your app. You can create a mood board of your favorite UI elements, or a high-level list of how you want your application to behave—and AI will identify the patterns in your input and translate them into a ruleset.

Or maybe you’re running a creative team. In a pre-AI world, in order to create consistency, you needed to reduce your work down to principles and systems so that writers and designers could consistently capture your brand's voice and style.