Why did software change the world?

In the 1900s, much of the work being done by knowledge workers was computation: searching, sorting, calculating, tracking. Software made this work orders of magnitude cheaper and faster.

Naively, one might expect businesses and institutions to carry out largely the same processes, just more efficiently¹. But rather, the proliferation of software has also allowed for new kinds of processes. Instead of reordering inventory when shelves looked empty, supply chains began replenishing automatically based on real-time demand. Instead of valuing stocks quarterly, financial markets started pricing and trading continuously in milliseconds. Instead of designing products on intuition, teams began running thousands of simultaneous experiments on live users. Why did a quantitative shift in the cost of computation result in a qualitative shift in the nature of organizations?

It turns out that a huge number of useful computations were never being done at all. Some were prohibitively expensive to perform at any useful scale by humans alone. But most simply went unimagined, because people don’t design processes around resources they don’t have. Real-time pricing, personalized recommendation systems, and algorithmic trading were all inconceivable prior to early 2000s.

These latent processes could only emerge after the cost of computation dropped.

Organizing Agent Societies

The forms of knowledge work that persisted to today are now being made more efficient by LLMs. But while they’ve enhanced the efficiency of human nodes on the graph of production, the structure of the graph has still been left intact.

In software development in particular, knowledge work consists of designing systems, implementing code, and coordinating decisions across teams. Humans have developed rich toolkits for distributing such cognitive work – think standups, code review, memos, design docs. At their core, these are protocols for coordinating creation and judgement across many people.

The forms that LLMs take on mediate the ways that they plug into these toolkits. Their primary manifestation – chatbots and coding agents – implies a kind of pair-programming, with one agent and one human working side by side. In this capacity, they’re able to write code and give feedback on code reviews. But not much more.

In this sense, we’re in the “naive” phase of LLM applications. LLMs might make writing code and debugging more efficient², but the nature of work hasn’t changed much.

The first software transition didn’t just make existing computations faster; it allowed for entirely new kinds of computation. LLMs, if given the right affordances, will reveal cognitive flows we haven’t imagined yet.

A new metascience

How should cognitive work be organized? For most of history, this has been a question for metascience, management theory and post-hoc rationalization. Soon, this question will be able to be answered by experiment. We’re curious to see what the answers look like.

1. This resembles the lump of labor fallacy, which is misconception that there is a fixed, limited amount of work to be done in an economy that can be distributed to create more or fewer jobs. Under the fallacy, if an individual does work more efficiently, someone else must lose a job, since work is zero-sum. In fact, economists have found that as more work is being done, the economy grow. This situation isn’t exactly analogous, but has some similarities. Here, the naive expectation about software’s effects implicitly holds the assumption that there’s a fixed amount of computational work to be done in the world. In fact, as this essay argues, there were many forms of desired computational work that weren’t being done prior to the introduction of software. ↩

2. This claim is highly contentious – see METR’s uplift study. But whether AIs actually make developers more or less effective is not central to this argument: the main point is that LLMs are working with our systems at a human-node level, which limits the ways they can be of use. ↩