Apes on fire

After Software

Jo Wedenigg

← Back
after-software

I remember the oak tree next to our house. As kids, we would sit underneath it in the tall grass, surrounded by flowers, looking at the clouds and deciding which of the imaginary things we were dreaming up at that time we were going to do next.

Technology existed. It was part of life, yet, it did not contain life. We had television. Movies were important.  You’d spent an afternoon outside and come home to intentionally watch a film together. Our houses had land line telephones. And radios, and calculators, vacuums and tools. Technology was utilitarian and neatly contained.

I belong to the slightly unfortunate generation that then spent adulthood adapting to technology over and over again.

First: the Internet.

Then: Mobile.

Social.

Blockchain.

SaaS.

Now AI.

Every wave came with another set of interfaces, skills, experts, workflows and businesses required to navigate the previous wave.

Which makes me wonder whether AI might be just another wave in this endless rollercoaster of change, or could it eventually produce a strange inversion:

What if this is the last adaptation?

Not because technology disappears.

Because it eventually, finally, becomes capable enough to submerge into what we actually wanted it to do.

An economy of cognitive friction

For most of human history, useful cognition has been scarce.

Knowing something was expensive. Finding the person who knew it could cost you as much. Transferring their knowledge was expensive. Applying it – expensive. Coordinating multiple knowledgeable people – really expensive.

Humanity has therefore spent centuries, millennia, constructing infrastructure around the limitations of human cognition.

Writing externalized memory. Professions concentrated expertise. Markets helped locate capability, while companies internalized it. Management then got to coordinate it. Software encoded workflows and processes, made knowledge execute. The internet distributed it and search helped us retrieve it.

None of this created or required particularly new economics.

Ronald Coase argued in 1937 that firms exist partly because using markets carries transaction costs, and that firm boundaries move according to the relative cost of coordinating transactions internally versus externally. Herbert Simon later observed that information abundance does not eliminate scarcity: it moves scarcity toward the attention required to process information. Robert Grant’s knowledge-based theory of the firm describes organizations as institutions for integrating specialized knowledge distributed across people.

Remarkably, newer economic work is returning directly to Hayek’s early insight of economically useful knowledge being dispersed among individuals. Erik Brynjolfsson and Zoë Hitzig ask what happens when AI can codify local knowledge that was previously tacit and massively expand the ability to aggregate, process, interpret and act on information.

The old problem doesn’t disappear, but the cost of solving it changes.

The interesting thing is how consistently technology has moved in one direction as a result:

towards lowering the opportunity and transaction costs of cognition.

Even the now-fashionable idea that judgment becomes more valuable as machine capability rises predates ChatGPT. In 2018, Ajay Agrawal, Joshua Gans and Avi Goldfarb explicitly separated prediction from judgment and explored what happens as the cost of machine prediction falls.

AI might simply be the asymptote of that history.

The internet made knowledge cheap to distribute, and AI makes it increasingly cheap to apply.

Research, synthesis, translation, analysis, prediction, coding, design, planning, and increasingly: action itself – all subject to prompting and harnessing the right model.

When the price of an input changes radically, the structures built to economize that input get repriced too.

Software is frozen cognition

For fifty years, computers could execute instructions beautifully but could not understand what we wanted. This entire conundrum led humans to translate intention into workflows, workflows into requirements, requirements into specifications, specifications into code, code into applications, applications into interfaces – and then onwards to train other humans to operate the interfaces.

That entire chain exists because cognition could not travel directly from intention to execution. Software is knowledge and cognitive process frozen in code. We got so good at building that translation layer that we started treating it as the product.

AI attacks the assumption underneath it.

SAP is now explicitly describing “batch size 1” applications: temporary interfaces generated around one user, one context and one problem. Palantir can generate ontology entities, design specifications and production frontend code from a natural-language description. Gartner estimates that agents bypassing UX-heavy applications could expose roughly $234 billion of enterprise SaaS spending by 2030.

Software doesn’t disappear – there could be even orders of magnitude more of it. Something more consequential is happening: software loses its conceptual primacy.

Andreessen may be completely right about software eating the world at compute speed. But ubiquity is almost the direct opposite of scarcity. Once software can be created, modified, operated and discarded at the whim of a moment, it becomes technically universal while at the same time becoming economically less interesting as the thing humans buy and navigate. Software does eat everything. Eventually, perhaps, including itself.

The customer persists. The account persists. Inventory persists. Identity, history, permissions and commitments – all of those elements of the transaction persist. The application through which we interface with the transaction, however, may not. Software becomes the generated and generative infrastructure beneath intention.

And then the shock moves upward. Because applications are not the only structures we built around cognitive friction.

The same is true for workflows: a workflow exists partly because no single actor can know, decide or execute everything required at once.

  • Export this.
  • Brief them.
  • Ask Finance.
  • Send Legal.
  • Update the system.
  • Schedule the meeting.
  • Wait.

Much of knowledge work is the machinery required to move limited cognition through an organization.

AI can automate those steps. More importantly, it can sometimes remove the reason for those steps to exist in the first place.

That is the beginning of the rupture. The technological singularity is usually imagined as the moment machine intelligence exceeds ours. But the institutional rupture need not wait for AGI. Cognition only needs to become cheap enough at the margin that the old build-versus-buy, human-versus-machine and coordinate-versus-execute equations begin flipping.

Perhaps the singularity arrives economically before it arrives intellectually.

Not as one machine suddenly becoming godlike. But as the infrastructure we built around human cognitive scarcity becomes progressively uneconomic. And then the question gets much bigger than software:

How much of modern civilization is infrastructure for managing the scarcity of human cognition – and what survives when that scarcity disappears?

 

Sources: