---
title: "The Autonomous Model"
date: "2026-10-10"
description: "A rebuttal to Alex Wang's article 'The model is the company'"
draft: false
---

[Alex Wang's article "The model is the company"][wang-article] argues that the next
generation of startup companies will have vastly reduced headcounts with the
knowledge located within a 'Model'. The specifics of the Model are left a little vague,
so I will make my best attempt at describing what the Model is. The Model is a machine
that takes in information and outputs either actions or directives.
Wang describes the Model as superintelligent, implying the extraordinary claim that
the Model would be capable of outcompeting a similar business[^scots].
He argues the Model would be capable of using the entire history of
inputs, actions, and results to make informed decisions about future actions. Another
feature of the Model is that it can work, learn and execute at any time.

[^scots]: I realise that this could devolve into a No True Scotsman fallacy. I will
define the comparative company as one with similar financial capital to
the company running the Model.

The article is exuberantly optimistic. Wang is arguing that the next generation of startup companies will each be a demigod in a bottle[^demi].   

[^demi]: Wang uses phrases like "superintelligent" and "complete context of the
business" in his article, so forgive my literary flair.

I consider myself a technology advocate however Wang's illustration fails to take into
the physical constraints of limited focus a data resolution. As any undergraduate
science or literature student can tell you, authoring a piece of work in not a unbiased
affair, nor is collecting data. There is inherent choices being made on what to measure
and what to keep. This is important when the customer is reliniqusing control to the
Model as the biases of the customer and the Model must align for a fruitful
relationship.

To simplify the picture, let us assume that the Model showcases complete
accuracy and capability to perform its function. In principle the Model has solved
its chosen domain. Let us call the owner of the Model The Company. The Company is now
liable for the Model and its operation. It requires sales, insurance and customer
relations teams: someone to sell and promote the Model, someone to negotiate liability
when the Model is misaligned with reality, and someone to maintain the relationship
through hard times. If asked, a techno optimist may argue that these roles ought to be
delegated to a large language model or incorporated into the Model. The key word,
though, is that it is a *delegation* of responsibility from *The Company*. Don't
believe for a second that the Model itself can hold responsibility, à la
[a computer can never be held accountable][computer-accountable][^slide].

[^slide]: Attributed to a 1979 IBM training slide, a photo of which was shared online
in 2017.

A techno optimist may concede that The Company will need the founder,
a lawyer, and an insurer. However, the argument becomes a little curly when you start
thinking about what the insurer needs. Will the insurer trust the Model to certify
itself? That would be a very poor financial move on the insurer's side. The task 
for The Company, then is to satisfy the insurer. How will the insurer be satisfied? I
believe that will only occur once the Model’s processes can be validated in a way that
is comprehensible and standardised. But who determines that the embedded logic is
standard? Who can comprehend the Model? This all starts to sound like it requires
expertise and knowledge from outside observers.

The techno optimist may argue that these validation services could be contracted out,
perhaps even to other models. That misses the forest for the trees. The insurer cares
that someone has made *representations* about how the Model behaves and is willing to
stand behind them. Most risks are priced from loss history, the record of past claims,
which is why freight and motor vehicle insurance are routine. A new Model has no loss
history, so it falls into a different camp, alongside autopilot in aviation: insurable
only once a certification regime, run by experts, vouches for it.


[wang-article]: https://alexwang.ai/posts/the-model-is-the-company/
[computer-accountable]: https://simonwillison.net/b/8470