We're a couple years post ChatGPT, and the market has finally converged: AI will be commercialized in the enterprise. Superintelligence isn't required to toast a piece of bread, but may very well be required for running a 300,000 person company, protection against foreign adversaries, or conducting novel research.
Coding agents established the product market fit of AI in the enterprise by showing unprecedented revenue ramps quarter after quarter. We're now entering a lumpy middle. Coding agents are magical, but are clearly not the only (or even the primary) product for enterprises to spend tokens to maximize return on investment. Router variants may help keep down costs but they don't answer the fundamental question: Where will the biggest companies in the world spend most of their tokens?
Company building and model building have a lot in common with one another. Abstractly, they are both functions that convert data and capital (compute) into objectives. Success usually involves optimizing the conversion of data and capital into objectives as quickly as possible, over and over again.
Two conclusions fall out of this insight:
- Company building is model building
- There are two fundamental model building tasks: identification of worthwhile objectives and objective optimization
It seems to me that identifying worthwhile objectives is a continuously changing, at times fundamentally human task. Objective optimization, on the other hand, strikes me as repeatable and autonomous. Productizing objective optimization, then, should be both possible and unassailably important.
Frontier model building today is most commonly associated with large language models and their generative outputs. Imagining a company as a language model falls apart upon examination primarily due to their cost, unless you imagine a superintelligent AI with infinite compute building novel internal tools (classic mainstream lab cop out).
The most meaningful difference between the model that will optimize everything and a language model is learning efficiency. In other words, this new model will be able to efficiently elicit feedback from its environment to incrementally improve on defined objectives. The solution may sound algorithmic, but algorithms are most likely only a small part of the solution.
Model training techniques today pose the wrong question to solve credit attribution. Optimization is not best thought of as a skill to post-train a language model to improve on, it's not really a task at all. Instead of training models to improve at the task of optimization, models should be trained to directly achieve the outcomes themselves, starting with model evaluation co-design. Email team [at] conway [dot] ai to learn more.
So where will enterprises spend tokens? There will be token expenditure on software agents. Relative to today, you can expect this spend to grow as company builders find more and more ways to use agents to build products and do research. That said, this spend will be dwarfed by autonomous systems optimizing every facet of a company everywhere and always.
Conway's Law is an old computer science adage pointing out the symmetry between team structure and their products. Today Conway's Law reads less as a saying and more like an axiom: Company building is model building. Optimize everything.