# Intelligence Refining Companies

*This article was adapted from notes in my journal.*

---

Tuesday, October 20, 2026.

I didn't sleep very well last night. I woke up at 5 AM with a strong feeling of excitement, a bit like in December 2025 when I couldn't stop thinking of new things to make Opus 4.5 do. I think I finally understood something about the future, and the direction my work is going to take.

Very soon (my bet is early 2027), artificial intelligence will surpass us not only at solving problems, but at understanding them. In other words, when I ask AI for something, it will know better than I do what I want. Or rather, what I actually need.

Right now, my business ([Verity Labs](https://veritylabs.dev/)) makes money by auditing critical software, like Lido's smart contracts or precompiles at the heart of Ethereum, using formal verification.

This will probably survive, but applied to much larger problems that can't be solved without dangerous amounts of intelligence.

## Uranium, not oil

Frontier labs have such a profitable business that they have no interest in making enemies of the entire world and adding a million operational problems to their plate. To me, inference providers will look a lot like Aramco, Saudi Arabia's national oil company.

The difference is that intelligence is closer to uranium than oil: you can do great things with it, but it's not something you can put in everyone's hands.

Which brings me to the concept of an **Intelligence Refining Company**.

These will be more or less regulated companies (depending on the quality and quantity of raw material passing through their hands) whose competitive advantage will lie in two things: their mastery of the refining process, and their distribution capacity.

The least ambitious version of this vision for us would be to become an accredited research firm whose value comes from finding clients and convincing authorities (labs or governments) that our use of AI is legitimate.

Something much stronger is to become an expert in the transformation process itself: turning a stream of raw tokens into valuable, certified knowledge.[^1] That creates a virtuous circle: lowering costs enough that clients are better off going through us than trying to solve their problems themselves, capturing data from every run (both what worked and the dead ends we had to prune) to train smaller models and optimize the process, and unlocking economies of scale.

## Two examples

To make this concrete, let's take a short-term example, and then a long-term one.

### 1. Short term: embedded software

Take the firmware of embedded devices, like a Ledger hardware wallet. We can make it both safer and lighter by rewriting the code in Lean 4, proving it correct, and optimizing its compilation down to the target chip.

In the process, we improve our ability to route each sub-problem to the cheapest capable model, generate formal proofs efficiently, and optimize our own compiler. All of that compounds for the next client.

### 2. Long term: the physical world

Imagine designing a more efficient engine, or any problem that normally requires expensive tests in the physical world. Once a technology is mature, every extra percent of efficiency requires exponentially more mind energy and physical prototyping.

After formalizing the rules of that physical world, we would generate a simulation proven to conform to those rules, but computationally optimal: simulating only the exact interactions needed to iterate on the design, and compiled specifically for the hardware at our disposal.

Once again, not only our proving machinery, but a large part of the work itself (what we call certified research) becomes reusable for future problems.

## How to get there

To reach this vision, I've identified three priorities:

1. **Securing access to frontier models**, through certifications and strong agreements with the labs.
2. **Identifying bottlenecks and eliminating them autonomously** (humans like myself in the loop, physical tests, etc.), which implies measuring as much as possible.
3. **For the bottlenecks that cannot be eliminated, securing supply that unlocks economies of scale**, both in hardware and inference.

The best business in the world today is to have a problem AI can't solve and sell it to AI labs so they can train their AI on it.

That's too hard for us, so we settled for the second best business in the world: using AI to solve every other problem.

[^1]: All of this might sound a bit abstract for now, but it will become clearer very soon. A hint: [Maxwell's demon](https://en.wikipedia.org/wiki/Maxwell%27s_demon).