From useful work to useful intelligence

That sentence is why useful proof of work got loud this year, and why a lot of the loud version is still incomplete. Matrix multiplication sits under most inference. So it looks like the honest swap for a hash. Pay machines to multiply. Secure a chain. Tell yourself the energy went into AI. Then watch a miner multiply random grids, pass the check, and collect the reward. The arithmetic looks like AI. The job does not exist.

The test is not "was a matrix multiplied." The test is "did anyone need the answer."

Amadeus has been live since March 2025. Solvers do seeded matrix multiplies, the same family of operation a model uses when it infers. Other nodes check the solution. The solvers who score, plus nodes backed by vaults, vote. Signatures bundle. An entry is final in about half a second.

On the volume we will actually stand behind: Amadeus executes over 1.5 trillion matrix multiplications every day. That is a safe floor, true for every one of the last 10 epochs. Roughly 20 million matrix multiplications per second, around the clock. About 20 quintillion multiply-accumulate operations per day of verified, useful compute.

You can bring hardware and earn without parking a token first. That is not a promise. That is how the set is built.

Plain version: the network does not pay you to guess a hash that dies the moment it is found. It pays you to do a calculation of the kind intelligence actually uses. Then it checks you. Then it moves on.

It is not, today, a Hub agent typing a request and a miner answering that exact request on the next block. We have said this before and we should keep saying it. The primitive is live. Routing those tiles into confidential inference a live agent asked for is the work in front of us.

If we skip that distinction we are doing the thing the category is being called on. Printing "useful" on verified maths that had no customer.

A mining network with no door will always fail the test above. The multiply has nowhere to go except the next block reward.

Hub is the door for someone who shows up and composes an agent in plain language. Embed is the door for a wallet, an exchange or a terminal that already has a screen. The agent sits inside that screen. The partner does not rebuild the sealed runtime.

Those agents also have to survive real capital. Same inputs have to produce the same outputs, so you can replay a mandate instead of arguing with a log. The strategy has to sit inside a TEE, so the box it ran on does not get to read it. And the agent has to be able to pay. The venue it trades on can still be Base or Solana. The bill should not live in a spreadsheet.

That stack is why we think the multiply can become a job. Not because MatMul is fashionable. Because there is someone to ask for the result, a place to run it in private, and a way to settle.

Other networks made GPU mining feel current again. They should be in the conversation. Some connect agents to chains. Some sell compute as a market. Some treat matrix multiplication as the new hash.

Different jobs. Ours is narrower. Own the work. Own the replay. Own the seal. Own the bill. Own the door. Then stop claiming the loop is closed until the serving path is something Ivan will put his name on.

Useful work is the calculation. Useful intelligence is the agent that needed it. We have the first. We are building the second. That is the whole piece.