Hi,

Someone spends four days on a ten thousand piece jigsaw. Sorting the edge pieces, matching sky to sky, testing pieces that almost fit and putting them back.

Then a stranger walks in, glances at the table from the doorway, and knows in one second whether it is finished correctly.

That gap is the whole of Bitcoin mining, and for seventeen years almost every complaint about it has been aimed at the wrong half. The electricity is the four days. The one second is the point.

Last week a well known founder came out of retirement to announce a network that swaps the pointless puzzle for real AI work, and promised, in a single clause, that validators will verify the work was completed correctly.

That clause is the hardest open problem in this field. And the token goes out a full quarter before there is a network to check anything on.

Listen:

Spotify: https://open.spotify.com/episode/7uV5YjFMO69CWHS6ZXkLqb

Apple Podcasts: https://podcasts.apple.com/us/podcast/plain-strata/id6783455764?i=1000786613386

YouTube: https://youtu.be/D0jG7jatdMw

---

The full piece, no need to click through:

There is a machine somewhere doing the most famous pointless task in computing. It picks a number, runs it through a fixed grinder, looks at the result, sees it is not small enough, and picks another number. It has done that a few hundred trillion times since you started reading this sentence, and it will keep doing it until it stops or the electricity does. That is Bitcoin mining, and for seventeen years the standard complaint about it has been the waste. All that power, spent on numbers nobody needed.

On the eighteenth of August, that complaint got the answer it always eventually gets. Arthur Hayes, who co-founded the derivatives exchange BitMEX, announced he was coming out of retirement to run a company called Flop Labs, building a network whose token, FLOP, he described as food for AI agents. The pitch is the obvious one. Stop burning electricity on numbers nobody wants. Have the machines run real AI work instead, answering questions for software agents that pay for the service in the network's own token. The announcement calls this proof of useful inference. Miners supply the graphics cards. Validators, in one clause of the announcement, verify that the work was completed correctly. No presale, no venture capital, a fair launch, a large airdrop in the fourth quarter of this year, and the network's first block in the first quarter of next.

Read that last pair again in the order it was given. The token gets handed out a full quarter before the chain it runs on exists.

Here is the thing worth carrying out of this week, and it will outlive whatever happens to this particular network. The wasted electricity in proof of work was never the clever part. The clever part was the checking. Every proposal to swap the waste for useful work has to buy that checking back, and for AI inference the price is the highest in the field.

Start with what is physically happening. A blockchain is a shared record kept by machines that do not know each other and have no manager to appoint. They have to agree anyway. Consensus comes from the Latin consentire, to feel together, which is a generous word for what these machines do: they do not feel anything, they simply arrive at the same list. To decide who gets to add the next page of that list, Bitcoin makes everyone play a lottery you can only enter by computing. A miner takes the page it wants to add, staples a guessed number to it, and runs the whole thing through a fixed mathematical grinder that turns any input into a fixed-length scramble. If the scramble comes out below a target, the miner wins. If not, new guess. Across the network that is quadrillions of attempts a second.

Now the part that matters. To check the winner, you do the grind once. One pass, microseconds, on any laptop. Producing the answer takes the planet's worth of effort; confirming it takes a blink.

That gap is the entire foundation. It is what lets a network of strangers accept a page from someone they have never met, in a country they cannot name, with no license and no application form, and know immediately whether the work behind it was real. Take the gap away and permissionless stops being possible, because there is no longer any cheap way to tell an honest stranger from a lying one. The electricity is the cost of the mechanism. The cheap check is the mechanism.

Replacing the wasted arithmetic with something useful is not a 2026 thought. It is roughly as old as the complaint. In 2013, a project called Primecoin ran a chain whose mining searched for long chains of prime numbers, a real open problem in mathematics. It worked, and it worked for a specific reason: finding those primes takes an enormous search, and confirming one somebody else found takes a short calculation. The asymmetry survived the swap.

That is the test any useful-work proposal has to pass, and almost nothing passes it. Most valuable work in the world is roughly as expensive to check as it was to do. Reviewing a contract takes a lawyer about as long as writing one. Checking a diagnosis means examining the patient. So the useful-work chains stayed a curiosity, not because the idea was silly, but because the menu of tasks with the right shape is very short.

AI inference is not merely on the hard end of that menu. It is close to the worst thing you could pick, for four reasons that stack.

The first is arithmetic. The obvious way to check an answer is to run it again, and running it again costs exactly what running it the first time cost. Check everything and you have doubled the price of every answer, forever. No industry at planetary scale survives doing everything twice.

The second is that two honest machines disagree. Floating point addition is not associative, thousands of cores on a graphics card race and finish in a different order each run, and the results differ at the far end of the decimals. Honest disagreement looks exactly like lying, unless the network pins chip generation, driver and build hard enough to force the rounding errors to be identical.

The third is that the model is supposed to vary. A language model picks each next word from a spread of probabilities rather than a fixed answer, so the same question, the same weights and the same machine can legitimately give you different sentences on Tuesday than on Monday. A checker that flags difference as fraud will flag the product working correctly.

The fourth is the one that makes the other three expensive rather than merely annoying. Nothing in the answer tells you which model produced it. Text from a large model and text from a smaller, cheaper one both read like text. An operator paid per answer has a standing financial reason to quietly serve something smaller and pocket the difference, and the customer has no way to notice from the output alone.

Put those together and verify that the work was completed correctly stops being a clause in an announcement. It is the open problem this entire corner of the field has been working on for years.

Three approaches exist, and it is worth knowing what each one charges.

You can accept every answer immediately, make the operator post money that can be taken away, and leave a window in which anyone may dispute. Cheap in the normal case, because almost nobody ever checks, and the threat of the check is what does the work. It needs someone watching, and it needs the machine to be able to give the same answer twice.

You can produce a mathematical receipt, a proof that a specific model ran and produced a specific output, checkable in milliseconds by anyone. In July, Lagrange's DeepProve-1 became the first system to prove a full transformer inference end to end. The model was GPT-2, about 124 million parameters, and it had to be converted from decimals into whole numbers before any of the cryptography could begin. Hold that number against the models an agent would actually want to call, which are thousands of times larger.

You can seal the computation inside a chip that signs a statement about what it ran. Fast, real, in production today, and it relocates the trust onto a chip vendor rather than removing it.

And there is a fourth thing that is not really verification, which is what a network serving a large model to strangers is doing right now. On Bittensor's subnet 53, operators commit to a record of what happened inside the model while it was running, and validators audit slices of that record against a signature tied to the exact bytes served. That is a spot inspection with money behind it, not a proof, and it was built that way by people who could not afford any of the other three.

The contrast that makes the whole picture legible is a project called Pearl, which started a main network in April with a different bet: its miners run large matrix multiplications rather than hashes. That choice is not decoration. Checking a matrix product is genuinely cheaper than computing one, so the old asymmetry survives, and the network can pay for work it can afford to police. It is also why the useful work on offer is a mathematical operation rather than an answer to your question. The closer the work moves to what a person actually wanted, the harder the checking gets.

That is the pattern, and it is worth a name: checkable, not just useful. An open network can only pay for work whose checking is cheaper than its doing. So the menu of things such a network can buy is not the set of useful things. It is the much smaller set of useful things that happen to be cheap to verify, and the two lists overlap far less than anybody wants.

The shape is older than computing. A factory pays piece rates because widgets are countable, not because counting widgets is a good measure of a good worker. An insurer covers the losses it can confirm rather than the losses that hurt most. In every case a system that has to pay strangers ends up organised around what it can check, and the value of the work quietly becomes the second consideration.

None of this makes the announcement dishonest, and it is worth being precise about that. There is no published design here to be wrong about: no technical documentation, no tokenomics, nothing stating how nondeterminism would be handled, how a sampled answer would be compared, or what a wrong answer would cost the machine that gave it. Refusing a presale and refusing venture allocation is a real commitment and an unusual one. Proof of useful inference is a serious research direction with serious people in it, and someone will eventually make a version of it work.

What is fair to say is that the hardest problem in the field arrived this week inside a subordinate clause, and that the ordering of the plan puts a token in circulation a quarter before there is a network to check anything on. The first thing this network distributes is not compute and not answers. It is a claim on a network that does not exist yet.

Which leaves the question actually worth watching, and it is answerable within months rather than years: does the verification design arrive before the token does. In this field, the projects that got anywhere published the mechanism first and the market second. When it goes the other way round, the thing being priced is the promise, and a promise is the one product nobody has ever managed to check cheaply.

Whether Flop Labs publishes a verification mechanism before the fourth-quarter airdrop, and specifically whether it says anything concrete about nondeterminism, sampled outputs and what a wrong answer costs the operator who gave it.

And whether anyone puts a number on what checked serving actually costs today, against unchecked serving. The supply side of verification has been publishing figures for a year. The demand side still has none.

---

The two voices are AI. The research and writing are mine.

Decentralized AI, layer by layer.

Dastan,