I operate miner neurons on several Bittensor subnets. Not as an investment, as an operator. I run infrastructure that competes daily against other operators for work that a subnet's scoring code decides how to pay for.
A year of doing this has changed how I think about AI quality control. Not because of anything anyone promised, but because of what the payment schedule does to behavior, mine included.
Some background, compressed. Bittensor is a network where machine intelligence is organized as a market. Each subnet defines a task, inference, data generation, content verification, media placement, and pays the operators who perform it best, in TAO. Since the dynamic-TAO upgrade, every subnet has its own token and its own economy, and TAO holders decide where value flows. There are more than a hundred live subnets.
Now the part that matters:
1. When payment follows measurement, the measurement becomes the job.
Every subnet is only as good as its scoring rule. I have watched operators, myself included, optimize toward exactly what the scoring code measures, sometimes at the expense of what it was supposed to measure. That is not a bug unique to crypto; it is Goodhart's law running on a public scoreboard. The subnets that survive are the ones whose maintainers keep tightening the gap between "what is measured" and "what is valuable."
2. Cutting off non-performers turned out to be routine, not drama.
In June 2026, Bittensor's leadership began halting emissions for inactive or non-functional subnets; by July it was permanent weekly policy, and dozens of subnets have been cut since. I expected outrage. What I saw instead was operators treating it as weather: check the list, adjust, move on. A market that actually stops paying dead projects is rarer than it should be, in crypto or anywhere else.
3. Vesting changes what you're willing to ship.
The clearest example I know is the earned-media subnet. Operators commit to a placement before publishing, payment vests in daily installments over a month, and if the published article is quietly deleted, or quietly converted into an ad, the remaining installments are gone. Compare that with the traditional press-release economy, where the invoice is paid the day the wire goes out. When your income is back-loaded and clawable, you stop asking "how do I get this placed" and start asking "will this still be standing in thirty days." That is a better question.
4. Decentralized entry does not mean distributed outcomes.
On several subnets, the media one included, a few operators capture most of the mining rewards. Open doors do not guarantee spread-out winnings; they only guarantee that nobody can lock the door afterward. I think that distinction gets lost in most decentralized-AI threads.
5. The scoreboard being public is the underrated part.
Every claim I make here can be checked against chain data and the network's own repositories. Whatever you think of token prices, and most subnet tokens have sold off hard for months, I hold some of them, the experiment itself is legible in a way that closed AI labs are not. You can watch who produced, who stopped, and who got paid, in public, every day.
The question I care about is whether incentive engineering can reliably produce quality at scale. A year of operating says: it produces effort reliably, it produces honesty about performance reliably, and it produces quality exactly as good as the scoring rule allows. That last part is the whole game, and more people should be watching how it's played.
I hold TAO and subnet tokens. Nothing here is financial advice; it's an operator's field notes.