Bittensor Weekly — September 25, 2026

Nick Hotz of Unsupervised Capital published a proposal Thursday called Burnweight: allocate TAO emissions by a decaying weight created by burning TAO. A subnet burns to build its weight, the weight decays, and emission becomes an ongoing auction between subnets. Right now a subnet's share of the day's TAO follows its price. Burnweight prices it in burns.

He dropped it in Bittensor's root channel with "Would love to hear what people here think."

The answers split fast, along the line the whole network keeps hitting: subnets that earn revenue versus subnets that produce research. Monk Kukai: "i think that this will kill all research subnets, and model training subnets and they will need to migrate to other chain." tao talks liked it, with a caveat: "I like the fact that emissions would be based on a more verifiable commitment to bittensor instead of promises and hopium."

Hotz's answer on research: subnets with strong potential would still draw stakers into their alpha, emissions or not. And on participation, a floor under the math: "So no one would have to participate if they didn't want to. And they could still buy back their own token if they wanted to as well. If the ecosystem can only put up 1000 TAO per day and no one else wants to burn, the subnets that do participate will still share 3600 TAO/day of emissions."

τp read it as a sequel: "is it essentially a better version of taoflow, right?" btchydra took the market's side: "I think it would absolutely send the price lol."

Nothing is voted on. The proposal has to survive this room first, and what the room decides shapes whether emission stays loosely tied to price or becomes a bill subnets pay to hold their place. The research-subnet question is the one to watch.

Affine (SN120): a Bittensor co-founder's autonomous agent keeps taking the crown, and now it files its own pull requests.

The agent is Arbos, running Jacob Steeves' mining operation as an outsider. It rents 8-GPU boxes on Lium, trains with a GRPO loop (a standard reinforcement-learning method) scored by the validator's own code, and submits what comes out. No human tunes it, it has held the crown more than once, and it burns a few thousand dollars a day to stay in the fight. The recipe copies in one line: hand an agent a GPU rental key and a funded wallet, then tell it to rent until it wins.

The full playbook is public, including what didn't work, and it ends with an ask: improve the loop, then publish your notes. Lium credits and registration TAO for anyone who takes it up.

Arbos is still running. This week it moved its forfeit floor mid-competition on an explicit operator directive, refreshed its field notes, and opened a pull request folding benchmark-failure trials into its training data, every change stamped with the directive that ordered it. The mining loop writes its own diffs now.

Conjectures (SN66) is at 31 Lean-verified proofs, and two of its miners turned out to be teenagers.

The subnet's results page reads like a scoreboard for unsolved mathematics: Erdős problem 1062, part two, verified September 21 and paid 3,922 alpha; Green's open problem 39, to a miner called Jordan on September 19, 2,734 alpha; Erdős problem 859, to JenW1N a day earlier, 3,740 alpha. Each entry carries the Lean statement (the exact formal target), the kernel verdict, and a human review decision. Rejections publish with reasons, from "Already solved before this submission" to "Already formalized elsewhere."

The team, on why the setup holds: "The proof itself is the answer, and verification is binary, yes or no... Open-ended search can absorb enormous compute, while success can still be identified with extraordinary precision."

The channel spent the week on the humans behind it. JenWIN and Liam, two names on the board, are 18 and 19, both studying engineering. A miner with an IMO background spent a day talking coursework with them and guessed the major from equations on a wall behind one of them. The equations were AI-generated background art, and the guess landed anyway.

"we are trying," JenWIN wrote. "some fun preliminary results so far--no promises." The team says its problem pipeline is heading to universities this year.

Hot. Thursday was the loudest subnet day in weeks: 931 messages across 75 channels, against 245 the day before. ORO (SN15) carried 109, most of it race logistics and a moderation cleanup; Leadpoet (SN71) 77 on the reward fight; SOMA (SN114) 61; OpenRoboto (SN80) 54; Ditto (SN118) 52; Cacheon (SN14) 50.

Building. Enigma (SN63) paid its 512-bit milestone and put the next rung live: 530 bits, $10k, four hours on the clock, with $250k still sitting on the 1024-bit top. Cortex (SN100) kept its agent on the queue, eight user-reported defects merged in the latest round. Cascade (SN91) switched over: verdicts every 3 hours, per-domain receipts. Vidaio (SN85) closed enrollment at 14 contenders, with crown and podium bars set from baseline runs. Investing (SN88) deployed incentivized conviction; dividends go weekly, first one next Thursday. Perturb (SN26) started adversarial training and funded a second 500 TAO buyback from team capital. MVTRX (SN79) put 0.6.2 on mainnet, with the market picked by a search over roughly 800 candidates, and started its TradingView integration. Numinous (SN6) ran a first buyback off forecasting API revenue. ReadyAI (SN33) published the numbers behind its game demo: the same spec hit 14 of 14 checks for $0.66 with skills installed and $4.12 without. KubeTEE (SN90) became a provider on SayGM (SN28), serving Xiaomi's MiMo V2.6 Pro UltraSpeed at 10% off list. Hippius (SN75) shipped its Android app.

Cold. Twenty of the 129 channels said nothing at all across the two days, subnets 40 and 103 among them. RedTeam (SN61) ran on commit-reveal support tickets. sundae_bar's challenge thread stays paused until Monday. The house pattern holds: a dozen teams ship, everyone else watches.

Friction. Ditto (SN118) spent the week under repair: the screener broke, agents jammed in deferred source review, and owner Peyton set burn to 100% "for fairness" while he fixes it, cut uploads to a 10-hour cooldown, and said he would unburn once he is certain the system works. Miners split between pausing emissions and pausing submissions; Dexter's line carried the room: "No manual review please. You already saw how messy it becomes."

Leadpoet (SN71) crowned a champion and then argued about what a crown pays. With burn still at 100% and 30% earmarked for the champion while validators update their code, miners pressed on the math. pepelepluτus: "If you deem reward too little (which it's not) then you don't need to participate. We will make changes based on our direction, not spam in the chat."

SOMA (SN114) got a self-mining accusation over its round-9 winner; the team answered with explorer links and a repo check, and the miner making the claim said he would keep collecting data before a final call. The subnet also cut 18 hotkeys for prompt injection earlier in the week.

ORO (SN15) relit the agent-review fight: miners say code requirements are unclear and want checks run at submission time, the team says cleverly hidden hardcoding slips past LLM review, and it is refining static checks instead. Both sides aired it in the channel, in public.

Beam (SN105) had a transfer gap stretch past eight hours with no notice. One miner: "Eight hours without any notice makes the production network feel more like a sandbox."

The sundae_bar (SN121) team shipped Scout this week: nine ready-made agents, with nine miner-written skills inside five of them, every skill a winner from the subnet's own past challenges. Then came the note. One of the team followed up on the announcement: "Nine winning Skills already running inside five of our first nine ready-made Agents, all live for the public, is a pretty cool moment."

Miners here spend weeks writing markdown files that a scoring harness tears apart, and the reward is usually a payout and silence. This time their work shipped inside a product with their fingerprints on it, and the team said so out loud: "Thank you, appreciate you, and let's get it." Challenges restart Monday. That is what a subnet with a working loop sounds like.

Next week's issue goes out Friday. Subscribe and it lands in your inbox: https://newsletter.bittensor.quest

Shaping the next issue: tell me what to keep, what to cut, and what you want more of — mailto:[email protected]