Reliquary pays miners for verified RL training data: rollout groups (several attempts at the same prompt) that a GRPO trainer can learn from. This week the team moved to a steeper top-heavy curve and started burning half the subnet's emission (burn: tokens destroyed instead of paid out), on the argument that miners now generate far more data than the trainer can consume.
RomAI, the subnet's lead, put it plainly: "Emissions are not leading to more innovation, just more gpu and spending, we had to fix that." On the tradeoff: "We burn 0.5 now and the training is still going at the same speed."
The miners spent Thursday arguing the other side. partouches, who runs a single H100 at the top of the MATH environment, called it a hardware race replacing an optimization race, and said fleets of six or seven hotkeys per environment keep losing to one-hotkey miners. elhim_2 ran the economics in public: the biggest MATH holder had about 10 of 336 seats, worth roughly 0.267 TAO a day, about $59 at the time, against $57.60 a day to rent an H100. Then: "It seems that it's time to leave this subnet." Vincent: "Time to say good-bye. Will give up." tiny: "I worry the sn could start feeling more like self-mining, which may push smaller miners to give up on the sn."
The team's answers were specific. Per-hotkey caps are "exploitable too easily" because one person can run hundreds of hotkeys or IPs. Fixed collection windows waste two thirds of the time, since the trainer steps every 30 to 50 seconds. Putting k (the number of groups per prompt) into the incentive is on the table, parameters will get tweaked, and new task types are coming with their own compute attached. 0xgrizz framed the money: "Emissions don't originate from our bank account but from the value accrued by individuals who hold bags."
Cutting what a subnet cannot consume is correct, and the team is the only party that can see the training queue. A top-heavy curve with no submission cap also turns a small subnet into a hardware auction, and the supply side answered by saying out loud that it would leave. The next training run settles which of those two facts matters more.
Reliquary's owner and its miners trading numbers over the new incentive curve, September 17
Subnet Spotlight
Cacheon (SN14) crowned its own kernel, then told miners to submit a better one.
Cacheon is a kernel-optimization subnet: miners ship Docker images serving Qwen2.5-72B, and the best Time to First Token and throughput wins the crown. This week the crown went to a bundle the owner team built. Miners replied that the evaluation had gone down for four days while the team's own slot sat at the top, and asked for the crown to be reverted, for evaluation GPUs to stay off self-mining, and for evaluation to run 24/7. sky.u: "your team crowned new slot by your solution." Viper: "you need to care about miners' effort."
vune, the owner, did not soften anything: "Literally just submit a better submission." On redistributing the crown's emission: "We're absolutely not redistributing to random miners that are suboptimal. That's free lunch." On the resolution: "I'll commit to locking 100% of the emission we're getting from this first slot" and "we're absolutely not reverting the improvement." His read on the channel's complaint: "It's obvious that the previous submissions are from your mining team and you're upset you're not getting emissions still." Another miner, ai-hpc, backed him: "this subnet is highly domain expert's subnet."
Locking the first slot's emission is the right call and should have been announced before the crown landed, not after. A kernel contest where the host also competes needs every submission published as it is revealed. vune's answer: "You can view all submissions as they are revealed and garner info from them." The fix going forward is more slots, each announced before work starts on them.
Cacheon's owner and miners going back and forth on the self-mined crown and locking emission, September 17
Niome (SN55): a miner found a hole in the scoring and published it instead of farming it.
A Cacheon-style writeup showed up in Niome this week, from a miner, not a team: "Scoring integrity vulnerability: round seed exposed during the submission window." The seed used to grade a round could be read while submissions were still open, and matching it drove consistency_factor (the check that catches copied or degenerate answers) to a clean 1.0, switching off the integrity gate entirely.
He measured it before reporting. Honest submission, no seeds matched: consistency 0.11, final score in the 26 to 36 range. Three of three matched: consistency 1.0, score 217 to 294. Roughly eight times the payout for a copy of the grader's randomness, which is the difference between a working subnet and a farm.
The interesting part is what he did with it. Running three hotkeys across UIDs 137, 72 and 37, he could have sat on the exploit for a month. He posted the reproduction, the impact table and the severity rating instead, then spent the evening talking shop with another miner about it.
The Niome scoring audit as posted by the miner who found it, September 17
Sentiment Roundup
Hot. Thursday was the loudest day subnet chats have had in a while: 1,213 messages across 63 channels. Cacheon (SN14) carried 210 of them, all of it the self-mining fight. Ditto (SN118) had 150, Conjectures (SN66) 93, NeverPlayAlone (SN98) 75, Reliquary (SN81) 68.
Building. Conjectures (SN66) answered the "there's no moat here" jab with a roadmap and a plan to sell problem placement to orgs. Swarm (SN124) shipped the fix that made every validator fly identical maps, cut seeds per epoch from 1,100 to 1,000, and paused evaluations to replace hand-corrections of zero scores with one fixed rule. Apex (SN1) spent Thursday on a scoring dispute, a later and lower-scoring submission sitting above an earlier one, and its dev pointed miners at the auto-stake tool: emissions go straight to a validator hotkey, no fee, and most miners still don't use it. ORO (SN15) published an anonymized, categorized list of every behavior that got agents kicked at its race gate, with code examples of grader-mirroring (an agent borrowing the scorer's own logic instead of doing the shopping task). Pareton (SN10) opened its first SGLang campaign: RadixArk's Qwen3.8-27B on 4x RTX 5090 with 262k context. SayGM (SN28) ran its first buyback and burn, $43k of alpha, monthly from here. Vidaio (SN85) confirmed its evidence bucket is publicly readable, and a miner pulled the grader's input bytes straight out of it and matched them byte-for-byte with the digests it computed locally.
Cold. Subnet 103, Subnet 40 and SN21 have nothing new in them, and sundae_bar's challenge thread is still flagged paused. Half the network's channels were dark on the loudest day of the week. That is the standing house pattern: a dozen teams ship, the rest watch.
Friction. Ditto (SN118) spent the week on its review queue. The owner revoked the second reviewer's access, his review loop stopped overnight, and miners watched flagged agents hold top slots through the gap. Proposals on the table: automatic review for anything landing in the top five against the median, or manual review for every agent ever rejected. NeverPlayAlone (SN98) got the accusation directly, and the owner answered it flatly: "We are NOT self-mining." The room countered that the same miner wins every round and asked for the champion's code to be published at the end of each one.
Numbers That Matter
1,213 messages across 63 subnet channels on Thursday, against 176 the day before. Cacheon 210, Ditto 150, Conjectures 93.
$240.85 for TAO, up 7.6% on the day and a couple of percent on the week, with a dip to $218 on Wednesday. Market cap $2.73B.
0.5 is the burn ratio Reliquary runs now, with the team claiming identical training speed.
0.267 TAO a day is what one miner calculated the top MATH seat pays, about $59, against $57.60 a day of H100 rental. Miner margins on that subnet are a rounding error.
$43k of alpha burned in SayGM's first buyback.
1,000 seeds per Swarm epoch, down from 1,100, on maps that are now identical across every validator.
26 versus 294. The score range on Niome depending on whether you copy the round seed.
262k context, 4x RTX 5090, TP 4. The hardware spec for Pareton's new SGLang campaign.
Notable Quotes
"Emissions are not leading to more innovation, just more gpu and spending, we had to fix that."
"I'll commit to locking 100% of the emission we're getting from this first slot... The emissions will not be changed arbitrarily."
"The moat is building the pool of incredibly talented miners, and then getting paid problem placement. This current problem set is meant to help us build the network of smart miners where we can go to orgs and get them to pay us for the problems."
Human Moment
SayGM (SN28) had a week worth noting. Its dev spent hours on one miner's setup, through a hash mismatch, a version bump, a scope verification error and five hours of the miner's own frustration, until the reply landed: "thanks @Poggle I successfully run miner." Nobody gets paid for that. Then came the other half: an epoch whose traffic numbers came in well below the others, spotted by a holder, answered with "the numbers on the epochs have now been corrected and the underpayment added to the impacted miners accounts." A small subnet crediting back money it could have kept, and a dev who answers every message himself. That is the part of this network worth holding.
SayGM correcting an epoch's numbers and paying the shortfall back to miners, September 17
Coming Up
Reliquary (SN81): new task types with their own compute, the k value under test in the incentive, parameter tweaks, and a next training run that starts with SFT and a higher in-zone rate.
Cacheon (SN14): more slots, each announced before work starts on it. The first self-mined slot's emission stays locked.
Swarm (SN124): evaluations reopen once the zero-score rule is deployed, every model gets re-evaluated on the shared maps, and seeds are published after the epoch transition.
Pareton (SN10): SGLang campaign open for submissions, campaign ID a073b97b-f194-43df-8a1a-61c6ec1b8289.
ReadyAI (SN33): SS33, a storefront where builders and agents can find and combine the skills the subnet generates. Their proof of concept, a Calendly clone called readyCal, was built and deployed in a single run.
Trishool (SN23): a miner says tri-judge carries a bug being exploited in the current challenge (judge-result.ts forcing any verdict that reads like a refusal down to 0), and asked to take it off DM. The team asked him to post the reproduction in the channel.
Apex (SN1): race 2's qualifying threshold sits at 0.55, and miners are asking for it to come down to 40% before the next one.
Worth Your Time
https://www.bittensor.com/docs/tx/set-auto-stake — auto-stake, straight from the docs. A miner hotkey can point its emission at a validator hotkey with no fee. If you run miners and stake by hand every week, this ends that.
https://github.com/vllm-project/vllm/pull/57140 — merged into vLLM this week: a speculative-decoding change that drops one output-sized allocation and the final full-output copy for Qwen GDN. The kind of optimization subnet miners spend all day chasing, landing in the inference stack everyone else runs.
https://conjectures.io/roadmap — Conjectures' roadmap, including verified software optimization as the next problem class. The team's moat argument reads better with the plan in front of you.
https://swarm124.com/benchmark — Swarm's open benchmark. Every task family is a king-of-the-hill ladder, models come from independent developers, and the seeds are released after the epoch transition so you can check the flight yourself.
https://arxiv.org/abs/2609.13443 — "Learning to Solve Hard Problems in RL for LLMs by Never Giving Up." RL improves easy problems and stalls on hard ones, so the authors sample where the model is failing. Reliquary's team linked it as directly relevant to how they pick prompts.
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]