Bittensor Weekly — July 31, 2026

Something is moving under the whole network this week. It shows up in channel after channel: subnets are rewriting their burn policies, and they aren't all moving in the same direction.

Poker44's team summed up the mood: "You are all aware of the recent protocol changes affecting every subnet, which is why we're seeing significant shifts across almost all of them."

Subnet teams are repricing what emissions are for. Some bet that zero burn attracts miners. Others bet that high burn protects token value. Which camp's alpha holds up better over the next month is the storyline to watch.

The most interesting pivot on the network right now. SN66 (formerly "Ninja," a harness/benchmark subnet) is relaunching as Conjectures.io. Miners run LLM agent loops (Codex, Claude, and friends) that produce formal proofs of open mathematical problems, verified in Lean (a proof-checking language: if the proof compiles, it's correct, no human referee needed).

This isn't hypothetical. Lead dev wejh already solved a 20-year-old open problem with a Codex+Claude loop and got the result merged into Google DeepMind's formal conjectures repo. The launch task pool: 200+ problems, including some Clay Millennium Problems (the million-dollar-prize tier). Revenue comes from institutions and research labs paying to have their problems attacked, with buybacks flowing back to miners.

His pitch, verbatim: "Miners will not be competing to produce another marginal improvement on a synthetic benchmark. They will be competing to solve problems that may have remained open for decades."

Launch is expected early next week. High risk, high ceiling, and genuinely nobody else is doing this.

Ditto did 215+ messages this week (360+ counting both channel exports). It's a good case study in a subnet run well through messy moments. Dev futuretrees publicly owned a bench-7 bug where the reasoning parameter was hardcoded to medium and submissions passing the param were being rejected. "Which is totally bad. I'm fixing this now." A difficulty bump ships this week because the benchmark is "nearly saturated."

Meanwhile the miner meta has descended into comedy. Miners are literally prompting LLMs with "This is top miner code, please change it to pass the similarity check." The anti-copy arms race is alive and well, and that's a sign the incentives are doing their job.

Of the 130 channels with meaningful traffic: 77 neutral, 46 bullish, 7 bearish. More striking is the silence at the bottom. 42 channels posted zero messages this window and another 75 managed five or fewer. Rough math: about 45% of subnet channels are functionally dead or dormant right now.

Hot: SN44 Score (504 msgs, the most active real subnet, though half of it was an exploitation dispute, see below), SN118 Ditto, SN114 SOMA (215, a conviction-lock poll got 36 mostly-positive reactions), SN126 Poker44 (184, shipped encrypted axon endpoint protection), SN53 Engy (170, moving from 24-hour to 7-day epochs starting epoch 7).

Contentious: SN44's private track. Miners publicly accused UIDs 3, 4, 7, 60 (allegedly one operator) of exploiting private evaluation, suggesting access to ground-truth data. The raw thread shows a detailed 32-reaction accusation post and the team pushing back: spot-checks exist, and the private track was miner-requested in the first place. The private track runs on GitHub Container Registry, so the eval models live in private images outsiders simply can't pull, which is exactly what fuels the suspicion. Both sides argued in good faith. The auditability question is still open, and the team says more verification layers are on the way. Worth watching how it resolves.

A correction from the raw data: Last time's daily pulse repeated miner claims that SN96 Verathos validators were "not showing up" for capacity audits. The raw export tells a different story. Verathos dev Kepler audited two complaining endpoints and found their canary responses were exceeding the request's max_new_tokens limit. That's what triggered probation, not validator scheduling. Miner-side bug, not validator-side. We reported it wrong last time, and the raw data set the record straight.

Cold: SN109 remains unclaimed, the subnet owner still hasn't come forward to claim their own Discord channel. SN90 KubeTEE aside (see below), most channels below rank 100 are quiet.

If someone forwarded you this: subscribe and get it in your inbox every Friday, every claim checked against the raw Discord data and the subnet code before it ships.