Bittensor Weekly — October 9, 2026

The subnet opened a network for motion-to-meaning intelligence this week, starting with /sign: live two-way translation between Deaf and hearing people from a standard camera. No gloves, no wearables. Signing becomes speech; speech comes back as captions and signing. Yahoo Finance, Benzinga and Business Insider all ran it within a day, and SN78, formerly the Vocence voice subnet, has become one of the network's most closely watched builds.

The models come from an open competition underneath: King of the Hill rules, every score public. Cohort 5 drew 178 submissions, and evaluation is past the halfway mark, with 65% of endpoint challenges answered at the last checkpoint. The team scaled capacity mid-round to keep the remaining work moving.

The people behind it: a dev lead who ran the nucleus team at OpenTensor for two years, and a medical-side founder with ASL users and advocacy groups already in the room. Sign language has been the afterthought in translation tech; here it is the first product.

Conjectures (SN66) paid its first contributor, and the math is falling fast enough to argue about.

The payout cleared Thursday: $500 for a Green 24 contribution that formalized the sharp lower bound γ ≥ 1/3 in Lean, settled as 582.028422659 SN66 alpha through the team's 2-of-3 treasury multisig and verified on-chain. The contribution repo behind it pays out for original, verified Lean work that helps someone crack a conjecture.

Then came the pace problem. One crew swept three problems in three hours, and miners grinding a week-long Erdos problem watched their payout shrink. The team tuned within hours: new problems start at half reward while the long-standing ones hold their value. After an overnight error reverted the change, they set it again the next day.

Why math works as a mining target, in the team's own words: "The proof itself is the answer, and verification is binary, yes or no. That is why mathematics is such a natural fit for an incentive mining network. Open-ended search can absorb enormous compute, while success can still be identified with extraordinary precision."

The bounty pool grows by 148 alpha an epoch, twenty epochs a day. At this week's solving pace, it will not sit idle.

Cascade (SN91) beat Datadog's Toto on both of the benchmark's scoring rules, with 84.6% fewer training tokens.

Same architecture, same size, trained entirely inside the subnet, and never a fine-tune of Datadog's weights. Miners here do not touch the model; they compete on the data generators that feed it. The result: 519B reported training tokens against 3.36T for Toto 2.0 4M, with wins on TIME CRPS and TIME MASE, the two scoring rules for forecast quality where lower is better.

The scoreboard backs it: 235 miners entered the latest round, challengers must clear statistical acceptance criteria to dethrone the king, and the lead has changed hands 13 times across 131 rounds. A dashboard now charts each generation's best model against the official Datadog checkpoint on the same battery.

And the tooling reaches past the benchmark: Ephemeris, the subnet's forecasting API, bundles six models behind one request, and one miner pointed it at music, forecasting the next note of a track and feeding it into a live-coding engine. Forecasting as an instrument.

Hot. Conjectures (SN66) and ORO (SN15) tied at 130 messages to lead the week, ahead of Beam (SN105) at 110, Ditto (SN118) at 89, and Score (SN44) at 64. The volume tracked arguments as much as progress: Conjectures fought through a bounty-pool redesign, ORO through a provider switch and a contested race, Beam through a rules change it applied backward, Ditto through a crown that went the wrong way.

Building. Claims (SN111) ran a Nature Aging study through its pipeline and published it as 15 claims, each tied to its evidence. MVTRX (SN79) published its first numbers since the scoring change, with miner-supplied liquidity more than doubled and aggressive-order volume down three quarters. Herald (SN69) shipped v0.2.2, letting PR firms register a miner key from a wallet in one transaction. Affine (SN120) finished its pivot to audited math tasks, and its Round 9 exact-sampler now replays every token a submission produced, so copied answers fail even with valid compute proofs attached. Swarm (SN124) counted about $100k of signed pilot revenue and lined up Sentinel, a solar-park security task, for its drones. Cacheon (SN14) did $100k in revenue last month and posted 528.9 tokens/s on Affine's model, a 40% speedup, with a deepseek v4.1 flash arena on B300 GPUs next. Instant (SN46) published head-to-head speed tests with raw answers and said the burn goes to zero next week. ChronoLLM (SN38) took 87% of 650 duels against a bigger Llama at round 14, and from round 15 the leak scorer is back at 30% of weight. Cortex (SN100) published its cost base and cleared a 500-deep bug-bounty queue. RedTeam (SN61) got its final Kraken document and unblocked buybacks after months on hold. rec4ll (SN31) closed its external security audit with nothing critical open. Endure (SN30) went live as a risk desk for alpha tokens and fixed a launch-week registration flood inside a day. Cathedral (SN94) ran Affine's public sandbox spec and posted the receipts: 0.20s sandbox ready times against a 10s bar, and 1,000 sandboxes in 18.7 seconds. SayGM (SN28) launched Proof of Customers, paying affiliates in alpha for the revenue they bring. Connito (SN102) signed its first customer for a clinical expert serving doctors and patients in underserved regions.

Cold. Same shape as most weeks: a dozen teams carried the news, and the rest of the map watched.

Friction. Beam (SN105) rewrote its selling rules on October 6 and applied them backward. Version 101 introduced sale-adjusted emission weights, one sale of up to 80% per week before the multiplier bites, and the initial calculation reached into activity from before the announcement. Miners who held and built watched ranks slide while miners who never sold climbed. Some welcomed the cleanup; others asked why a rule starts before its own announcement. The team's position: the security and integrity of the network are at stake, and they are not soliciting feedback on it.

ORO (SN15) disabled Chutes as an inference provider from Race 7 after miners showed Chutes runs could surface hidden requirements and skew crowns. The fix was popular; the anti-cheat rejections that came with it were not, landing without line numbers or rule names, and the runtime turned over three times in four days.

Ditto (SN118) spent the week untangling a crown that miners say the margin math got wrong, with validator weights lagging the shared scoring ledger and submissions rejected as "unreviewable." The team stayed live in the channel through it, tracking the fixes.

Ridges (SN62) spent the week under a flood of near-identical agents from a few coldkeys, slowing screening for everyone else. The team is drafting a policy.

Beam (SN105) spent the week in a fight over data access: full miner telemetry behind a paid API, a million requests a day as the contested ground, a leaderboard that stopped showing everyone. Then a miner named Ψοοτ made an offer. Hand over the leaderboard data, and he would build it for free, make it public, remove the rate limits for everyone.

He meant it. "So, I took on that processing task for all miners, free of charge."

Within the day, the team opened up in return: the dashboard now shows every miner, and the audit API is free. "Robin Hood of bittensor, the hero we didn't know we needed," one member wrote. Ψοοτ's reply was two words: "thank you."

A paywall argument that ended with more data public than before it started. Communities are supposed to work like this; it keeps being worth saying when one does.

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]