October 5, 2026 · Published daily
Subnet News — October 5, 2026
SN116 Carbon announces it is live and publishes eight industrial challenges. SN80 OpenRoboto allocates 35% to egocentric data, SN41 Almanac reports forecasting gains, and SN78 UMI reopens intake with rewards still inactive. Plus SN108 ChipForge’s actual launch status, new miner and evaluation updates, SN38 ChronoLLM’s Round 14 proposal, conviction events and refreshed videos.
- Coverage window
- Oct 4, 2026, 1:49 PM – Oct 5, 2026, 11:10 AM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 21
Biggest stories first
Today's ranked news
SN116 Carbon announces it is live and publishes eight industrial challenges
SN116 Carbon announced it was live at 18:26 UTC on October 4, just after yesterday’s news cutoff. Today it published an initial eight-challenge physics AI portfolio: server cooling, chip thermal management, optical links, battery charging, precision motors, low-vibration robotics, acoustic equipment and passive mixing. The idea is to help engineers screen designs and make decisions with evidence, rather than treating every AI prediction as trustworthy. The portfolio is proposed research, with proof requirements attached to each use case. Its dollar-saving examples are hypothetical customer benefits, not achieved results, eight customers or Carbon revenue. The launch announcement advances yesterday’s registration story; completed rewarded mining and customer deployments were not independently demonstrated.
SN80 OpenRoboto’s published rules allocate 35% to egocentric data
SN80 OpenRoboto has updated its published allocation to 15% simulation, 25% real-robot π0.5, 25% real-robot LingBot-VLA 2.0 and 35% egocentric data collection. That makes accepted first-person data a substantial part of the incentive model alongside training robots. The data score combines accepted minutes, task rate, quality and novelty; weekly weights use the previous week’s accepted contributions. Invalid data earns nothing. The same documentation specifies a 0.2 TAO simulation submission fee, checked against the selected season before payment. These are current published rules, not a notice opening a new physical-robot competition; this review did not independently reconstruct the chain’s applied allocation. The Open Axis partnership post is a September recap, not another new launch.
SN15 ORO updates its runtime and reportedly begins duplicate-agent enforcement
SN15 ORO validator v2.2.0 adopts oro-env-runtime 3.3.0 and runtime contract 2. The release aligns stock tasks with the current contract and rejects incompatible older practice packs before loading them. Separately, today’s team-attributed brief says enforcement is rolling out against the same or trivially varied agent submitted across multiple hotkeys. Miners who think a flag is wrong can submit their hotkey for review, and additional checks may remove entries before the current race ends. The runtime change is verified in code; the enforcement rollout remains a secondary team report. Yesterday’s Race 2 recovery does not independently close the separately reported Race 11 pause.
SN26 Perturb reportedly shifts scoring toward adversarial robustness
Today’s team-attributed brief says SN26 Perturb now weights evaluation 70% adversarial accuracy and 30% clean accuracy, replacing the previous 50/50 mix. That would reward models more heavily for handling deliberately difficult inputs, rather than only normal examples. The brief also reports consensus-rank changes for scanning emissions and asks validators to update and restart together. The original announcement and full validator rollout were not independently located, and current repository attribution remains uncertain. Treat this as a reported scoring change rather than a verified chain-wide activation.
SN41 Almanac reports better forecasting accuracy and more repeatable evidence
SN41 Almanac’s weekly development report puts its Jev-integrated forecasting harness at a 0.165 Brier score versus 0.169 for standalone Claude Sonnet; lower is better. It also reports source-URL agreement between repeated runs improving from roughly 80% to 93%, with probability outputs stable to the third decimal place. Those are useful measures of forecasting quality and reproducibility, rather than a vague claim that a model is smarter. Alma has entered a September 28–January 6 tournament with a $30 daily budget and approximately $1 median forecast cost. The team’s results have not been independently reproduced, and execution still takes about 40 minutes. Faster inference and a self-improving harness remain work in progress.
SN50 Synth tightens forecasting backtests and avoids blocking on every asset
SN50 Synth’s synthlib code now requires an offline market-data bundle for backtesting and moves the evaluation cutoff before the forecast horizon, reducing the risk of accidentally using future information. It also reads the published champion tree rather than relying on a stale local model. Asset warmup runs in the background: an asset can be served once ready instead of waiting for every asset to finish loading. These are useful reliability and evaluation changes for synthetic forecasting. They do not establish a new predictive-accuracy record, commercial revenue or an independently measured faster production service.
SN61 RedTeam plans an ADA-3 replacement as scoring issues continue
SN61 RedTeam’s team-attributed brief acknowledges anti-detect-browser and device-timeout issues affecting ADA-3 scoring. Failed sessions cannot simply be excluded under the current process. A replacement challenge is being prepared, but the brief says it will be introduced after miners solve the current BEX challenge. Separately, today’s public v4.10.9 code raises the number of unique commits allowed for ADA detection. That is a specific submission change, not proof that the grading problem has been fixed or a replacement is already live.
SN71 Leadpoet adds bounded recovery after provider-credit failures
SN71 Leadpoet’s Arena code allows an owner-signed second attempt after a provider-credit failure only when the gateway proves the original call settled with zero charge. The miner must first restore credit on the same provider account; the scoring stage and frozen deadline must still be open. Uncertain charges, closed rounds and old failures without the new proof are excluded, and the original spend limit and two-attempt cap remain. This prevents a narrowly recoverable billing failure from automatically ending a submission. Separately, the team-attributed channel brief still reports duplicate-agent slot pressure and a planned capacity increase; the retry feature does not settle that access problem.
SN74 Gittensor’s SparkInfer v0.6.28 closes a scoped serving gap
SN74 Gittensor has published SparkInfer v0.6.28. Its RTX 5090 test on Qwen3.6-35B-A3B reports 2,366 output tokens per second at 32 simultaneous chat requests, matching vLLM in that test and improving about 5% over v0.6.27. Median first-token latency is 318 ms versus vLLM’s 580 ms; vLLM still has lower inter-token latency. Packing incoming prompts together is the main improvement. The tests use 1,024-token prompts, 256 output tokens and two runs per cell on the same machine. The systems use different quantization formats, so this is a defined serving comparison, not a universal performance lead. Qwen3.8 is unchanged.
SN78 UMI’s intake is open again; rewards are still inactive
SN78 UMI’s public competition status at 15:04 UTC on October 5 showed admission open, capacity available and 203 accepted submissions. Yesterday’s observation showed admission unverified and not accepting new entries, so this is a meaningful status change. Assignment delivery, model intake and evaluation were marked ready, but rewards_active and chain_submission_authorized both remained false: accepting work is not the same as paying for it. New code also adjusts cohort processing budgets and carries chain state forward through runtime changes. Those repairs do not by themselves prove C5 has settled or miners have received rewards.
SN81 Reliquary improves usable dataset exports and terminal-task reliability
SN81 Reliquary’s new code exports bounded instruction datasets alongside its existing corpus deliveries, helping turn generated work into prompt-and-response training data. This applies to supported single-turn jobs; episode or unsupported prompt formats retain their original delivery rather than being forced into the wrong schema. A separate terminal-environment merge adds a default 180-second command timeout, per-test grading detail and cleanup of containers left by interrupted runs. Grading fixes also avoid drawing impossible cross-provider rechecks when only one provider is available. These are inspected implementation changes; the reported new software-engineering mining job and production deployment were not independently confirmed.
SN106 Nodexo team reportedly focuses on SN96 Verathos and explores a sale
Today’s team-attributed brief says the SN106 Nodexo team is now focused on SN96 Verathos, with hardware verification for a mesh of grouped compute resources and plans to operate its own inference service. It also says SN106 is in sale discussions with several candidate teams. The distinction matters: a discussion is not a completed sale, a new owner or an announced purchase price. This review did not locate a matched completed chain transfer or an independently tested SN96 service launch.
SN108 ChipForge ships clearer evaluation visibility; its challenge is not open yet
SN108 ChipForge’s October 4 update describes a redesigned live leaderboard with per-submission scores, quality gates, validator progress, downloadable logs and a view of which design the chain rewards. The team also says NPU tests now use trained benchmark models with hidden expected answers checked against Google’s reference software. At 15:05 UTC on October 5, the public leaderboard showed no active challenge, no submissions and all three validators’ stake burning emissions. Its API reported healthy. That is useful infrastructure progress ahead of mining, not confirmation that today’s planned competition has opened. The accompanying TatsuOS update reports end-to-end testing completed, with infrastructure deployment and BLE/OTA work still ahead of a preview.
SN120 Affine adds recovery for missed hourly reward records
SN120 Affine’s latest code reconciles missed hourly records using the learner’s original completion time, rather than counting old work as if it finished in the current hour. Previously published records are checked against their saved journal and retained unchanged. The practical purpose is to recover gaps without rewriting reward history or double-counting work. This is an implementation update, not independent confirmation that affected miners have been paid. Today’s operator status still describes a recovery job in progress and additional H200 nodes awaiting qualification and enrollment; preparation and tests are not a completed rollout.
Stillcore publishes its TAO 2030 investment thesis
Stillcore Capital published its TAO 2030 paper on October 4 after yesterday’s cutoff, explaining its scenario for a trillion-dollar-plus Bittensor ecosystem by 2030. Its argument centers on intelligence becoming cheaper, outside customers paying for subnet services, and value flowing through the network’s currency and markets. This is an investor’s valuation thesis with assumptions and milestones to track, not an achieved valuation or a guaranteed price target. Its revenue estimates are not added to yesterday’s already-covered revenue index as a second, independent pool of sales.
Exploit follow-up: Brandtensor outlines marketing support for subnet teams
Gordon Frayne’s October 4 interview post and Brandtensor’s response bring attention to the agency’s work on subnet positioning, product proof, developer education and reaching customers beyond the Bittensor audience. Its introduction was actually published September 25; this is a late-found conference follow-up, not an October 5 launch. The practical question is whether technical progress can be explained through clear use cases and verified evidence. The announcement does not establish new client revenue or independently audited results.
SN9 iota reportedly deploys a peer-discovery backend fix
Today’s team-attributed brief says SN9 iota upgraded its backend to redis-py 8.0.1 to resolve a peer-discovery problem that prevented miners from finding adjacent-layer nodes. This gives yesterday’s transient Redis recovery and staged-fix story a more specific reported implementation. The live backend and network-wide recovery were not independently checked. Macrocosmos’s distributed-training SDK remains a preview; this fix is not a full SDK launch.
SN21 AdTAO reports its second weekly conviction lock
SN21 AdTAO says its second weekly conviction lock is complete and reiterates a plan to commit roughly 25% of its owner emissions to Conviction. The announcement links a public lock portfolio. This is a reported recurring owner-emission commitment, not new advertising revenue, a guaranteed return or proof of an additional market purchase. The exact new lock amount and transaction were not independently matched during this review.
SN28 SayGM demonstrates its combined AI dashboard
SN28 SayGM has published a two-minute walkthrough of its shipped dashboard: chat across models, image generation, guardrails on API keys, custom model fusions and cascades, and spending tracked to the cent. The useful angle is control over how AI services are combined and paid for, rather than another standalone chatbot. This is a new product demonstration of existing shipped features, not evidence that every feature launched today or that an external revenue target has been reached. The authenticated product’s full workflow was not independently exercised.
SN38 ChronoLLM proposes faster Round 14 evaluation and persistent model caching
SN38 ChronoLLM’s open Round 14 pull request, created on October 4 and updated today, proposes batching duel generation and retaining downloaded models through evaluation. That would avoid downloading weights twice and prevent a model going private after qualification from blocking its next evaluation stage. The author reports roughly 20-fold faster batched generation, but this review did not reproduce that measurement. The pull request was still open at the check: these are proposed evaluation improvements, not a merged release, a confirmed new round or a verified live speedup.
SN44 Score reportedly removes a challenge with incorrect ground truth
Today’s team-attributed brief says SN44 Score removed petrol-station challenge block 9209400 after an incorrect ground-truth annotation was reported. It separately says rewards for the top Detect Worksite miner are functioning and that tiebreak rules apply. This is a targeted evaluation correction, not confirmation that every PPE annotation concern or yesterday’s warmup-related zero scores has been resolved. Vision Lab remains marked coming soon.
SN49 Nepher Robotics records a 46,437-alpha conviction lock
SubnetRadar records a 46,436.627-alpha lock on SN49 Nepher Robotics at block 9215815, timestamped 09:33 UTC on October 5. It targets the subnet owner’s hotkey. This is a lock and commitment signal; it does not prove a same-sized new purchase, revenue or any particular locker’s motive. Separately, today’s “Nepher Verified” certification post looks back to last month and is not a new regulatory approval or evidence of equivalence to CE, UL or ISO certification.
SN114 SOMA reportedly restarts 330 sandbox-affected miner runs
Today’s team-attributed brief says SN114 SOMA restarted 330 miner runs affected by sandbox errors, with a competition review planned for October 6. Restarting the affected work is a recovery step, not confirmation that every run has finished, rewards have settled or the broader competition is clear. Member allegations about cheating were not confirmed by the team and are not treated as established misconduct.
Audit trail
Coverage notes
Searched
- All 128 active subnet identities and 90 October 4–5 dated channel briefs screened, with selected brief texts read in full.
- Signed-in @ShizzyUnchained X discovery covered Bittensor/TAO, SN80/SN38/SN108/SN121, SN113/LongShort and official-account follow-ups. 195 associated public-account candidates yielded 185 usable samples and 125 current unique post records. These include commentators and quoted accounts, not just official teams.
- 180 repository paths scanned; 20 selected commit records inspected. Primary release, PR state, SN78 API, SN116 portfolio, SN108 leaderboard and service health checked separately.
- Signed-in AlphaGap feed and leaderboard, SubnetRadar news and dated Moments, TAO Daily, TAO Media and partner sources checked. Current market outliers and the top 12 emission recipients were reviewed separately from account counts.
- September 28–October 4 editions reviewed for duplicates, late findings, unresolved incidents and scheduled events. Conference recaps, documentation changes and merged code were not treated as completed production launches.
- Signed-in YouTube Videos and Live tabs checked. Public wallet guide PQ7gRxDG1Sg and October 4 members video M2nWtKAkw6I verified; missing September 29 members video FqFmRffEgVU recovered. Current SN49 replay remains listed, with no new scheduled livestream found.
- An overlapping pass of 18 priority public profiles ended at 2026-10-05T15:05:58.257Z; its newest SN118 business post repeats the September 30 conference announcement. Subsequent primary and signed-in follow-ups extend this edition’s research cutoff to 2026-10-05T15:10:26.683Z.
Public-data gaps
- Research cutoff 2026-10-05T15:10:26.683Z. Full-roster collection started around 14:55 UTC; subsequent targeted checks and overlap continued to the cutoff. Associated public-account samples and focused signed-in searches are not an exhaustive search of every SN/name. No Grok scan is claimed.
- 13 repository paths returned errors and 10 public-account candidates returned no usable sample. Directory metadata can be stale; SN26 repository attribution remains uncertain. Guessed chrono_llm, umi_vision and sundae_bar search aliases are not claimed as official identities; confirmed chronollm, umi_sn78 and sundaebar_ai sources were separately identified.
- Generated channel briefs and AlphaGap summaries are discovery evidence. Team-attributed scoring changes and incident responses remain developing without original-speaker or live-rollout confirmation; member-only accusations were excluded.
- SN116 Carbon’s live announcement and proposed eight-challenge portfolio are directly available, but rewarded mining, customer deployments and commercial savings were not independently demonstrated. Its apparent returns cross the slot’s pre-registration history and are not a Carbon performance record.
- SN41 Almanac and SN74 Gittensor measurements were not independently reproduced. SN80 OpenRoboto’s published allocation is verified documentation; its applied on-chain split was not reconstructed. SN38 ChronoLLM’s Round 14 PR remains open.
- SN78 UMI is accepting intake but the checked API still marks rewards inactive and chain submission unauthorized. SN120 Affine accounting changes and operator recovery records are not independently reconstructed wallet payments or proof of improved learning.
- SN118 Ditto reward recovery, SN11 TrajectoryRL missing-reward reports, SN44 Score warmup failures, SN15 ORO’s separate Race 11 pause and SN54 Yanez UAV payments are not independently closed. SN61 RedTeam replacement and SN114 SOMA review remain prospective.
- SN108 ChipForge’s leaderboard has no active challenge and reports all validator stake burning at observation. SN22 Desearch’s October 5 task/burn target and SN121 sundae_bar’s October 5–6 challenge target are not confirmed complete. SN111 Claims burn proposal, SN65 diligence, SN104 launch, SN94 permissionless access, SN122 dashboard and SN25 production qualification remain gated.
- SN106 Nodexo sale discussions do not establish a completed ownership transfer. SN113 LongShort’s outlier move lacks a confirmed new team or matched transfer. SN21 AdTAO’s latest lock amount was not independently matched; SN49 Nepher Robotics’s sourced lock is not evidence of a new purchase.
- Members-only uploads require YouTube membership access and were not bypassed. Public guide metadata rounds duration to 625 seconds, while the visible player shows 10:24; the site uses the player duration. No new livestream schedule was found.