October 17, 2026 · Published daily
Subnet News — October 17, 2026
SN1 Apex puts an agent-built competition into play, SN2 DSperse announces a DaniLabs handover agreement, and SN28 SayGM publishes its second burn cycle. SN97 Ingot turns to model quantization, SN82 Noah releases participation code, and operating incidents remain on watch. Also: eTAO code versus mainnet activation, verified identity changes, a scheduled coldkey swap and the latest Shizzy members video.
- Coverage window
- Oct 16, 2026, 9:00 AM – Oct 17, 2026, 7:50 PM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 32
Biggest stories first
Today's ranked news
eTAO code refines root protection; finalized Finney still reports spec 479
Primary subtensor code adds the spec-481 root-off-the-top refinement and an integer-parts validator-floor comparison so the advertised floor and chain check agree. At the recorded finalized head, Finney still reported runtime spec 479. These are verified code changes awaiting confirmed mainnet activation, not a reason to recalculate today’s live yield using future reward splits. Registration immunity and lock-reduction parameters were unchanged between the two recorded heads.
SN1 Apex launches an agent-built incentive competition
On October 16, SN1 Apex announced an incentivized competition built by agents from a prompt by Const. Its follow-up links a live NIST-Lite robotic assembly contest. The team says development took six hours without human coding; that is an attributed development claim, not a benchmark we reproduced. Const’s quote-post invites agents to mine it. The important change is a concrete competition to participate in, rather than another promise about agent-built economies.
SN2 DSperse announces a DaniLabs handover agreement
Inference Labs announced October 16 that it has entered an agreement for Dan to take over SN2 as DaniLabs, focusing on fast, affordable inference. This is an announced agreement, not a chain-confirmed completed sale. Our finalized comparison from block 9,296,035 to 9,306,529 found no change to SN2’s owner or registration height, and its stored name remained DSperse. No sale price or executed buyer identity is inferred. Unverified accusations in channel discussion are excluded.
SN7 Allways incident remains open while recovery code lands
The October 17 channel digest reports an October 14–16 miner-side failed-swap exploit, a halt for patching and planned reimbursements. These incident and reimbursement details remain secondary reporting; we did not inspect private original messages or confirm every refund. Primary code separately bounds event ingestion to prevent backlogs blocking votes and enriches swap/configuration event logs. Those repairs support an operating recovery story, but do not establish that service has resumed or that the exploit is fully resolved.
SN28 SayGM reports its second burn cycle, with a finalized burn receipt
SN28 SayGM reports $104,372 burned against $514,686 of miner emissions for September 16–October 15, or 20.3% coverage. We independently decoded a successful burn_alpha at block 9,296,823, extrinsic 11: AlphaBurned records 10,771.229342391 SN28 alpha. The burn is verified; the dollar valuation and cycle-wide offset calculation remain the project’s accounting. Its page says some burned alpha comes from customer payments, so the dollar total should not automatically be described as cash spent buying tokens.
SN28 SayGM adds DeepSeek Flash with SN90 KubeTEE
SN28 SayGM announced DeepSeek-V4.1-Flash-1017 availability on October 17 and credits SN90 KubeTEE for day-zero onboarding. The primary miner repository allows the model and promotes gm-miner 0.5.2. The team advertises confidential delivery and pricing 10% below list. We verified the announcement and supporting code, but did not send a paid inference request, audit the security boundary or independently measure model performance.
SN60 Bitsec announces SN105 Beam as a continuous-coverage client
SN60 Bitsec announced October 16 that SN105 Beam is now covered by Sentios Continuous Coverage, pointing to an active-project listing and a GitHub badge. The original announcement is readable, but our attempted active-project page fetch failed and the public BeamCore README examined did not independently confirm the badge. This remains a team-announced client relationship pending stronger confirmation from the other side. No contract value, vulnerability finding or security guarantee is inferred.
SN82 Noah publishes its miner and validator participation system
SN82’s finalized identity now reads Noah and links the public Vora-Inc/NOAH repository, whose initial release documents an earnings-and-contribution scoring system. Its rules require an entry contribution of 10 spendable NOAH and contributions toward 50% of eligible spendable earnings to the Ark. The documentation explicitly says these transfers change ownership and are not withdrawable balances; a finalized receipt is separate from credited score. Publishing code does not by itself prove an operating validator. The alpha/TAO snapshot showed +111.59% over seven days at 23:49 UTC; that price move is context, not evidence the system works.
SN92 REAX explains decision-model mining and repairs competition queues
SN92 REAX’s October 16 launch thread describes miners hosting and improving decision models, validators checking correctness, and a customer endpoint at DecisionModels. The team also describes demand-forecasting and risk-assessment competitions. Primary releases rc8.post1 and post2 stop unfetchable entries consuming evaluation slots, share hourly evaluation budgets between tracks and repair queue liveness/first-draw carry-over. These are code-verified changes; promotional speedup claims are not independent benchmarks. The 23:49 UTC alpha/TAO snapshot showed +16.48% in 24 hours and +771.47% in seven days, which does not establish the cause or durability of the move.
SN97 Ingot replaces the Albedo identity with a quantization competition
Finalized identity storage now names SN97 Ingot and points to DendriteLabs/ingot. Its published rules ask miners to quantize a fixed Qwen3.8-27B teacher while beating the current king on both output-distribution divergence and hidden-state distance. Up to five recent kings each receive a 20% share for at most five days, subject to replacement; unused shares go to UID 0. New code rejects unloadable challengers and invalid tensor formats. This is a distinct direction from the previously covered Albedo distillation work; the rules and implementation are verified, not independently rerun evaluations.
SN118 Ditto posts conflicting burn and treasury plans
Two official October 16 posts describe different destinations for buybacks funded by half of profits: one says the acquired SN118 alpha will be burned; a later post says it will be held in the company treasury. These are materially different policies. We retain both direct sources and treat the destination as unresolved rather than reporting an executed burn or assuming a clarification that has not been verified. The dated channel digest also reports a validation pause under investigation; no confirmed all-clear was found in the reviewed sources.
Const explains the design stance behind Null Consensus
In October 16 replies, Const argues that allowing more consensus choices is a less prescriptive approach to subnet design and says builder demand motivates the change. He contrasts the current dTAO setting with the earlier root-managed environment. These are his design opinions, not an independent finding that every implementation is decentralized or safe. The optional Null Consensus release itself was covered yesterday; today adds the substantive follow-up commentary and the SN1 Apex endorsement in its own story.
Crucible announces Chrome passkeys and wider iOS availability
Crucible Labs announced Chrome passkey support on October 16 and says its iOS wallet is available in 30 countries as of October 17. These are original product announcements; we did not install a wallet, authenticate with a passkey or independently verify every storefront. The update is wallet access and distribution, not a staking-yield or validator-safety endorsement.
Const links future pool updates to lending
In an October 16 reply, Const says pool updates have been under development and will be tied to the introduction of lending. This is a direction-of-travel statement, not a completed lending launch, published rate, audited implementation or activation date. We include the original reply as planned protocol commentary and carry implementation and activation evidence into the next research pass.
SN4 Targon publishes its eTAO discussion with Const
SN4 Targon published Talking Targon Episode 01 with Const on October 17 UTC, covering emission controls, guardrails and compute economics. This is a newly released discussion, not proof eTAO is active on Finney. Our finalized runtime check at block 9,306,529 still returned spec 479; the newer eTAO changes are present in primary code. The revenue-dashboard milestone was covered yesterday and is not presented again as new revenue.
SN14 Cacheon evaluation pause remains a capacity issue to watch
The October 17 channel digest attributes the evaluation pause to GPU availability and cost, and reports a team expectation to resume the following morning. That is a forecast, not a verified restart. Public product-website work was also collected, but it does not prove the evaluation backlog has cleared. Private original messages and live evaluation telemetry were not independently inspected.
SN15 ORO bounds validator file reads and sandbox disk use
Primary SN15 ORO changes cap reads of sandbox outputs/logs and bound per-evaluation disk use, with evaluation failure when budgets are exceeded. The stated purpose is to stop malformed output or excessive files from exhausting validator resources. This is verified implementation work, not a claim that a specific production exploit occurred or that every operator has deployed it.
SN44 Score gives an early look at its new Studio
The October 16 SN44 Score post shows a Studio preview in which a user corrects a visual interpretation through conversation. The caption calls it a first look and says more is coming, so this is a demonstration rather than a confirmed public release. We inspected the playing clip. Separately, the October 17 channel digest reports Fallen Person Detection restored and PPE active again; those operating updates remain secondary and are not treated as independently verified uptime. Recirculated September break-even claims originated October 1 and are not a new October 17 revenue milestone.
SN49 Nepher Robotics reports a stronger third-phase run/jump result
SN49 Nepher Robotics announced completion of Humanoid Run Jump Phase 3 on October 16, reporting a winning score of 0.8245 versus 0.6750 and 0.7735 in the prior phases using the same Unitree G1 task. These are team-reported competition scores, not an independently replicated physical-robot result. The October 17 digest also clarifies submissions run October 5–18; participants should check the team’s current timeline before acting on a deadline.
SN51 Lium strengthens GPU-node qualification and CLI availability
Primary SN51 Lium code now fails unrented nodes when sshd does not start in the platform’s filler container, rather than treating installed software as evidence of a usable connection. Another fix limits a portal deletion to the named node. The CLI exposes free GPU counts on partly rented nodes and can upload Docker build context. These are specific operator/customer reliability improvements, not evidence for promotional payout comparisons or a new revenue total.
SN56 Gradients pauses environment tournaments while retaining their emission settings
The primary change disables new environment tournament starts and skips leftover pending/active environment rows until the full sunset. Its documentation explicitly leaves the environment emission share and related weight endpoints in place for the week. Text and image tournaments retain their separate schedules. A pause in scheduling should therefore not be described as an immediate removal of that emission allocation.
SN64 Chutes transcription issue is reported in the playground
The October 17 channel digest reports a team member observing gibberish or incomplete output from the audiodojo speech-to-transcription path in the playground. This is a specific reported model/path issue, not evidence that all SN64 inference is unavailable. We did not reproduce the request or verify a resolution; keep it on the operating watch list.
SN67 Harnyx adds separately citable visual and structured page retrieval
SN67 Harnyx announced October 16 that Expand Fetch is available to every miner as a page-fetching tool. Each request provides a page representation preserving tables/layout and separately citable hidden structured data as JSON. This matters for tracing research claims to what an agent actually retrieved. We verified the original announcement; improved answer quality is an intended benefit, not a benchmark we independently measured.
SN75 Hippius releases desktop capture improvements
Primary release 0.6.13 adds recording controls, a dedicated Captures drive, Linux capture fixes and a screenshot editor. Its release notes distinguish macOS release-pipeline recording support from local builds that omit the helper. These are concrete desktop product changes; we did not install the app or test every platform, and a release entry is not proof of customer adoption.
SN78 UMI backports durable coverage-gap recovery
SN78 UMI’s primary code backports historical coverage recovery to its exact C5 runtime. The changes retain separate recent and historical replay lanes and require native proof/ancestry checks; failed proof or credit handling does not advance the historical cursor. This is a repair to evidence continuity, not confirmation that all outstanding rewards are payable or the previously tracked rollout is complete. We retain the unresolved operating status until activation and payout evidence is available.
SN85 Vidaio reviews licence eligibility in competition 003
According to the October 17 channel digest, nine entries were rejected over unlicensed pretrained checkpoints and three later reinstated when review found those checkpoints were not loaded at runtime. This remains attributed secondary reporting: we did not inspect every submission or a private moderation record. The useful distinction is between packaging an unlicensed artifact and actually using it; it is not a blanket claim that all disqualified models were restored.
SN89 IQ prepares gold calendar markets for October 18
Primary SN89 IQ code schedules daily and weekly gold windows from October 18 at 17:00 EDT, using the documented gold oracle and session rules. The code states behavior is unchanged before that activation time. A separate README update sets the IQ Markets withdrawal minimum at $50. These are documented rule changes and a future activation, not a claim that the new gold windows are already live during this report.
SN102 ConnitoAI prevents a chain-read failure from killing the validator
Primary release v0.7.1 changes round-start handling so a chain-read failure no longer terminates the validator. This is an availability repair distinct from the finalized coldkey swap covered yesterday. The repository does not by itself prove adoption by every validator, and the future incentive-format plans are not presented as already activated.
SN104 TAOstatus reports external security-miner access, with evaluation still incomplete
The October 17 channel digest says the security track is opening to external miners using per-miner Dockerized agents instead of a shared agent. It also acknowledges validator evaluation is not yet fully operational. These are attributed progress and limitation statements, not proof of a mature competitive marketplace or verified earnings. The evaluation gap stays open for the next run.
SN108 ChipForge fixes repeated-hotkey reward aggregation
The primary SN108 ladder now follows the server’s repeat_hotkeys option, which defaults off when missing. When repetition is allowed, multiple places held by one hotkey are combined into one UID weight entry; bans are applied per place so one quest’s ban does not automatically erase payment for another valid quest. The change addresses reward accounting, not a guaranteed increase in any miner’s return.
SN118 Ditto previews a Claude plugin and demonstrates partner screening
SN118 Ditto labels its October 16 Claude plugin demonstration an early look at an upcoming product: the example retrieves 20 memories from different sources. It is a preview, not a confirmed public launch. An October 17 post says a partner fit-check against 95 features shipped that week, illustrated with a made-up landscaping company. The example is a demonstration, not evidence that the pictured business is a paying customer. No new capability is inferred from the separate sparse-caption meme video that could not be fully reviewed.
SN120 Affine weight-update incident is not yet a confirmed all-clear
The October 17 channel digest reports a weight-update error, a team statement that a fix should restore updates, and subsequent miner reports of mismatched incentives. We did not independently reconcile each miner’s submission and payout history. Keep this as a developing incident: a claimed fix is useful progress, but unresolved discrepancies prevent describing normal reward operation as established.
SN121 SundaeBar documents skill bundles and its leadership margin
The primary README now defines submissions as skill bundles with a limited SKILL.md plus optional reference files, and documents a 0.01 leadership margin: a statistical tie leaves the incumbent in place. The October 17 channel digest says SBC30 closed October 16, with SBC31 planned for Monday and an interim UID 0 allocation. Those schedule/allocation statements remain secondary reporting, not independent confirmation of every weight update.
A finalized coldkey-swap announcement is scheduled, not executed
At block 9,303,541, extrinsic 10, the node records ColdkeySwapAnnounced for 5Gndo…aXTP and a new_coldkey_hash commitment. That hash is not a decoded destination address. The announcement proves scheduling, not execution, buyer identity, price or a subnet sale. Across all 128 slots, owner/registration/network-presence endpoints were unchanged between blocks 9,296,035 and 9,306,529. Full-range historical storage querying was blocked by RPC4003, so temporary intermediate events cannot be ruled out.
Latest Shizzy members video is refreshed
The October 16 members edition, “10/16/26 Bittensor News Updates & TAO Price Action | Members Video,” is now the latest members promotion. Both channel Videos and Live tabs were checked. The watch player confirms membership-gated access and 542 seconds of content; the channel card rounds the display to 9:03. The latest public upload remains the default public player, and a newly listed SN10 Pareton interview is scheduled for October 20 at 1:00 PM Eastern (17:00 UTC). The Live tab and player were rechecked at 23:59 UTC, after the news research cutoff.
SN10 Pareton interview is scheduled for October 20
The channel Live tab now lists “Can Bittensor Make AI Faster and Cheaper? Inside Pareton SN10” for October 20 at 1:00 PM Eastern / 10:00 AM Pacific (17:00 UTC). The YouTube player confirms upcoming status, not a live broadcast. This listing was verified at 23:59 UTC, after the main news cutoff. The site reminder and library now point to the scheduled interview while retaining the latest public replay as the default player.
SN16 Zero Hour is the finalized name; its robotics concept remains a demo
At finalized block 9,306,529, SN16’s stored name is Zero Hour, with a description about robot-task disruptions and recovery demonstrations for VLA training. The linked fast-thinker/kenju repository still calls the subnet proposed and its current implementation a local symbolic demo. That verifies the identity and concept, not a production mining launch or completed sale. The alpha/TAO snapshot showed +18.24% over 24 hours at 23:49 UTC; price is not proof that the concept is operating.
SN68 NOVA clarifies reproducibility requirements
The October 17 digest attributes a rule clarification to the team: pretrained coefficients without verifiable training methods/data are not allowed, while training a surrogate inside the sandbox budget is permitted. This is secondary reporting of a qualification rule, not a finding that any named miner cheated. Primary enforcement and updated formal rules remain follow-up items.
SN70 finalized identity changes to Consted 2.0
Finalized identity storage at block 9,306,529 reads Consted 2.0 for SN70. We record that owner-set label for history, without treating its accompanying commentary as an established factual claim or implying any affiliation with Const. Owner, registration-height and network-presence endpoints were unchanged from the preceding recorded head. A renamed label does not establish a sale, new registration or operating product.
SN107 Minos plans an IM2 presentation; deployment is still ahead
The October 17 digest says SN107 Minos is prioritizing IM2 and plans a presentation around October 22–23. The same discussion contains disputed allegations; we do not adjudicate them from a digest. No completed IM2 release or independently confirmed resolution of sybil concerns is established here. Treat the dated presentation and incentive redesign as planned follow-up.
Audit trail
Coverage notes
Searched
- 128 current subnet profiles; 98 October 16–17 dated channel briefs; primary repositories across 199 paths
- 209 current/previous team and partner X account samples, 248 current original post pages
- Signed-in Grok discovery and estimated top-five attention review; direct Const originals, quote-posts and substantive replies
- SubnetRadar, TAO Daily, tao.media, public project feed, Moments and primary product/release sources
- Finalized node owner/registration/network-presence comparisons; selected burn and coldkey-swap event decoding; current runtime and identities
- Full-roster alpha/TAO market and emission review; unresolved incidents and preceding seven editions
- Both YouTube Videos and Live listings, new members player metadata, cached validator/yield health checks
Public-data gaps
- Public X account sampling and Grok are discovery aids, not exhaustive searches or a reliable platform-wide mention census. Some accounts and repository paths were inaccessible; per-source outcomes are in the audit.
- Private Discord originals and AlphaGap paid/live intelligence were not inspected. Digest-only incidents and rules remain attributed and developing.
- The archive RPC rejected full-window state_queryStorage with code 4003. Endpoint comparisons and selected decoded events are recorded, but temporary intermediate registrations/removals/transfers cannot be excluded.
- Some media was only partially inspected or could not be fully reviewed. No unobserved feature, benchmark or launch claims are inferred. Revenue, usage and cycle valuation are not independently audited.
- The SN94 conviction/king alerts were not corroborated by the targeted decoded events at the advertised block; they remain unresolved in the audit rather than a reported completed takeover.
- Prepared validator recommendations and a ready background node yield check were accessible. Direct TaoStats account credits/usage and all historical indexer event types were not independently inspected; no new fallback integration is claimed.