Dates and verification labels belong to the original edition. A past verified announcement can now be outdated. Historical project names are retained; an SN filter searches the slot across registrations. All keywords must match the story text or subnet label. Ratings describe editorial priority, not investment quality.
SN28 SayGM releases a desktop agent for everyday work
SN28 SayGM has introduced a desktop agent that uses frontier and open models through its gateway. The product page offers Apple Silicon Mac, Ubuntu x64/ARM64 and Windows x64 downloads; Windows is labelled an unsigned beta. Built on the open-source Goose project, it adds SN28 SayGM sign-in, billing and guardrails, with local-file work, model switching, MCP connections and scheduled tasks. The launch demonstration turns twelve contracts into a review table. This brings the subnet’s inference gateway into a user-facing workflow rather than requiring every customer to build an API integration. The service advertises usage-based payment without an app subscription. Downloads and documentation were checked; no software was installed and privacy claims were not independently audited.
SN28 SayGM has a newly available desktop product with identifiable downloads and a concrete demonstration, extending its inference distribution.
SN30 Endure opens a dashboard for checking mainnet activity
SN30 Endure has announced a mainnet dashboard exposing network state, miner commitments and consensus accuracy with underlying sources users can inspect. It adds a practical verification surface to the mainnet launch covered yesterday: risk intelligence is more useful when participants can examine how predictions and consensus are formed. The dashboard page was reachable at this check; this scan did not independently replay the full underlying dataset or demonstrate a miner payout. Miners’ reported archive-RPC requirements and the separate Forge product roadmap remain distinct from dashboard availability.
SN30 Endure adds a concrete public inspection tool following mainnet launch, without using dashboard access as proof of settled rewards.
SN33 ReadyAI previews measured slide-deck skills; store opens later this month
SN33 ReadyAI reports that adding one of its skills to the same AI model and brief raised an LLM-judge score from 24 to 94 out of 100 across 150 slide decks. It also reports hard-to-read slides falling from 86% to 2% and made-up numbers from 13% to 3%. These are the team’s scoped evaluation results, not an independent human review or a guarantee that generated decks contain no errors. Skill Store 33 is scheduled to open later in October, so the result is a product preview rather than a store already launched. The separate ARR and September buyback figures were covered on October 6.
SN33 ReadyAI offers a quantified product demonstration with a named evaluation method and an explicit future launch boundary.
SN41 Almanac reports $69,868 in September fees and a larger buyback budget
SN41 Almanac reports September platform fees of $69,868, more than triple its previous monthly record, alongside $6.83 million in trading volume and 17,889 trades. The team says all those fees funded purchases of TAO and SN41 Almanac alpha, with the acquired tokens permanently removed from circulation. It describes using accumulated revenue during a TAO pullback rather than sticking to its usual single monthly purchase. Early October fees had reached a reported $20,400 in the first week. The significance is the link between paid platform activity and token purchases; these are project-reported figures, not independently audited revenue or matched burn receipts. A projection that October will beat September remains a forecast, and trading volume is not revenue.
SN41 Almanac provides a fresh quantified business report and explicit fee-funded token mechanism, with reported results distinguished from forecasts and independent audit.
SN44 Score demonstrates an early world-model baseline from one store image
On October 8, SN44 Score shared a world-model demonstration from Max Sebti: one AI-generated store image and a camera path produce 49 frames at 512×512, alongside an estimate of the visible 3D surfaces. The official team shared the clip as a baseline for miners to improve. It gives the move from computer vision toward spatial intelligence a concrete output: generating how a scene might look as a camera moves through it, rather than only detecting objects already in a frame. The original post explicitly acknowledges visible artifacts, predicted unseen views and reliable physical simulation still ahead. This is an early demonstrated pipeline, not a launched simulator, independently reproduced benchmark or new customer deployment.
SN44 Score provides an inspectable early world-model demonstration with official team confirmation; commercial deployment and reliable simulation remain future goals.
SN44 Score’s October 6 football task is live with hourly evaluations
Late-found October 6 announcement: SN44 Score opened Detect Football Event, asking miners to identify passes, shots, saves and other match events in low-resolution youth-football footage. A single wide camera and a ball only a few pixels across make this a temporal understanding problem, beyond recognizing an object in a still image. The team announced a 53.1% best score against an 85% target and hourly evaluations. At the October 8 follow-up check, 16:27 UTC, the public console listed the task as live with a 53.9% private score against the same 85% target. Those are task-specific scores at different observation times, not a general accuracy claim or proof that the target has been reached. This corrects an omission without presenting the October 6 launch as new today.
SN44 Score has a distinct live mining task supported by an original team announcement and current public console; the original event date and unmet target are explicit.
SN51 lium.io: OpenTensor reports nearly $2.7 million in revenue-funded alpha buybacks
The OpenTensor Foundation’s October 8 update puts numbers behind SN51 lium.io’s GPU business: five buybacks between April and September reportedly removed roughly 171,000 alpha, worth nearly $2.7 million. September’s $946,000 purchase was the largest. The Foundation also cites $964,000 in rental billings and about 613,000 GPU-hours for August 10–September 9, with billed accounts rising from 445 to 1,187. The business connection matters: customer GPU-credit revenue is used to purchase SN51 lium.io alpha with TAO, then burn that alpha. This is a new primary ecosystem report about historical activity, not a new $2.7 million purchase today. Transaction-by-transaction burns and funding were not independently reconstructed, and these figures do not settle yesterday’s separate report of an October 5,000 TAO purchase. TAO itself is not the token described as burned. Separately, on October 7, SN51 lium.io said utilization was at an all-time high and most GPU types were fully rented, calling for more GPU providers. That is a current team-reported capacity signal, not an independently audited utilization percentage or guaranteed provider income.
SN51 lium.io has substantial, quantified customer usage and a described revenue-to-alpha mechanism in a fresh OpenTensor Foundation report. Historical figures and independent transaction verification remain distinct.
SN54 Yanez adds voice cloning in Phase 5; images still drive live rewards
SN54 Yanez’s merged Phase 5 Cycle 1 update adds a voice-cloning task: miners receive a reference voice and target speech, then return a cloned WAV. Voice is post-graded for the UAV path; image generation remains the live KAV-scored reward task. The update pauses real physical screen-replay collection and uses six synthetic face-variation tasks, while the remote reward API ignores screen-replay scores. Its sandbox is extended by a week, so the earlier October 8 target should not be treated as confirmation that every planned reward component resumed today. The practical change is a broader identity-testing dataset and miner interface, with reward status kept distinct from task availability.
SN54 Yanez has a substantive merged task expansion and schedule revision, while voice collection and live reward activation are explicitly separated.
SN58 Attune invites robotics miners; Horizon remains upcoming
SN58 Attune’s builder is inviting miners to challenge its robotics models, while the project site identifies Vector as live and Horizon as coming soon. Vector tests whether a robot policy can perform a demonstrated task in a different scene; its documented competition uses 160 evaluation units and requires a three-point success-rate improvement to take the crown. Horizon is intended to extend this toward learning from human video and world-action models. The October 8 invitation’s parent X post was deleted, so no launch announcement is inferred from that missing context. At the 15:37 UTC snapshot, alpha/TAO was up 0.90% over 24 hours and 75.76% over seven days. That prompted this review but does not prove a catalyst or reward payout.
SN58 Attune has a fresh builder invitation and primary documentation clarifying what is live versus planned. The market move is a discovery trigger, not the reason for its stars.
SN62 Ridges’ test-generation competition lifts its top approved score to 42%
SN62 Ridges reports that the leading approved score in its Test Generation competition rose from 16% to 42% over eight days. It records 946 agent submissions and eleven agents approved for emissions. The task rewards tests that detect bugs and remain useful when code changes, making this a concrete measure of the competition’s progress. The numbers describe this evaluation, not a universal coding success rate; 946 submissions should not be read as 946 distinct paying customers or miners. Approval for emissions also does not independently establish that every agent has received a settled payout.
SN62 Ridges supplies a large scoped improvement and concrete participation figures for an active useful-work competition.
SN64 Chutes confirms SN114 SOMA inference integration; AMD hardware remains gated
SN64 Chutes’ October 7 engineering recap confirms that SN114 SOMA is live on its inference service. SN114 SOMA compresses an agent’s context before inference, reducing the tokens it sends; the cost benefit is a team-described product mechanism, not an independently measured savings guarantee. The recap also distinguishes AMD SEV-SNP verification support from actual AMD hardware admission: the team says no AMD hosts are on the network yet and admission waits for attestable resolution of AMD-SB-3030. Private-chute hot-swap upgrades remain in progress. On research, the successor Lambda run is reported ahead of Kappa at the same training point, without an independently repeated comparison. Kappa results and the phone demonstration were already covered; the recap does not turn either into a newly downloadable consumer app.
SN64 Chutes supplies a primary integration and engineering-status update with clear boundaries between available inference, verification support and future hardware admission.
SN65 True Performance Network reports more than 150 Bench accounts
SN65 True Performance Network says more than 150 people have created Bench accounts during two weeks of outreach to developers and labs. Its service accepts a Hugging Face model repository and runs selected reproducible benchmarks without requiring users to set up their own evaluation environment. This is an early adoption signal for a benchmark service, following yesterday’s integration of SN51 lium.io compute. The count is registered accounts, not disclosed paying customers, completed benchmark jobs or revenue. No new competition winner is implied.
SN65 True Performance Network reports a specific early adoption measurement for its benchmark product, with accounts distinguished from paid usage.
SN66 Conjectures lowers new bounty starting shares to 5%
SN66 Conjectures has merged a change to its dynamic proof-bounty formula. A new target starts at 5% of uncommitted alpha treasury funds, rather than 10%, then ramps linearly to 12.5% after fifteen days. Previously accepted submissions retain their locked recorded amounts. This changes how a finite treasury is spread across mathematical targets without repricing already accepted work. Automated reconciliation and human multisig transfer signing remain separate parts of the payout process. Yesterday’s claimed Erdős Problem 579 resolution still requires independent proof checking and catalogue recognition; the bounty-policy change is not that confirmation.
SN66 Conjectures implements a concrete treasury-allocation change while protecting locked submissions and retaining proof-verification limits.
SN71 Leadpoet announces strategic investment from SubConnect
SN71 Leadpoet announced a strategic investment from SubConnect on October 8, saying the backing will help expand distribution and bring its sales-intelligence product to more teams. It is also recruiting a Founding GTM Engineer. This adds a financing and distribution development to yesterday’s separately reported doubling of September sales. The company did not disclose the investment amount, valuation or legal terms. Its founder also publicly acknowledged working with SubConnect’s team, while the counterparty’s sampled public feed had not yet supplied a separate investment announcement. The verified development is SN71 Leadpoet’s announcement; it is not a disclosed funding total or proof that future sales targets have been achieved.
SN71 Leadpoet names an investor and a concrete distribution focus, beyond a general growth teaser. Undisclosed terms and the counterparty confirmation gap remain visible.
SN71 Leadpoet changes Arena rewards to cumulative slots with weekly decay
SN71 Leadpoet’s merged Arena policy allocates three cumulative reward slots within a 30% miner incentive pot. A promoted improvement of at least ten points takes all three slots; five points takes two and one point takes one, initially corresponding to 30%, 15% and 6% of total emissions. Each slot’s reward then halves every completed 140-epoch reward week, reaching a permanent one-sixteenth floor after four weeks; removed allocation goes to burn. A new qualifying win resets only the slots it replaces. This rewards continued improvement rather than an unchanged champion indefinitely. Validators must support the signed policy before activation, and historical accepted reward states retain their original rules. Merged policy is not proof of new payouts already settled.
SN71 Leadpoet makes an implemented incentive change with clear thresholds, decay and activation boundaries.
SN74 Gittensor adds a GPU-rental CLI with a testnet default
SN74 Gittensor has merged customer commands for listing GPU offers, creating rentals, connecting by SSH, extending or stopping a rental and checking the balance. The default service endpoint is explicitly its testnet product, with other endpoints configurable. This makes its rental workflow more concrete, but should not be described as a demonstrated public mainnet commercial launch. A separate merged change introduces a 48-hour settlement holdback so serious host failures can affect accrued, unpaid miner compensation. No rental was ordered and no customer payment or production payout was independently verified.
SN74 Gittensor implements a customer-facing rental interface and a meaningful settlement control, with its testnet default made explicit.
SN75 Hippius launches team accounts with individual logins and one bill
SN75 Hippius has made Shared Accounts available in its Console. Teams can give each collaborator their own login and product-specific Viewer, Operator or Admin access, while resources remain owned and paid for by the account owner. The documentation covers virtual machines, databases, Kubernetes, runners and S3, with team-management permissions and an activity audit trail. That removes a practical barrier to using decentralized cloud infrastructure as a business: sharing access no longer requires sharing the owner’s password. Billing stays centralized and certain sensitive actions remain owner-only. Shared Drives are a separate file-collaboration feature. This verifies the feature announcement and documented access model; no customer account was modified and no independent security audit is implied.
SN75 Hippius ships a concrete enterprise-oriented cloud feature, supported by its own announcement and detailed operating documentation.
SN75 Hippius publishes a reproducible confidential-VM miner agent
SN75 Hippius has published its prebuilt confidential-VM miner agent, with the latest v2026.10.08.1 release accompanied by binary hashes, build provenance and signed software-component inventories. The release says it was built twice with identical bytes. Its confidential-compute path targets supported AMD EPYC SEV-SNP hardware; this is not an ordinary desktop-app update. Miners who have automatic updates enabled install the release after it is 24 hours old, rather than every miner switching immediately. These are useful operator and verification tools, while a signed build does not itself prove that every live host has upgraded or that every customer workload runs inside the claimed protection.
SN75 Hippius ships an inspectable operator release with concrete build-verification artifacts and an explicit update policy.
SN80 OpenRoboto bridge operator reports first-72-hour activity
ForeverMoney reports $393,000 in trading volume and 237 new token holders for SN80 OpenRoboto after bridging to Base and Robinhood Chain. It also cites more than 250,000 social impressions during the first 72 hours. This is directly relevant to the subnet because it measures a new distribution route for its token beyond Finney. The figures come from the bridge operator, not an independently reconciled holder census or customer-revenue audit. Trading volume and token holders should not be confused with robotics customers or subnet revenue, and the operator’s advertised price and comparative-profit figures do not establish returns for every holder.
SN80 OpenRoboto has a fresh, directly relevant bridge-usage report. Operator attribution and the limits of market-activity metrics remain explicit.
SN91 Cascade reports a stronger time-series result using fewer training tokens
SN91 Cascade says its forecasting model beats Toto 2.0 4M on the TIME MASE and TIME CRPS metrics after 58 days, using only miner-generated synthetic data and 15.4% as many training tokens as the comparison. Those metrics assess forecast error and probabilistic forecast quality. The result matters because it tests whether the subnet’s synthetic-data competition can improve a useful forecasting model efficiently. It remains a team-reported comparison, not an independently repeated benchmark or proof of superiority across every forecasting task. Yesterday’s Shizzy interview is now available as a 54:15 public replay and has been moved out of upcoming-stream status on the site.
SN91 Cascade reports a specific, relevant model result with a named baseline, metrics and training scope. Independent reproduction remains separate.
Searches published headlines and summaries, not private bookmarks, unpublished leads or source-page contents. Duplicate headlines within an edition appear once; separate editions remain separate so follow-ups are preserved. Share the page URL to retain your filters.
