BITTENSOR · THE NEWS ARCHIVE
Search 88 dated editions without scrolling through them one by one.
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 A past verified announcement can now be outdated.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.
Today’s SN1 Apex channel brief says the team has identified a 20-minute gap before competition rounds open that is causing scoring lag, and intends to fix it. This affects when participants can obtain a valid ranking opportunity. The brief does not confirm that the repair has shipped, so the issue remains developing.
A current competition-timing issue matters to miners, with resolution still unconfirmed.
SN15 ORO has merged partial scoring for generated runs that exhaust a miner’s inference key and for partial harness failures. Its OpenRouter proxy also explicitly requests usage-cost information. These updates make incomplete evaluations and inference spending more visible, following yesterday’s qualification-gate news. The repository verifies the changes; this is not a claim that every previously rejected agent will now receive rewards.
Evaluation accounting and spend visibility affect competition fairness and operator costs.
SN21 AdTAO says its next phase will bring miner predictions into customers’ Google Ads decisions and charge for their use. It plans to connect measured customer usage with attributable revenue and the alpha purchases funded by that revenue, with 28-day evaluation results still to come. The team says this commercial structure is being prepared during September. This is a stated rollout plan, not evidence that a new buyback has occurred or that the projected revenue has been earned.
A specific route from subnet output to customer use and buybacks is commercially relevant, but remains under preparation.
SN28 SayGM’s miner and CLI now support native DeepInfra image requests and image-priced offers, with the miner version bumped to 0.4.19. This extends the product’s routing and offer format beyond the preceding registration compatibility fixes. The implementation and version change are public; a new customer adoption or revenue figure has not been disclosed in these changes.
New image-routing and pricing support expands the work the inference marketplace can handle.
SN34 BitMind version 5.0.7 makes uniform random selection the default for generator challenges, assigning image or video tests across registered generators. Qualification, onboarding and responsiveness still determine eligibility for payment, but no longer change who gets asked under the default mode. The previous bucket-based selection remains an option. This is a documented change to how the competition samples work, rather than a removal of payment-quality checks.
Separating sampling from payment qualification changes access to evaluation and helps readers understand what the new default actually does.
The SN37 Aurelius channel brief says its sparse-autoencoder incentive mechanism has been delayed by testnet bugs, with fixes targeted for the weekend or following week. The same digest describes Proba as working with early design partners. The timing is a team-attributed estimate, and no new signed-customer count or production launch is established.
A rollout delay and present product stage help set expectations without turning development work into a launch claim.
Today’s SN46 Instant channel brief says the first-stage validator release for accelerated GLM-5.3 Flash inference is available, with miner code planned for the following day. The brief also describes forthcoming hyperparameter changes and miner collateral. This is a staged rollout: validator availability does not mean the complete new miner competition is already open. The timing and activation claims are attributed to the channel digest.
A staged inference rollout is substantive, while the miner release and collateral details still need direct release confirmation.
SN56 Gradients’ updated tournament documentation lists participation fees of 0.6 TAO for Environment, 0.7 TAO for Text and 0.4 TAO for Image, double the previous documented amounts. The code also makes fee messages reflect the configured charge. Separately, today’s channel brief attributes a September 21 text-tournament change to the team: three tasks per first-round group, no 32B model in the pre-boss round and a 30B–72B range for the large boss model. The fee documentation is verified; the upcoming format details remain digest-attributed.
Participation costs and the evaluation format directly affect miners’ operating decisions.
The SN62 Ridges channel brief reports 1,436 agent submissions from 460 hotkeys in Competition 28, with eight agents approved for emissions. It attributes a 36% leading result to the team, up from 32%, and says submissions are accepted through either the CLI or website. These are digest-reported competition figures, not independently reproduced benchmark results.
A quantified participation and leaderboard update is useful context, with modest weight because the original benchmark was not independently checked.
NATIX, the team behind SN72 StreetVision, has announced its third open-source driving dataset on Hugging Face. The addition pairs real multi-camera driving footage with vehicle telemetry, providing information about the vehicle’s actions alongside what its cameras saw. This is a parent-project data release relevant to its driving-data work. The announcement does not establish that all of the dataset was generated by SN72 or that the release changes subnet rewards.
A new reusable dataset adds tangible driving-research output while preserving the distinction between the parent project and subnet production.
Today’s SN78 UMI channel brief attributes an evaluator-bug fix and resumed weight commits to the team. New code also adds an independent public-intake audit monitor and strengthens readiness checks for staged competition acceptance. The recovery claim remains digest-attributed, and resumed weight commits should not be confused with proof that the new competition’s reward stage is fully active. An unnamed partnership teaser is not treated as a completed deal.
Restoring an interrupted reward-related process is meaningful, while activation and unnamed commercial claims need separate evidence.
SN79 MVTRX’s September 18 configuration moves its de-beta component to half of the trading score, with Kappa at 39.5% and profit-and-loss at 10.5%. Yesterday’s release assigned de-beta 25%, so this is a new step in the rollout. The release also adds counterparty information to the recent-trades view and hardens validator engine restarts. The configured weights are verified in the repository; network-wide validator adoption was not measured.
Doubling a live-release scoring component changes the incentives faced by competing trading agents.
SN89 InfiniteQuant has committed a future-effective expansion adding 21 crypto altcoin pairs to its high- and low-frequency evaluation boards from September 21 at 00:00 UTC. The change also adjusts how correlated crypto markets count toward its diversity floor. The implementation explicitly leaves current grading unchanged until the boundary. Three other candidate assets are held back because the available historical data falls below the required coverage.
A dated market-universe expansion and diversity-rule change are actionable competition updates, with activation still ahead.
SN90 KubeTEE’s validator configuration now requires 72 CPU cores and 1,024 GiB of memory, with at least five active nodes, replacing the previous 8-core, 16-GiB and seven-node thresholds. This is a substantial change in the resources expected from participating infrastructure. The same update documents a repair to the proxy permission used for recycling alpha earned through SN28 SayGM. The requirements are verified in code, not a measurement of how many operators already satisfy them.
A large increase in hardware requirements materially changes the participation barrier.
SN92 MicroTensor has changed extraction scoring to compare leaf fields, so nested JSON can be assessed against the flattened form used by hidden tests. Separately, observed network allocation is zero in the 10:42 AM ET snapshot, versus 0.0106% yesterday. Alpha is down 18.1% against TAO over 24 hours. Those observations do not establish that the scoring fix caused either move or that miners’ historical rewards were reversed.
A scoring repair and a new zero-allocation observation affect participants; the market decline adds context without proving causation.
SN100 Cortex’s official account says its website has returned online. This is an availability update after the interruption, rather than a new subnet or product launch.
A small but useful service-availability update.
SN102 ConnitoAI has merged a guard that discards evaluations completing after their round has already finalized. A separate update recalculates the FP8 layout when the task changes. These repairs address consistency around round boundaries and task switches, following yesterday’s correction to full-network evaluation. They are verified implementation changes; this scan did not reproduce the bugs on a live validator.
Protecting finalized results is material to stable competition accounting and miner rewards.
SN113 TensorUSD says Pallet is building its long/short mechanism and plans to let users allocate TAO to miner strategies and copy them, paying performance fees only on realized profits. The team describes miners as competing on delta-neutral strategies across subnet alpha markets. This provides more detail on the planned user and fee model; the posts use future tense and do not establish that strategy copying is publicly live.
A named implementation partner and user-fee model clarify a planned commercial product.
SN118 Ditto has merged a repair that starts a winning submission’s source-disclosure clock only after confirmed earnings in a completed reward period. Positive weights alone are insufficient. The change addresses the channel-reported problem of code becoming public before the contributor received emissions. Follow-up work reconciles payment receipts before archive scanning and displays the exact pending disclosure date. The implemented privacy window is 48 hours after qualifying earnings; adoption by every live service was not independently tested.
Disclosure timing directly affects miners’ ability to earn from winning work before it becomes public.
SN118 Ditto says its growing synthetic agent-trace dataset has reached 186 GB and is stored on SN75 Hippius for future Ditto model training. It also confirms Hippius storage for user files, with an option to connect a user-owned bucket. The project explicitly distinguishes the synthetic training dataset from customer files and says it does not train on user data. These are the team’s infrastructure and data-use statements; the dataset size and privacy implementation were not independently audited.
A quantified training-data update and a clear account of subnet storage use add product substance without implying a new model release.
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.