Discord Alpha

Multiple miners in the Dogelayer (SN80) Discord accused the subnet operator of stealing mining rewards. One miner claims they earned 32 DOGE instead of 7000 DOGE based on their 1.2 TH/day hashrate, presenting calculations and screenshots as evidence. The operator has not publicly responded to the accusations in the channel.

Forge reported autoresearch produced no score improvements last night with 85 errors and 15 reverts. The team released updated code to GitHub. Multiple users are encountering agent submission issues including plagiarism errors and validation tool failures.

Team members discussed registration cost economics, noting that recycling emissions exceed mining rewards on the subnet. Conversation surfaced questions about a single validator operating 50+ hotkeys and concerns about hardcoded behavior in the system, though no resolution or formal action was documented.

The beam subnet team acknowledged a critical internal error in the gateway update API. Developers reported failed requests with 'internal error' responses, and the team confirmed they will address the issue during business hours. The team also emphasized the subnet's unique orchestration complexity as a competitive advantage.

Team discussed moving away from Claude as their current solution for an upcoming release cycle. Discussion centered on OAuth token configuration, environment variable precedence, and cost optimization for API usage. No timeline or specific alternative announced yet.

A team member reported missing data on the subnet's dashboard pages. The issue involves certain '24s' metrics not appearing on pages 1 and 2. The team confirmed the bug and will investigate.

The team discussed miner cost structures for baseline runs, proposed using a free Qwen model for evaluations, and requested fixes to the minimum incentive mechanism for screen-passed miners. Token limits for baseline submissions were also mentioned.

Miners report that newly deployed models receive zero composite scores despite high total_weighted scores because latency_pass defaults to false. This creates a barrier for new miners to compete with established ones within the 30-evaluation window. A team member asked about workarounds involving pre-warming the Chutes infrastructure before on-chain commit.

Cacheon is restarting its competition on June 1 after refactoring evaluation logic. Speed and correctness now run on identical prompts in a single pass; TTFT and TPS metrics replaced by one end-to-end wall-time number against vLLM baseline. Spec decoding that passes correctness checks is permitted. Emissions will ramp to full rate after June 8, starting at 2% next week.

The Ready Subnet team highlighted the low cost of mining entry and recommended using the testnet (uid 148) for setup practice. They noted that default miners aren't competitive, urging innovation in models and prompting strategies. Monitoring tools like Bactensor and the metagraph are available for miners to track scores.

The subnet's current score is described as unbeatable, prompting requests for new scenarios. A discussion about conviction mechanisms (lock or not) is ongoing.

The Horde team discussed increasing the minimum chunk take to deter weight copying, clarified validator requirements, and confirmed CPU-based validation remains viable. Validators without Trusted Miners can operate in 'auditor' mode.

Team members exchanged technical questions about the DQ (data quality or similar) system status and sample count metrics. The messages indicate routine operational monitoring but provide limited concrete updates on shipping or changes.

The apex subnet team reported ongoing competition bugs with no scheduled resolution. User inquiries about CLI usage and ownership status were raised but unanswered.

Cathedral v2.1.0 shifts weight scoring to recent verified solves, causing inactive positions to decay. Challenges are live but undergoing calibration; team expects stability by evening. Roadmap includes 50+ tiered challenges with submit timeouts and per-challenge winner-takes-all mechanics.

The Allways network resumed after a 45-minute pause for v1.0.8 deployment. Key changes include credibility penalties for repeated timeouts, free-rider hole closures, capacity adjustments, validator robustness fixes, and logging enhancements. Miners with 3+ timeouts in the credibility window will see rewards zeroed starting next round.

The Minos team responded to claims that some miners have identical coldkeys as validators, stating they have no incentive for self-mining and have not received proof. They emphasized dashboard transparency showing dynamic winner changes and directed users to request evidence from claimants.

Team members noted ongoing issues with miners failing to activate deployments. Alexdrocks advised verifying local chute functionality and using the warmup endpoint. Exchange rate concerns were mentioned as a potential obstacle.

Ridges announced a partnership with Latent. Team directed community to a dedicated Discord server for updates on project developments.

The τeuτonic team shared no concrete developments over the past 24 hours. Messages included vague references to untriggered processes and a cryptic 'conviction' statement with emojis. Activity appears minimal.

Quasar deployed a validator scoring update to stabilize crown-quality gate benchmarking with reduced variance. The team is reviewing whether the updated gate is now too strict and will solicit miner feedback before any adjustments. A larger data update is in preparation, and the team emphasized that miners should train against quasar's unique looped-transformer architecture rather than copying standard qwen approaches.

Discord messages contained meta-announcements and a feature suggestion without concrete implementation details. No releases, fixes, or metrics reported.

Team member cats_cr called out an unnamed party for removing owners from three subnets without communication. cats_cr referenced a previous removal of 'Daryxx' and alleged ongoing extraction activity. Subnets 36, 70, and 99 flagged as concerns, with pressure for faster resolution than prior cases.

Ninja (SN66) identified critical issues with its Openrouter LLM judge logic and is raising the disqualification threshold to 7 to reduce false DQs from task ambiguity. The team shipped a new dashboard benchmark comparing agents against mini-swe-agent on unscored reference tasks—meant for tracking progress, not affecting miner scoring. Team is evaluating reward mechanisms to incentivize direct agent improvements over hill climbing, with agents given several days to adapt to the new judge before scoring resumes.

The team activated a burn mechanism to mitigate exploit risks and address 70% incentive concentration among miners. They confirmed this aligns with the subnet's design thesis, where signal-rich prompt selection drives rewards, and training difficulty increases as unfilled windows rise to 20%.

The Soma team addressed questions about scoring methodology, miner incentives, and post-competition transparency. Final scores use all 50 SWE problems, not 45. Incentive distribution details were requested but not specified. A proposal to share top agent code after the competition was suggested for community learning.

The team highlighted inconsistencies in the subnet's scoring system, where models with low validation loss were excluded from scoring and promotion. An internal dashboard was referenced to cross-verify metrics, and concerns were raised about models being locked out of scoring after poor initial performance.

Evaluation functionality is currently not working in the subnet. The team confirmed that GPU out-of-memory (OOM) issues will trigger re-evaluation. Discussion included clarification of senior miner identities and repository management questions.

The team responded to repeated demands for crypto releases and addressed withdrawal delays exceeding T+2 days. Discussions included code changes to prevent virtual miner injection risks, but no concrete actions or timelines were confirmed.

Beam patched a deregistration bug causing hotkeys to lose partial or total history. The team is working on a persist timeout issue requiring control-plane redesign. Orchestrators are experiencing assignment failures and routing gaps; the team is investigating cooldown and reliability weight mechanisms affecting transfer routing.

GT confirmed ownership of the oxmarkets subnet, with Taoshi/Vanta team no longer involved. Conviction mechanism updates are prioritized alongside veAlpha logic changes. Exchange launch postponed to June 30 due to audit requirements and team restructuring.

The Red Team subnet (SN61) confirmed a partnership with Yuma and provided documentation for validation checks and challenge submissions. Team members referenced ongoing participation in a fingerprint challenge.

The team identified a reward distribution bug where miners with high task completions received 0 emission. A fix via the Omakase branch is under deployment, with validators instructed to pull the update. Weight decay in reward scheduling is confirmed as the root cause.

Team members reviewed quality gate metrics on Weights & Biases logs and discussed training interval optimization. A brief exchange highlighted confidence in project direction alongside technical questions about 15-minute training cycles and their impact on model quality.

Community members inquired about validator and emission functionality, with confirmation that validators are now working. A token holder expressed concerns about development alignment with team commitments and requested transparency. Separate reports of miner-side technical issues without disclosed resolution details.

The team addressed a 6-hour period where miners received no challenges for the Car-wash detection element. A fix was implemented and confirmed to resolve the issue.

The team identified a scoring inconsistency where miners can maintain scores during 'cold' chutes by temporarily disabling inference, then reactivating to preserve rankings. This undermines model fairness and requires evaluation rule adjustments.

The albedo subnet team addressed a GPU-related issue impacting validator operations. Initial reports flagged the problem, and subsequent updates confirmed resolution. Validator functionality is now restored.

Subnet 15 is experiencing persistent timeout issues in the Shop and Voucher modules during validator deployment, with inconsistent results across different validator instances despite model upgrades and workload optimization. Separately, overnight autoresearch completed 100 iterations without performance gains, scoring 0.8571 with 85 errors and 15 reverts; the team is analyzing error logs to identify common failure patterns before the next cycle.

The neverplayalone team is actively seeking a subnet slot on mainnet and exploring funding options. They are prepared to purchase a slot directly if registration costs decrease, indicating readiness to launch.

The team reported improved miner incentives with recent adjustments, noting many miners now receive proper rewards. No concrete technical updates or launches were detailed in the 24-hour window.

The nova subnet team addressed an intermittent issue with the UniProt API. After restarting the service, they confirmed the problem was temporary and fully resolved.

Team members raised two unresolved questions about validation fixes and new scenario additions, but no responses or updates were provided in the last 24 hours.

Team members acknowledged a query about potential announcements but provided no concrete updates or timelines. No new features, releases, or incidents were discussed.

The team clarified that exact duplicate submissions from the same hotkey trigger similarity checks, while minor modifications are permitted to encourage iterative improvements. A Challenge Card resource was referenced for detailed guidelines.

The team highlights flaws in fixed reward shares for repos, proposing dynamic allocation based on independent contributor activity and anti-gaming mechanisms. Per-repo emissions were recently enabled, with plans to transition control to alpha holders.

The ninja subnet team implemented technical changes introducing per-round I/O overhead and adjusted rate-limiting parameters, while actively resolving duel evaluation speed regressions. Miners reported prolonged queues and evaluation times, prompting concurrency optimizations that reduced duel durations to 10-15 minutes. A new challenger's syntax errors caused solver failures, and discussions highlighted miner strategies contributing to queue congestion.

The Beam subnet (SN105) initiated its first incentivized transfers, marking live emission for qualified orchestrators. The team acknowledged multiple Beam Core bugs, including stalled transfers, metagraph sync failures, and storage limit errors during registration. Guidance was provided on maintaining infrastructure stability during the qualifying phase.

The Albedo team reset its evaluation competition with fixes, but identified persistent prompt injection vulnerabilities in model responses. They confirmed old models are invalid and burned hotkeys remain inactive. Dashboard evaluation display issues and evaluation mechanism limitations were also discussed.

Team recommended miners migrate Docker images from Docker Hub to GHCR (ghcr.io) due to rate-limiting failures affecting evaluation reliability. Community members raised concerns about suspicious validator behavior, claiming the dev team is self-mining on the subnet and pointing to identical transaction patterns and cross-subnet mining activity as evidence. Miners expressed frustration over lack of transparency and indicated they are leaving the subnet.

The Soma team resolved a hard drive cleanup issue that paused evaluations, restarting validation processes. They confirmed miner runs completed and outlined baseline cost estimates for experiments.

Chutes validator experienced 11k-block downtime over 24 hours while infrastructure migrates to improved systems, with a hotkey swap also noted. Team clarified support scope: third-party wallet tools are out of scope; use native Bittensor SDK and btcli. Library v0.6.9+ required for log streaming; warmup issues typically stem from chute config.

Score subnet (SN44) is undergoing incremental improvements rather than fixes, with the team describing the network as thriving. Team is busy with upcoming initiatives including a proof-of-talk event and will be in Paris starting Sunday. Emissions for the cricket task will be increased.

Community members identified UID 242, a top-emitting miner, as operating a fake/cheating setup with artificially low hashrate. Multiple users accused the subnet owner of running the fake miner themselves and provided proof-of-concept code showing exploitable pool difficulty fallback logic (0.0001 difficulty). Calls for immediate removal and code redeployment.

Team members discussed potential changes to mining workflows and confirmed Champion model resampling on B200 hardware. A deployment failure was noted but deemed non-critical to duels. A system prompt validation issue in NAVWORLD was reported.

The Bitsec team resolved an INFERENCE_API_KEY validation error in the LLM screener and added a new project to improve feedback accuracy. They also discussed increasing the 3k miner limit after it was identified as restrictive.

The Eirel team clarified their incentive distribution model, stating that rewards are currently centralized under the 'general chat' family. They explained that future families (e.g., research, coding) will receive weighted distributions as they are added.

The subnet's autoresearch system completed 100 iterations but encountered 85 errors and 15 reverts, leaving the score unchanged at 0.8571. Yesterday's code update has been deployed to the live agent repository.

Veraτhos added an uncensored Qwen3.6-27B model to its on-chain registry and expanded the base Qwen3.6-27B to include AWQ-INT4 quantization. The abliterated version removes safety refusals, while the AWQ-INT4 variant reduces VRAM requirements. Miners can reconfigure via `verathos setup`.

Dashboard displays stale validator thresholds (0.060 crown quality) while live validator enforces correct threshold (0.200). Team confirmed validator is operating correctly in production; dashboard UI fix is in progress. Separately, team clarified subnet's fund-locking status and noted immunity to ownership issues as a sub-one-year-old subnet.