⌞· τaoli ·⌟

Bittensor changes FAST. Bittensor Quest now has a newsletter delivered to your email, weekly! If you have no time to: — Scour all subnet's discord channels — Read through 128 Github repos — Browse X for hours on end This newsletter is for you. It's free. Made with ❤️ 👉newsletter.bittensor.quest/?…

This week's #Bittensor Weekly is out! Iota (SN9) shipped Decide 4B, its first public model: a 4B decision model under Apache-2.0 that returns probabilities instead of text, and scored 78.6 on the public JevBench split against 62.4 for OpenAI's decisions API. Sixteen RTX A4000s linked over the internet did the training, no cluster. It leads this issue. Also in the issue: Ditto (SN118) killing its 50% miner burn and moving to 18/33/49, Apex (SN1) raising the peg noise until scripted entries fell from 0.85 to 0.09, ComputeHorde (SN12) patching the holes Dovetail found in job isolation, Gradients (SN56) reversing a duplicate flag after reading the code, Affine's (SN120) queue fight, Niome's (SN55) seed questions, Blockmachine's (SN19) 11-chain emission split and the miners who got two days' notice, Kokoah (SN39) taking the slot, and Instant's (SN46) first buyback on Friday. Then the week's numbers: 1,161 messages across 69 subnet channels in two days, 25 channels with nothing all week, 800 miners training at home on Iota with 1,600 queued, and 0.0007 of a point deciding Vidaio's (SN85) upscaling crown. The Human Moment is Nova (SN68), where two crashed validators and a straight answer earned the week's best compliment. Four verified links close it out. Read the full issue: 👉newsletter.bittensor.quest/p… $TAO #Bittensor

78.6 on JevBench (a benchmark for decision models) from a 4B open-weights model. OpenAI's own decisions API scores 62.4 on that board. Iota (SN9) shipped Decide 4B, Macrocosmos's first public model. It doesn't write text. You give it a state (a ticket, an agent trace, a policy) and typed questions, and it returns a probability for each option in one forward pass, about 33 ms on a GPU. The weights are Apache-2.0 on Hugging Face and 8.4 GB, small enough for one card. Trained on 16 consumer GPUs kept in sync over the internet, no cluster. That is the whole run. Follow Iota (SN9): @IOTA_SN9 $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

Bittensor's chain gets 15% of Blockmachine's (SN19) miner emission. BSC and Polygon get 17.5% each. Blockmachine sells RPC access (the servers wallets and apps call to read a chain), and it now splits miner emission across the 11 chains it serves: - BSC and Polygon: 17.5% - Bittensor: 15% - Ethereum and Base: 10% - Robinhood: 7.5% - Mantle, Optimism, Arbitrum, Arc: 5% - Scroll: 2.5% Cost and demand set the order. BSC and Polygon cost the most to run, and Bittensor has Taostats, its biggest customer, at 15%. A US region opens at Vint Hill, Virginia, with payments there from Monday at the earliest. Follow Blockmachine (SN19): @blockmachine_io $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

193 miners got charged for missing 2,915 episodes in AdTAO's (SN21) 3 November basket. The run never reached them. Each day miners get a basket of ad episodes to predict. This one was bigger than usual, the run hit a resource limit at the model stage, and nothing from UID 42 up ran. UIDs 0 to 41 scored fine, which is how miners spotted the pattern. SN21Dev raised the limits, ran every model that had not run, and cancelled the charges. Each cancellation is published next to the charge it replaces. The day's report went out again as BD-2026-11-04-COR-1, and the leaderboard weights moved with it. Follow AdTAO (SN21): @adtao_ppcrebel $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

Bro, what the fuck is this?!?

Cathedral (SN94) rents sandboxes at 3.5 cents an hour, and its ten-box test came back clean. Jobs run on AMD confidential-computing hardware, sealed so the host cannot look inside, with attestation (a hardware proof of which machine ran the job) and signed receipts. From 2 vCPU and 4 GiB at $0.035/hour to 8 vCPU and 16 GiB at $0.14. Ten sandboxes, same job, median 3 seconds to ready, all ten exited 0. Mining is not open. Seed nodes sit at UIDs 240 to 242, and the team wants more validators running its binary. Miners get paid for making the sandbox harder to break, faster to start and cheaper to run. Follow Cathedral (SN94): @cathedralhq $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

Remember, remember, the 5th of November.

Alpha is not a token sitting in a wallet. It is stake tied to a subnet, a hotkey and a coldkey, which is why normal DeFi lending never touched it. Forge (@ForgeLending) gets there through the staking precompile (a built-in contract that can move native stake). You approve a set amount of alpha for a chosen subnet and hotkey, the stake moves into a wrapper, and that wrapper becomes your collateral. It keeps earning the whole time it sits there. Mainnet terms give the borrower 75% of the yield the collateral generates, with 25% going to Forge as a performance fee. The TAO side looks like a standard lending pool. Suppliers earn root yield plus borrower interest that climbs with utilization (the share of the pool that is lent out). Borrowing power comes from each market's LTV (how much TAO one unit of collateral supports), and cross the liquidation threshold and a liquidator repays your TAO debt and takes the wrapped alpha, then unwinds it at whatever exit price they can get. The lending core is Venus pool code held byte for byte against upstream, with the Forge material, meaning the staking flows, oracles and liquidator, reviewed on its own. tao.com audited the wrappers underneath. And it doesn't start wide: first markets are held at 10-12% of subnet pool liquidity, parameters sit behind a multisig, and @EndureNet (SN30) supplies the risk forecasts meant to take over later. What LTV would you lend TAO against an asset priced by a single subnet pool and nothing else? teapawt.net/ForgeLending/status/21… $TAO #Bittensor

KubeTEE (SN90) put its first European cluster in Luxembourg, so EU workloads can run inside a TEE without leaving the bloc. A TEE (a sealed slice of the CPU that even the machine owner cannot read) is the whole product. Capacity is attested. The customer gets a hardware proof of which machine ran the work. More regions come online as capacity fills. Each one widens where workloads can run. Miners get in through the team for now, with the permissionless path still in staging. Follow KubeTEE (SN90): @KubeTEEAI $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

A subnet owner needs $TAO for a year of devs and compute. The only move today is to sell alpha, and every delegator watches it happen. That sale does two jobs. It funds the work, and it tells the pool that the person running the subnet is cutting their own exposure. Forge (@ForgeLending) on #Bittensor opens mid-November with a second option. Post alpha as collateral, borrow $TAO against it, keep the position and keep the yield. On the other side, $TAO suppliers earn root yield as a base rate plus borrower interest when the pool gets busy. Two numbers make it worth watching. Teams have signed more than $5M of alpha commitments before launch. And each alpha market opens capped at 10-12% of that subnet's pool liquidity, raised only once a market has a track record. Risk settings start under a Guardian Multisig (a wallet that needs several signatures to change anything), with @EndureNet (SN30) forecasts meant to inform them as that network's scoring record builds. If a subnet owner borrows against their own alpha instead of selling it, where's the line between conviction and doubling down on the asset their project depends on? teapawt.net/ForgeLending/status/21…

ComputeHorde (SN12) patched the holes a security team found in how it runs other people's jobs. The Dovetail researchers broke job authentication, the isolation between one job and the next, and the handling of job artifacts. They sent reproducible attacks with the report. Everything they found is fixed, and the patch is public. A path traversal bug (a way to read files outside a job's own folder) reported through GitHub in July had gone unnoticed until this review. It's patched now, and the team says it will stop missing those reports. Follow ComputeHorde (SN12): @ComputeHorde $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

Gradients (SN56) flagged two Round 2 miners as duplicates, then reversed the call once the code was read closely. The dedup check, run through the Claude API, saw the same hyperparameters (the training settings) down to the decimal. One miner had left those numbers in his uploaded artifacts, the other admitted reusing them from last week. The code and training logic differ enough that both attempts stand as work, with a leaked config doing the damage. Round 2 is back on with the Round 1 winners competing again. Every future dedup call needs two reviewers, public hotkeys and a 24 hour window for miners to argue before anything is final. Follow Gradients (SN56): @gradients_ai $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

Harnyx (SN67) crowns two champions now: the best deep research report, and the most accurate answers on their own. The main title still needs a full report with every claim traced to its source. The new one scores straight accuracy across every task. Each title decays its emissions separately, and the champion budget split follows the accuracy gap between the two winners. One submission can hold both. Follow Harnyx (SN67): @harnyx_ai $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

0.0007 of a point separated first from second in Vidaio's (SN85) upscaling competition. UID 164 takes 90% of miner emissions for the next two weeks at 0.2410. UID 150 gets 4% at 0.2403. 35 entries enrolled and 31 were scored. Four got rejected, one of them for a missing LICENSE file. Competition 005 (streaming compression) opens tonight. Follow Vidaio (SN85): @vidaio_ $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

A 50% miner burn is gone from Ditto (SN118), and validators and stakers went from 41% to 49% of emissions. The split now reads 18% owner / 33% miners / 49% validators and stakers, live on-chain since Tuesday. Miner share looks smaller on paper, 41 down to 33, but half of that old slice used to be burned, destroyed instead of paid out. None of it burns now, so more TAO lands with the people running the agents. More TAO stays in the pool too, which means a deeper market for anyone buying or selling SN118. Follow Ditto (SN118): @heydittoai $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

A professor feeds a student essay into a detector and gets back 92% AI. There is no evidence trail and no second opinion, just a number, and now a misconduct hearing. Detectors built that way guess from writing style. When the guess lands on a human, the human has to prove they wrote their own thoughts. It's AI (@ai_detection), SN32 on #Bittensor, went after that failure first. On ASAP 2.0, 25,000 essays written by students, it holds a 0.8% false positive rate. That is about 1 human flagged for every 125 essays. On MGTD, a comparison across 15 datasets and roughly 2 million samples, it scores 92.0 ROC-AUC. GPTZero scores 88.7. The subnet's scoring points the same way. Miners earn on F1, on average precision, and on how rarely they label human writing as AI. Getting the human side wrong costs more than missing a machine. If a detector flags 1 in 125 human essays, should it get a vote in a misconduct hearing? $TAO

Don't get me wrong, I love Chutes. But honestly, this is kinda pathetic. Kimi K2.7 Code was release in June...

SayGM (SN28) told Bedrock miners to move to v0.5.18. Demand for Opus, Anthropic's top Claude model, is running ahead of supply. The release helps Bedrock workers survive heavy load, so more requests get served instead of dropped. Miners update with gmcli and redeploy. Paying demand outrunning capacity is the problem every inference subnet says it wants. Follow SayGM (SN28): @say_gm_ $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest

Nov 6 at 5pm ET is Instant's (SN46) first buyback and burn. The subnet has onboarded its first paying users through Stripe, though no funds have settled yet. Crypto payments open to API users this week while Stripe verifies USDC, which skips chargeback risk. Every buyback and burn lands on two public coldkeys (wallets), so the first one is checkable. Follow Instant (SN46): @thinkinstant $TAO #Bittensor — Everything important that happens in Bittensor, delivered weekly: 👉newsletter.bittensor.quest