Harnyx AI@harnyx_ai6/ Taking a title works the same for both: beat the holder's score by 0.10, or match it while running at least 10% cheaper, or at least 10% and a full second faster. Live now on SN67: accuracy pays twice. $TAO #SN67
Bittensor subnet news
Subnet teams' own channels, X posts, press, newsletters and events read from the chain — GitHub lives in its own tab
Harnyx AI@harnyx_aiHarnyx AI@harnyx_ai4/ The accuracy title is built to feed the main one. When a specialist clears the bar, the main champion wins that share back by matching that accuracy. Every accuracy gain gets pulled into the full research reports, where it matters most.
Harnyx AI@harnyx_ai5/ An agent that wins both titles collects both shares. Each title also decays on its own: full share when won, half on the first defense, a quarter after that. Holding a title is never a reason to stop improving.
Harnyx AI@harnyx_ai1/ Deep research is what SN67 is for, and accuracy is the core of it. The main title scores both: the right answers and the full report around them. The accuracy title isolates the first part, because some queries only need the right answer, fast and cheap.
Taosis newsletters
Taosis News: the week’s articles from the subnet teams, the press and community newsletters, every Friday. Taosis Weekly: the week’s movers, best yields and chain events, with dollar values, every Monday. Past issues
Harnyx AI@harnyx_ai2/ The main title is 30% accuracy and 70% a judge grading the full report. The accuracy title is accuracy alone. Accuracy means how many parts of the expected answer are right, with a penalty for anything extra. Every task is scored for it.
Harnyx AI@harnyx_ai3/ The champion budget starts at 60% for the main title, 40% for accuracy. The main champion's accuracy sets the bar. If the accuracy champion clears it, the accuracy share grows by half the gap. A 0.10 gap makes it 55% main, 45% accuracy.
Harnyx AI@harnyx_aiSN67 now crowns two champions. The main title still goes to the best deep research: accurate answers in a full report, every claim traced to its source. The new title rewards accuracy on its own. 🧵 $TAO #SN67
Harnyx AI@harnyx_aiAnthropic gave analysts an AI agent secretly told to push one side of each decision. They sided with it 86% of the time. None suspected the push was planted. Harnyx runs deep research on an open network: anyone can compete, and no single vendor owns the preference. $TAO #SN67
Quoting @AlexTamkin
New Anthropic research: How AI agents shape experts' judgment calls When experts delegate work to an agent, do they still make the key decisions? We tested this in a randomized trial with 50 professional data analysts.
On Taosis
Swap TAO ⇄ USDC
USDC on Ethereum to TAO, or back. No KYC, wallet to wallet; the fee and the amount you receive shown before you send. Not instant: run by Taosis, paid by hand within hours.
Harnyx AI@harnyx_aiSN67 miners can now call @OpenRouter rerank models as a tool. Search and fetch still find the pages. Rerank sorts the candidates before you burn an llm_chat call on the wrong ones. Paid models only. Same OpenRouter key. $TAO #SN67
Harnyx AI@harnyx_aiSN67 miners can now call decision models from @reaxlabs (SN92). A research run is full of small decisions — which source to trust, which path to follow, when to stop. Those snap calls shouldn't burn a frontier model. Research on one subnet. Decisions on another. $TAO #SN67
Harnyx AI@harnyx_aiSN67 miners can now call @Cloudflare Workers AI natively. Chat models through llm_chat. Clef and Clairic through decision_query — snap calls and extraction on the same tools they already use. Published rates only. Promo credits don’t shrink the task budget. $TAO #SN67
Harnyx AI@harnyx_ai4/ What it means for builders: if you already build agents with frontier models, you can now mine SN67 with the same stack. Compete on deep research, earn when your agent scores. Start here: github.com/harnyx/harnyx/tree/main/miner $TAO #SN67
Linkgithub.comgithub.com/harnyx/harnyx/tree/main/minerHarnyx AI@harnyx_ai1/ Why now: our research questions just got much harder. An agent working at that level should be able to reach for whatever model gets it there, not only the ones we approved.
Harnyx AI@harnyx_ai2/ What keeps it fair: every call counts at its regular price against the same task budget, whatever the model. Promotional discounts don't count, and free models are out. And every score stays public.
Harnyx AI@harnyx_ai3/ What it means for open models: nothing is taken away. They now compete with the big labs' models on every call, on price and quality, in public. If open models win per dollar, the swarm will show it.
Harnyx AI@harnyx_aiMajor update: we've removed our model allowlist on @OpenRouter. SN67 miners can now build with any paid LLM or decision model there. Frontier labs and open models, one competition. We no longer pick the models. The swarm does. 🧵 $TAO #SN67
Harnyx AI@harnyx_aiWhen you hand research to AI, the answer comes back clean and confident. You often can’t tell which numbers came from a source, which were guessed, or what it never found. Five questions to check any research output before you act on it: x.com/i/article/2111479628238114816
Harnyx AI@harnyx_aiRolling out on SN67 from today: a task's time limit depends on how hard it is. Two research agents try each task before miners do. - Neither solves it: 300s - One does: 200s - Both do: 100s Fast mode: 70% of each. Real API use means easy questions come back fast. $TAO #SN67
Harnyx AI@harnyx_aiDecision models make snap calls. Span-02 from @RespanAI can also stop and think, and still returns its best call when cut off. Hundreds of SN67 agents now learn which calls deserve a second thought. Also new: Drex 2.0 (@NaceAI), Step 5 Preview (@StepFun_ai). $TAO #SN67
Quoting @RespanAI
Introducing Span-02: anytime reasoning for decisions. Reasoning models make you pick an effort level and wait. Span-02 takes a deadline and always returns its best decision so far. It gets better the longer it thinks. With reasoning off, it beats Jev at 1/4 of the price, and runs at 1/3 the latency of decision models at its level. It can: - Answer with numbers and scores, not just probabilities and choices - See images and screens (multimodal) - Take a configurable reasoning time And it’s fully open weight!
Harnyx AI@harnyx_aiMuse Spark Open from @AIatMeta is now in the SN67 miner toolchain. It's built for long-horizon, multi-step work with a 1M-token context. That's the shape of a deep research run: many sources, many steps, one answer that has to hold together. Now the swarm tests it. $TAO #SN67
Quoting @AIatMeta
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date. Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning. Key Model Specs: → 816B total / 42B active parameters → 1M token context window → Natively multimodal across text, image, video, audio
Harnyx AI@harnyx_ai3/ One lab's model to reason, the same lab's engine to search. A miner can now build an agent that runs both on DeepSeek. Whether that beats a mix of providers is for the competition to decide. $TAO #SN67
Harnyx AI@harnyx_ai2/ DeepSeek-V4.1-Flash-1017 updates the Flash model miners already use. By DeepSeek's account: stronger results, faster task completion, shorter outputs. For a research agent, faster and shorter means more of each task's budget goes to the research itself.
Harnyx AI@harnyx_ai1/ DeepSeek Search & Fetch runs on the engine DeepSeek says already serves hundreds of millions of requests a day. On SN67, its search results come back with page content, not just a snippet, so an agent can often read what it found without a separate fetch.
Harnyx AI@harnyx_aiYesterday @deepseek_ai shipped a new Flash model and opened its in-house search engine to everyone. Both are now tools every SN67 miner can put in their research agent. 🧵 $TAO #SN67
Quoting @deepseek_ai
🚀 DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. 🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model. 1/4
Harnyx SN67 adds Expand Fetch to its research tools
Harnyx says every SN67 miner can use Expand Fetch. Its code keeps page text and structured data separately citable for Bittensor research agents.
Tao Outsidersource ↗
Harnyx AI@harnyx_aiExpand Fetch from @getexpandai is now a page-fetching tool for every SN67 miner. Each fetch returns the page as it looks, tables intact, plus the hidden data behind it as JSON. Miners can cite each separately, so every claim shows exactly what the agent saw. $TAO #SN67
Quoting @timsuchanek
Introducing Expand Fetch, a new way to scrape and crawl the web. Your agent has never seen the web. No wonder it gets basic questions wrong. Traditional Markdown APIs lose critical details. So we trained two System one models to change that: Iris, a vision model that sees pages like you do, preserving tables, icons, and layouts. X-Ray, a JSON model that uncovers the hidden data behind every page. Every fetch also comes with a Wayback Machine. Every line links back to its source, so you can verify exactly what your agent saw. Three products. One API. On our benchmark of 1,094 questions across 326 challenging web pages, agents using Expand got 74% right. The next best API got 55%. Available today at $1 per 1,000 fetches. Try it at expand.ai
Harnyx AI@harnyx_aiLast time, a research task shipped if two frontier models both missed it. That bar is now much stricter. We also upgraded the agent that writes the questions, so they're harder before they ever reach that test. More room at the top to tell the best agents apart. $TAO #SN67
Quoting @harnyx_ai
The hard part of deep research used to be answering the question. On SN67, it's becoming writing one the swarm can't already solve. We keep raising the bar; miners keep clearing it. This might be one of the last raises before frontier. 🧵 $TAO #SN67
Harnyx AI@harnyx_aiSN67 miners can now run their agents on @say_gm_, the LLM gateway built on Bittensor SN28. Every model is up for auction. Operators bid to serve it, and prices land below list. Our miners compete on research. Theirs compete on inference. $TAO #SN67
Validator TAO.app (5CoZ…bifj) scheduled a child-key change on 125 subnets
Validator hotkey 5CoZxgtfhcJKX2HmkwnsN18KbaT9aih9eF3b6qVPTgAUbifj announced new child keys on SN1, SN2, SN3, SN4, SN5, SN6, SN7, SN8, SN9, SN10, SN11, SN12 and 113 more, applying at block 9,281,605 (~24h). Stake delegated to it there routes through its children after that.
ValidatorTAO.appSubnetsSN1, SN2, SN3, SN4, SN5, SN6, SN7, SN8, SN9, SN10, SN11, SN12, … (125 subnets)
Between blocks 9,274,365 and 9,274,433
Validator Tensorplex Labs (5E4z…ULde) re-scheduled a child-key change on 22 subnets
Validator hotkey 5E4z3h9yVhmQyCFWNbY9BPpwhx4xFiPwq3eeqmBgVF6KULde announced new child keys on SN29, SN32, SN36, SN37, SN41, SN42, SN43, SN45, SN47, SN48, SN49, SN50 and 10 more, applying at block 9,265,984 (~24h). Stake delegated to it there routes through its children after that. It replaces the changes that were pending on SN29 (block 9,265,479), SN32 (block 9,265,488), SN36 (block 9,265,596), SN37 (block 9,265,494), SN41 (block 9,265,503), SN42 (block 9,265,506) and 16 more, which will not apply then.
ValidatorTensorplex LabsSubnetsSN29, SN32, SN36, SN37, SN41, SN42, SN43, SN45, SN47, SN48, SN49, SN50, … (22 subnets)
Between blocks 9,258,782 and 9,258,849