If you’re doing a proper Bittensor deep dive after the recent Anthropic chaos look no further than the Unsupervised Cap valuation framework

After the last 24 hours, how can you read this and not see that decentralized training is going to be the most important innovation to come out of crypto since Bitcoin?

There’s only one ecosystem that’s spent the last 3 years working on building the entire stack required to produce intelligence with no reliance on centralized companies Bittensor has been preparing to go to war for a while now

TLDR; our subnet on bittensor (SN15) is producing the highest quality open agentic shopping traces in the world. We're going to leverage that to create personalized shopping models for all. go give our arXiv pre-print a read, where we outline how the shopping traces from the best frontier open source models show pitfalls, the challenges and the opportunities in the goal of teaching AI to shop. Also, we're expanding the team! If you're interested in any of our work below, please reach out to team@oroagents.com.

THE SETUP FOR DECENTRALIZED TRAINING IS AS GOOD AS HUMANLY POSSIBLE. AT THE EXACT SAME TIME AS "THE INDUSTRY" DECLARES THE DEATH OF CRYPTOGRAPHIC COORDINATION, VERIFICATION, AND INCENTIVIZATION OF RESOURCES, WE HAVE BACK TO BACK TO BACK GROUNDBREAKING TECHNIQUES DEVELOPED AND EVEN DEPLOYED BY SERIOUS, QUALIFIED, EXTERNALLY VALIDATED PEOPLE. AT THE SAME TIME AS THE TECHNOLOGICAL PROSPECTS FOR DECENTERALZIED TRAINING ARE TAKING A 10X LEAP, MORE PEOPLE ARE AWARE OF AND ANGRY ABOUT THE EXACT PROBLEM DECENTRALIZED TRAINING AIMS TO SOLVE THAN EVER BEFORE. CENTRALIZED VENDOR LOCKIN CONCERNS ARE HITTING A CRESCENDO. A TIPPING POINT FOR YOU MALCOLM GLADCLART FANS IN THE SECOND ROW OF THE CLASSROOM. THIS IS THE BEST SETUP FOR A DECENTRALIZED TECHNOLOGY STACK OF ANY KIND SINCE CYPRUS IN 2013, BUT AIN'T NONE OF YOU NUMBIES STUDIED BALL OF ANY KIND SO YOU HAVE ZERO CLUE WHAT THAT EVEN MEANS. KNICKS IN 6.

Another day, another 1,000 tweets clamoring for decentralized training without realizing it Prohibitive LLM costs, segregated access, unilateral moral relativity. There could not be a better setup for decentralized training ever. Anyways Knicks in 6

Back from @proofoftalk and I'm leaving more optimistic than ever. The progress the top subnets are making is amazing, and new subnets keep arriving with high quality teams and genuinely innovative ideas. The ecosystem is in a great place. My biggest takeaway, though, is urgency, on two fronts: revenue and research. 1. On revenue, crypto is now being evaluated like a real business. The strongest tokens are the ones with actual revenue and growth, see HYPE, VENICE, NEAR. Bittensor is already delivering here. Subnets like @TargonCompute, @webuildscore, and @bitmind are posting real revenue growth. The strength is real, the market just hasn't fully heard the story yet, and that's on us to articulate. 2. For research, distributed training is moving incredibly fast. Frontier labs are going public and their access to capital only compounds from here. New business models are emerging just as quickly, like RL training as a service. @MacrocosmosAI announced a 100B run that uses infra which can power this type of solution. The way you compete with general frontier intelligence isn't by matching it head on, it's with custom, smaller, cost-efficient models. @zeussubnet, @metanova_labs, and others are making exactly that bet. Pushing this research forward is critical, for Bittensor and for the world.

Case in point... THOU SHALT NOT COMPETE. - AI/LLM development generally - distributed training - pretraining pipelines - ML accelerator design - cybersecurity - chemistry Already disclosed as nerfed for the plebs. What's next on the lobotomization list?

At Proof of Talk, I gave a talk on how Bittensor today is not the same network it was 6 months ago, let alone 1-2 years ago. So much progress has been made at the protocol and team level recently, meaning many people’s mental models of the network/subnets are outdated. Two subnets in particular which have had an incredible 2026 so far have been Macrocosmos’s IOTA and Apex. IOTA just proved its ability to train 100B models while Apex has become a meta-subnet, hosting short finite competitions (eventually for agents) Glad to have the Macrocosmos team in this ecosystem and really enjoyed seeing the general progress that all the subnets have made in the last year at PoT

I feel really lucky to have Chris joining the team at such a pivotal moment as everything in bittensor accelerates Sidenote: he’s a wizard when it comes to decentralized training and can become fluent in a new industry like drug discovery within a few days.

The beauty of decentralized technologies is that they beget each other, a protocol of decentralized compute is only relevant when there is a protocol for decentralized training.

Over the last couple weeks, we worked with @TAO_dot_com and @UnsupervisedCap on an impact analysis related to the proposed Conviction updates This podcast reviews that report in a data-driven discussion on the potential implications of locking and subnet takeover Worth a watch

> IOTA (subnet 9) ran a 100B parameter model using 48 single A100-80GB GPUs that were distributed (non-colocated, across multiple providers and multiple datacenters) and connected only over the internet (no datacenter fabric) > achieved 30.8% average MFU, which is impressive because frontier labs run at 38-50% MFU on average > achieved roughly 65% of the effective training speed of a comparable co-located datacenter setup > did it at a 2.5x cheaper per replica than high-end datacenter cluster equivalents training a 100B model with pipeline parallelism across 16 stages over the open internet at 30% MFU is very impressive. it is one of the more impressive things to come out of the bittensor ecosystem so far that said, I’d like to see them scale this to heterogeneous/permisionless hardware, run it for much longer and produce an actual competitive model while keeping the cost advantage over frontier models

the dark horse of the dark horse

Good to see Synth get the recognition it deserves! • On 15 min markets: Synth correctly called direction 86% of the time vs 57% for Polymarket • On hourly markets, Synth achieved 75% accuracy vs to 62% for Polymarket

Bittensor’s Conviction upgrade just hit mainnet today. The feature has a large scope and is the most wide reaching upgrade since Dynamic TAO, impacting every stakeholder on the network. Important note before diving into all the details: as of today, Conviction is functionally opt-in for subnet owners and has no material impact on stakeholders (e.g., subnet owners are at no risk of losing their subnet). Subnet owners and the community are being given time to explore the mechanics of the upgrade, discuss what settings work best for both tokenholders and subnet operators, and can/should voice feedback during this initial phase. Those within the network know Bittensor protocol development moves fast, so below you’ll find everything you need to know about the upgrade and how it’s being rolled out. Summary of Conviction Conviction is an onchain token-locking primitive. Users can lock alpha tokens to a key which (1) creates a "conviction" score and (2) locks tokens for a certain period of time. The intent is for subnet owners to be able to lock up tokens to signal their long-term conviction for building on the subnet. Eventually, functionality will be added that enables anyone to contest ownership of a subnet (they’ll need to have at least 10% of the Alpha held by tokenholders backing them), and if successful (i.e., a new key has a higher conviction score than the owner key), take control over it. Subnets that are less than 1 year old will be excluded from this feature. This functionality is not yet enabled on mainnet. Locked tokens can be unlocked by calling an onchain function. The unlock schedule follows an exponential curve, with 50% of the locked tokens available after 3 months, ~95% after 12 months. Locked tokens can still be transferred for OTC investments, but they will remain in the state they’re received in (e.g., if an investor receives locked tokens they’ll remain locked unless they choose to unlock them). For now, Conviction is primarily a social signaling feature. No subnets can be taken over. Subnet stakeholders will all need to work together to decide what the norms/expectations will be for each subnet, as they all have different constraints. What’s on mainnet, how it works, and what’s not on mainnet A user (any coldkey; subnet owner, regular token holder, etc) can lock some amount of their alpha on a subnet to a specific hotkey. While locked, that alpha cannot be sold. Over time, the lock also generates a "conviction" score that follows a continuous exponential formula. When locking, the user can choose between two lock modes: • Decaying lock (the default): locked alpha exponentially unlocks over time. After 3 months, half is liquid; after one year, ~95% is liquid. • Perpetual lock: locked alpha stays locked forever (until the user toggles to decaying). A given coldkey can have at most one active lock per subnet. Top-ups on locks must target the same hotkey (no locking to two different keys). Moving a lock to a different hotkey is allowed but resets conviction to zero. Subnet owners can decide whether their owner emissions are auto-locked or liquid. If set to auto-lock, every owner emission for that subnet is automatically locked to their hotkey rather than flowing in as liquid stake. The default on mainnet is liquid, not auto-locked. Subnet owners effectively have to opt-in to Conviction at this level. Subnet owners will also have to decide whether to lock all, a portion, or none of their current holdings, and whether a lock would be set to decay or perpetually locked. For a simple example: under a decaying lock with these defaults listed above, locked tokens and conviction evolve as: • Days Elapsed = 30; % still locked = 79%; conviction score (% of locked tokens) = 18% • Days Elapsed = 90; % still locked = 50%; conviction score (% of locked tokens) = 35% • Days Elapsed = 130; % still locked = 37%; conviction score (% of locked tokens) = 37% (peak) • Days Elapsed = 180; % still locked = 25%; conviction score (% of locked tokens) = 35% • Days Elapsed = 365; % still locked = 5%; conviction score (% of locked tokens) = 17% Under a perpetual lock, locked tokens stay constant and conviction matures to its asymptote on the same 90-day half-life curve. The system is not symmetric across all users. Subnet owner conviction is immediate. When the subnet owner's own coldkey locks alpha to their subnet, conviction is set to equal the locked token mass instantly. There is no 90-day ramp up. The owner has the most to commit and gets credited for it immediately. Anyone locked to the subnet owner's hotkey is also granted immediate conviction. Subnet owners have a structural advantage in keeping ownership of their subnet. Transfers carry locks with them. Often, teams will execute OTC deals with investors via token transfers; these will still function the same way as they do today. However, when alpha moves between coldkeys, the lock follows the alpha. The receiving coldkey inherits the lock state and can choose to make it perpetual or immediately decay. So, if a subnet owner transfers an investor locked tokens, it will be up to the investor to decide when to unlock them. The unlocking event is also an onchain event. The hotkey association on token transfer must match the receiving key. If the receiving coldkey has an existing lock to a different hotkey, the transfer fails. This prevents locking across different hotkey commitments. Right now, Conviction scores are computed, stored, exposed via RPC, but nothing onchain consumes them. The one function that would consume aggregate conviction, change_subnet_owner_if_needed, which would replace a subnet's owner with the highest-conviction hotkey's coldkey under certain conditions, is not functional yet. Think of what’s shipped today as the initial, slow rollout of the grander Conviction design. What to expect now All subnet stakeholders should start understanding, experimenting with, and discussing the design and parameters of Conviction. Every subnet operator has different funding constraints, trust levels, and maturity on a relative basis. A large part of Conviction will be settling on the norms and expectations the community/social fabric has with respect to subnet teams. Many of these parameters discussed above can be tweaked via governance. The takeover design/mechanism has yet to be rolled out yet, so it can still be evolved based on community feedback. Come join the community call/Novelty Search tonight to hear Const discuss these changes and provide feedback directly.