With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Bitsota is a decentralized research network (Subnet 94 on the Bittensor blockchain) designed to evolve machine learning algorithms through competitive, distributed computing. In essence, it transforms the process of AI research into a global mining competition: participants (called miners) around the world devote their CPU/GPU power to solving open machine-learning challenges, and only receive rewards when they successfully produce a better model (i.e. achieve a new state-of-the-art result). This is a departure from traditional crypto mining or static AI tasks – Bitsota focuses on “real research” progress rather than consuming compute on fixed workloads. Researchers (or “problem owners”) can post an AI challenge (with a dataset, metric, and baseline), and miners compete to beat the baseline. If a miner’s model surpasses the current best performance (verified on a hidden test set), the network confirms the breakthrough and automatically pays out a reward. In summary, Bitsota serves as a decentralized AutoML platform where “compute evolves intelligence” collaboratively on Bittensor.
Key characteristics of Bitsota’s approach include:
Rewarding breakthroughs, not just compute: Bitsota only issues payouts for “verified breakthroughs” that advance the state-of-the-art, rather than paying for mere participation or idle hashing. In other words, miners earn token rewards only when a new high score is achieved and validated, aligning incentives with genuine AI progress (instead of wasting GPU hours on unproven runs). This ensures every unit of compute contributes to measurable improvements in AI models.
Open participation: The network is open to potentially unlimited miners, lowering the barrier to entry with a one-click desktop miner (no complex setup or coding required). Anyone with a CPU (and eventually a GPU) can join the mining process with minimal friction – “no command line, no configs, just start and go.” This broad inclusivity means Bitsota can harness a globally distributed pool of CPUs/GPUs of all sizes, turning distributed compute into “a massive search engine for AI models.”
Trust-by-default validation: Every claimed model improvement is subject to rigorous verification. Bitsota employs hidden test datasets and a quorum of validator nodes to independently evaluate submissions. Only if a majority of validators confirm that a miner’s model truly beats the previous benchmark on these secret tests will the result be accepted. This validator quorum (N-of-M) consensus, combined with “reproducible receipts” (audit logs of each validation), makes each win fully auditable and credible. The use of hidden test cases prevents miners from overfitting or gaming the metric, enforcing honest progress. In short, Bitsota “trusts but verifies” every breakthrough before any reward is released.
In practice, Bitsota’s first live challenge is based on Google’s AutoML-Zero concept – an evolving algorithm task nicknamed “AI that creates AI.” Miners start with basic mathematical operations and use genetic programming to evolve new machine learning algorithms from scratch, striving to outperform a baseline on a simplified CIFAR-10 image classification task. This means instead of training a fixed neural network, miners are actually searching for better algorithms automatically. The network coordinates this search at scale: miners continuously generate and test algorithmic “children” locally, and whenever someone’s algorithm scores above the current best (on a validation set), they submit it for verification. Validators then re-run that model on hidden test data to check the claim. If the improvement is confirmed, Bitsota declares a new state-of-the-art (SoTA) and immediately releases a reward to the winning miner. If not, the claim is rejected and no reward is given, ensuring that only genuine progress gets rewarded.
Crucially, Bitsota is built on the Bittensor blockchain (layer-1), which provides the token incentive structure (the $TAO token emissions) and secure, decentralized governance for the subnet. Bitsota leverages Bittensor’s dynamic tokenomics so that its miners and validators are rewarded in proportion to the verified value they add. In Bitsota’s case, the “value” is improved AI performance, and the network’s rules are tuned to favor breakthroughs: for each block cycle, ~90% of the allotted mining reward is burned (or remains unspent) unless a new SoTA is found, and up to 10% is allocated to the winner of a confirmed breakthrough. This “results-gated” reward mechanism means Bitsota isn’t constantly paying out tokens unless real progress is happening – essentially “buying breakthroughs, not compute”. Over time, this creates a competitive dynamic: miners are incentivized to innovate quickly (to capture the prize before others), and problem owners get solutions only when milestones are met, paying only for success rather than effort. The end result is a provably efficient research marketplace: by aligning economic rewards with research outcomes, Bitsota aims to crowdsource AI innovation at scale, turning wasted idle compute worldwide into a collective engine for finding better algorithms.
Bitsota is a decentralized research network (Subnet 94 on the Bittensor blockchain) designed to evolve machine learning algorithms through competitive, distributed computing. In essence, it transforms the process of AI research into a global mining competition: participants (called miners) around the world devote their CPU/GPU power to solving open machine-learning challenges, and only receive rewards when they successfully produce a better model (i.e. achieve a new state-of-the-art result). This is a departure from traditional crypto mining or static AI tasks – Bitsota focuses on “real research” progress rather than consuming compute on fixed workloads. Researchers (or “problem owners”) can post an AI challenge (with a dataset, metric, and baseline), and miners compete to beat the baseline. If a miner’s model surpasses the current best performance (verified on a hidden test set), the network confirms the breakthrough and automatically pays out a reward. In summary, Bitsota serves as a decentralized AutoML platform where “compute evolves intelligence” collaboratively on Bittensor.
Key characteristics of Bitsota’s approach include:
Rewarding breakthroughs, not just compute: Bitsota only issues payouts for “verified breakthroughs” that advance the state-of-the-art, rather than paying for mere participation or idle hashing. In other words, miners earn token rewards only when a new high score is achieved and validated, aligning incentives with genuine AI progress (instead of wasting GPU hours on unproven runs). This ensures every unit of compute contributes to measurable improvements in AI models.
Open participation: The network is open to potentially unlimited miners, lowering the barrier to entry with a one-click desktop miner (no complex setup or coding required). Anyone with a CPU (and eventually a GPU) can join the mining process with minimal friction – “no command line, no configs, just start and go.” This broad inclusivity means Bitsota can harness a globally distributed pool of CPUs/GPUs of all sizes, turning distributed compute into “a massive search engine for AI models.”
Trust-by-default validation: Every claimed model improvement is subject to rigorous verification. Bitsota employs hidden test datasets and a quorum of validator nodes to independently evaluate submissions. Only if a majority of validators confirm that a miner’s model truly beats the previous benchmark on these secret tests will the result be accepted. This validator quorum (N-of-M) consensus, combined with “reproducible receipts” (audit logs of each validation), makes each win fully auditable and credible. The use of hidden test cases prevents miners from overfitting or gaming the metric, enforcing honest progress. In short, Bitsota “trusts but verifies” every breakthrough before any reward is released.
In practice, Bitsota’s first live challenge is based on Google’s AutoML-Zero concept – an evolving algorithm task nicknamed “AI that creates AI.” Miners start with basic mathematical operations and use genetic programming to evolve new machine learning algorithms from scratch, striving to outperform a baseline on a simplified CIFAR-10 image classification task. This means instead of training a fixed neural network, miners are actually searching for better algorithms automatically. The network coordinates this search at scale: miners continuously generate and test algorithmic “children” locally, and whenever someone’s algorithm scores above the current best (on a validation set), they submit it for verification. Validators then re-run that model on hidden test data to check the claim. If the improvement is confirmed, Bitsota declares a new state-of-the-art (SoTA) and immediately releases a reward to the winning miner. If not, the claim is rejected and no reward is given, ensuring that only genuine progress gets rewarded.
Crucially, Bitsota is built on the Bittensor blockchain (layer-1), which provides the token incentive structure (the $TAO token emissions) and secure, decentralized governance for the subnet. Bitsota leverages Bittensor’s dynamic tokenomics so that its miners and validators are rewarded in proportion to the verified value they add. In Bitsota’s case, the “value” is improved AI performance, and the network’s rules are tuned to favor breakthroughs: for each block cycle, ~90% of the allotted mining reward is burned (or remains unspent) unless a new SoTA is found, and up to 10% is allocated to the winner of a confirmed breakthrough. This “results-gated” reward mechanism means Bitsota isn’t constantly paying out tokens unless real progress is happening – essentially “buying breakthroughs, not compute”. Over time, this creates a competitive dynamic: miners are incentivized to innovate quickly (to capture the prize before others), and problem owners get solutions only when milestones are met, paying only for success rather than effort. The end result is a provably efficient research marketplace: by aligning economic rewards with research outcomes, Bitsota aims to crowdsource AI innovation at scale, turning wasted idle compute worldwide into a collective engine for finding better algorithms.
Bitsota’s product is essentially a full-stack decentralized research platform consisting of specialized software, smart-contract mechanisms, and integration with the Bittensor network. On a technical level, the Bitsota team has built a suite of components that enable this novel mining-and-validation process to run autonomously:
Desktop Mining Client (GUI): Bitsota provides a user-friendly one-click mining application for Windows, Mac, and Linux. This GUI app is the core product for miners – it bundles all the logic needed to participate in the subnet without requiring any coding or command-line usage. Upon installation, a miner simply inputs their Bittensor wallet key, chooses a mode (direct or pooled mining), and starts the process. The client then runs the AutoML genetic programming loop locally, generating and evaluating candidate algorithms each generation. Initially, Bitsota’s miner app runs on CPU only (no GPU required) for accessibility, but it is expected to incorporate GPU support as the project evolves (to accelerate training). The key feature is ease-of-use: the “Runs on Bittensor L1” backend is abstracted away, so miners anywhere can contribute compute in seconds. This lowers the entry barrier dramatically compared to traditional Bittensor mining, aligning with Bitsota’s mission of broad open participation.
Mining algorithm & pipeline: Under the hood, Bitsota’s miners implement a variant of AutoML-Zero genetic search. Each miner’s client interacts with a fixed evaluation pipeline (currently a CIFAR-10 binary classification task). The miner randomly initializes simple programs (made of basic math operations) and evolves them through genetic operations (mutation, crossover), seeking an algorithm that improves accuracy on the training data. The pipeline uses a defined metric (e.g. classification accuracy) and a baseline score to beat. Notably, miners only submit a model for network-wide validation if it’s near or above the current best score (SoTA) – this avoids spamming validators with mediocre models. This design essentially creates a global race: many miners in parallel explore the space of algorithms, but only when someone finds a potential breakthrough does it get broadcast for verification. The “miner” software includes logic to interface with Bittensor’s substrate chain (through a relay node) to submit these candidate models to the validators when triggered.
Validator nodes and quorum: The “validator” software is another critical part of Bitsota’s build. Validators are nodes run by independent participants (they can be community members who stake TAO and want to earn by verifying work). Bitsota’s codebase provides a standardized validator node program which automatically listens for new model submissions on the subnet, downloads the submitted model, and re-evaluates it on the secret test dataset. Multiple validators do this in parallel (forming an N-of-M quorum) and cross-check results. Bitsota uses a multi-signature voting mechanism (nicknamed Yuma consensus) where validators collectively decide whether the submission is a valid win. If a sufficient quorum agrees the model beats the target on the hidden tests, they then execute the reward payout via blockchain. In practice, validators achieve this by setting on-chain “weights” in Bittensor’s consensus – effectively casting votes that direct the token emission to the winning miner’s account. Bitsota’s code specifies that upon a confirmed win, 90% of that cycle’s TAO emission is burned and 10% is assigned to the winner (by updating Bittensor’s weight matrix). This on-chain vote acts like a smart contract ensuring the outcome is tamper-proof: either enough validators sign off and the reward is released, or if consensus isn’t reached (or the model fails the tests), the reward “capacitor” remains locked and rolls over. (A “capacitor” contract in Bitsota parlance is essentially the pooled prize pot that accumulates until a valid breakthrough occurs.) Each successful validation triggers an “audit receipt” – a publicly verifiable record (including the achieved score and validator signatures) that anyone can inspect to confirm that the new model indeed outperformed the old one.
Problem definition and challenge management: Bitsota is designed to be problem-agnostic and extensible. The first challenge (Track 1) is AutoML Zero on CIFAR-10, as described, and it serves as a “flagship benchmark”. The platform architecture, however, allows the addition of new tracks or problem statements. A problem owner can define a challenge by specifying the dataset, evaluation metric, and baseline model performance. This definition is then distributed to all miners (so they know what to train for) and to validators (so they know how to test submissions). The Bitsota website front-end includes a Leaderboard page that displays the current top score and recent winners for each track. It shows, for example, the “Global Top SoTA” and previous reward amount for Track 1, as well as a ranking of miners (by wallet address) who have achieved wins, with timestamps and scores. This live leaderboard is part of the product, providing transparency and competition – miners can see what the current target is and how long a particular model has held the top spot.
Integration with Bittensor L1: As a Bittensor subnet, Bitsota’s entire infrastructure is built on and registered within the Bittensor blockchain. This means Bitsota has its own subnet “alpha” token (governed by the Bittensor dTAO mechanism) and relies on Bittensor’s consensus for security. The Bitsota software (miners and validators) uses Bittensor’s SDK to communicate with the chain – for example, miners use a Bittensor hotkey (keypair) to authenticate and submit work, and validators stake TAO and update weights on-chain via Bittensor’s substrate calls. From a user perspective, this is mostly behind-the-scenes, but it ensures that rewards and validations are immutable and decentralized, not controlled by any single server. Bitsota’s smart-contract-like behavior (the capacitor logic) is implemented through this on-chain weight setting and a Solidity contract in the repository (likely for managing certain reward conditions). All token transfers (emissions, burns, payouts) occur on-chain, and anyone can inspect the blockchain or use tools like TaoStats to see Bitsota’s performance metrics (e.g. how many TAO have been allocated or burned in Subnet 94).
In summary, the “build” of Bitsota consists of the open-source codebase (hosted on GitHub under AlveusLabs/SN94-BitSota), which includes the miner GUI, the validation node program, and auxiliary scripts/contracts. The Bitsota platform also encompasses the web interface (bitsota.ai) for presenting challenges and stats, and the community infrastructure (Discord, etc.) for coordination. Technically, Bitsota is pushing the boundaries of decentralized AI by combining several innovations: genetic AutoML algorithms, distributed computing, hidden test validation, and blockchain incentives. All these pieces work together so that any researcher or enthusiast can contribute to AI research (by running the software) and be fairly rewarded in real time for any measurable progress they contribute. The product is therefore both a research service (for those who want AI problems solved) and a mining application (for those who want to earn by contributing compute and creativity), with the Bittensor network providing the backbone.
Bitsota’s product is essentially a full-stack decentralized research platform consisting of specialized software, smart-contract mechanisms, and integration with the Bittensor network. On a technical level, the Bitsota team has built a suite of components that enable this novel mining-and-validation process to run autonomously:
Desktop Mining Client (GUI): Bitsota provides a user-friendly one-click mining application for Windows, Mac, and Linux. This GUI app is the core product for miners – it bundles all the logic needed to participate in the subnet without requiring any coding or command-line usage. Upon installation, a miner simply inputs their Bittensor wallet key, chooses a mode (direct or pooled mining), and starts the process. The client then runs the AutoML genetic programming loop locally, generating and evaluating candidate algorithms each generation. Initially, Bitsota’s miner app runs on CPU only (no GPU required) for accessibility, but it is expected to incorporate GPU support as the project evolves (to accelerate training). The key feature is ease-of-use: the “Runs on Bittensor L1” backend is abstracted away, so miners anywhere can contribute compute in seconds. This lowers the entry barrier dramatically compared to traditional Bittensor mining, aligning with Bitsota’s mission of broad open participation.
Mining algorithm & pipeline: Under the hood, Bitsota’s miners implement a variant of AutoML-Zero genetic search. Each miner’s client interacts with a fixed evaluation pipeline (currently a CIFAR-10 binary classification task). The miner randomly initializes simple programs (made of basic math operations) and evolves them through genetic operations (mutation, crossover), seeking an algorithm that improves accuracy on the training data. The pipeline uses a defined metric (e.g. classification accuracy) and a baseline score to beat. Notably, miners only submit a model for network-wide validation if it’s near or above the current best score (SoTA) – this avoids spamming validators with mediocre models. This design essentially creates a global race: many miners in parallel explore the space of algorithms, but only when someone finds a potential breakthrough does it get broadcast for verification. The “miner” software includes logic to interface with Bittensor’s substrate chain (through a relay node) to submit these candidate models to the validators when triggered.
Validator nodes and quorum: The “validator” software is another critical part of Bitsota’s build. Validators are nodes run by independent participants (they can be community members who stake TAO and want to earn by verifying work). Bitsota’s codebase provides a standardized validator node program which automatically listens for new model submissions on the subnet, downloads the submitted model, and re-evaluates it on the secret test dataset. Multiple validators do this in parallel (forming an N-of-M quorum) and cross-check results. Bitsota uses a multi-signature voting mechanism (nicknamed Yuma consensus) where validators collectively decide whether the submission is a valid win. If a sufficient quorum agrees the model beats the target on the hidden tests, they then execute the reward payout via blockchain. In practice, validators achieve this by setting on-chain “weights” in Bittensor’s consensus – effectively casting votes that direct the token emission to the winning miner’s account. Bitsota’s code specifies that upon a confirmed win, 90% of that cycle’s TAO emission is burned and 10% is assigned to the winner (by updating Bittensor’s weight matrix). This on-chain vote acts like a smart contract ensuring the outcome is tamper-proof: either enough validators sign off and the reward is released, or if consensus isn’t reached (or the model fails the tests), the reward “capacitor” remains locked and rolls over. (A “capacitor” contract in Bitsota parlance is essentially the pooled prize pot that accumulates until a valid breakthrough occurs.) Each successful validation triggers an “audit receipt” – a publicly verifiable record (including the achieved score and validator signatures) that anyone can inspect to confirm that the new model indeed outperformed the old one.
Problem definition and challenge management: Bitsota is designed to be problem-agnostic and extensible. The first challenge (Track 1) is AutoML Zero on CIFAR-10, as described, and it serves as a “flagship benchmark”. The platform architecture, however, allows the addition of new tracks or problem statements. A problem owner can define a challenge by specifying the dataset, evaluation metric, and baseline model performance. This definition is then distributed to all miners (so they know what to train for) and to validators (so they know how to test submissions). The Bitsota website front-end includes a Leaderboard page that displays the current top score and recent winners for each track. It shows, for example, the “Global Top SoTA” and previous reward amount for Track 1, as well as a ranking of miners (by wallet address) who have achieved wins, with timestamps and scores. This live leaderboard is part of the product, providing transparency and competition – miners can see what the current target is and how long a particular model has held the top spot.
Integration with Bittensor L1: As a Bittensor subnet, Bitsota’s entire infrastructure is built on and registered within the Bittensor blockchain. This means Bitsota has its own subnet “alpha” token (governed by the Bittensor dTAO mechanism) and relies on Bittensor’s consensus for security. The Bitsota software (miners and validators) uses Bittensor’s SDK to communicate with the chain – for example, miners use a Bittensor hotkey (keypair) to authenticate and submit work, and validators stake TAO and update weights on-chain via Bittensor’s substrate calls. From a user perspective, this is mostly behind-the-scenes, but it ensures that rewards and validations are immutable and decentralized, not controlled by any single server. Bitsota’s smart-contract-like behavior (the capacitor logic) is implemented through this on-chain weight setting and a Solidity contract in the repository (likely for managing certain reward conditions). All token transfers (emissions, burns, payouts) occur on-chain, and anyone can inspect the blockchain or use tools like TaoStats to see Bitsota’s performance metrics (e.g. how many TAO have been allocated or burned in Subnet 94).
In summary, the “build” of Bitsota consists of the open-source codebase (hosted on GitHub under AlveusLabs/SN94-BitSota), which includes the miner GUI, the validation node program, and auxiliary scripts/contracts. The Bitsota platform also encompasses the web interface (bitsota.ai) for presenting challenges and stats, and the community infrastructure (Discord, etc.) for coordination. Technically, Bitsota is pushing the boundaries of decentralized AI by combining several innovations: genetic AutoML algorithms, distributed computing, hidden test validation, and blockchain incentives. All these pieces work together so that any researcher or enthusiast can contribute to AI research (by running the software) and be fairly rewarded in real time for any measurable progress they contribute. The product is therefore both a research service (for those who want AI problems solved) and a mining application (for those who want to earn by contributing compute and creativity), with the Bittensor network providing the backbone.
Bitsota is being developed by an independent team operating under the name Alveus Labs. The project’s code repository is maintained by the AlveusLabs organization on GitHub, and the team launched the Bitsota Twitter/X account in February 2025, indicating that development ramped up around early 2025. So far, the individual team members or founders have not been publicly identified on official channels – the website and documentation focus on the product and technology rather than the people. In the spirit of many crypto projects, the emphasis is on the open-source community effort. The Bitsota developers engage with the community through official channels like the Bitsota Discord server and X (Twitter). For example, the team has been active on Discord, inviting early participants to join and collaborate – “this is where we coordinate, share progress and bring our first operators together. Join us and plug into what we’re building”. This suggests a community-driven approach where early miners, validators, and researchers are working directly with the Bitsota team to refine the subnet.
While we don’t have names and titles, we do know that Alveus Labs is behind Subnet 94 (Bitsota), and that the project is closely aligned with the Bittensor ecosystem but not run by the core Opentensor Foundation (it’s a third-party subnet initiative). The team appears to be composed of individuals with expertise in machine learning research and blockchain development, given the complexity of implementing AutoML and smart contract mechanics together. Bitsota’s public communications (tweets and forum posts) often come from the voice of the project or organization rather than a single person, reinforcing that it’s a group effort. They have also signaled openness to community contribution: the GitHub explicitly welcomes bug reports, feature requests, and code contributions via pull requests, and encourages discussion on Discord. This means that, much like other open-source projects, the “team” includes not only the core developers at Alveus Labs but potentially community developers and Bittensor enthusiasts who contribute to improving the subnet.
In summary, the Bitsota team remains somewhat behind the scenes, preferring to highlight the platform’s capabilities. They operate under the pseudonym Alveus Labs, and interact with the Bittensor community through social media and chat. As the project matures, more information about the team may emerge (for instance, through published research papers or conference presentations, the authors might be revealed). For now, the focus is on building a robust network and achieving technical milestones, with the team’s credibility resting on delivered results (a functioning product) rather than personal branding. Interested community members can get involved via the official Discord and the GitHub repo to become part of Bitsota’s development and validation process.
Bitsota is being developed by an independent team operating under the name Alveus Labs. The project’s code repository is maintained by the AlveusLabs organization on GitHub, and the team launched the Bitsota Twitter/X account in February 2025, indicating that development ramped up around early 2025. So far, the individual team members or founders have not been publicly identified on official channels – the website and documentation focus on the product and technology rather than the people. In the spirit of many crypto projects, the emphasis is on the open-source community effort. The Bitsota developers engage with the community through official channels like the Bitsota Discord server and X (Twitter). For example, the team has been active on Discord, inviting early participants to join and collaborate – “this is where we coordinate, share progress and bring our first operators together. Join us and plug into what we’re building”. This suggests a community-driven approach where early miners, validators, and researchers are working directly with the Bitsota team to refine the subnet.
While we don’t have names and titles, we do know that Alveus Labs is behind Subnet 94 (Bitsota), and that the project is closely aligned with the Bittensor ecosystem but not run by the core Opentensor Foundation (it’s a third-party subnet initiative). The team appears to be composed of individuals with expertise in machine learning research and blockchain development, given the complexity of implementing AutoML and smart contract mechanics together. Bitsota’s public communications (tweets and forum posts) often come from the voice of the project or organization rather than a single person, reinforcing that it’s a group effort. They have also signaled openness to community contribution: the GitHub explicitly welcomes bug reports, feature requests, and code contributions via pull requests, and encourages discussion on Discord. This means that, much like other open-source projects, the “team” includes not only the core developers at Alveus Labs but potentially community developers and Bittensor enthusiasts who contribute to improving the subnet.
In summary, the Bitsota team remains somewhat behind the scenes, preferring to highlight the platform’s capabilities. They operate under the pseudonym Alveus Labs, and interact with the Bittensor community through social media and chat. As the project matures, more information about the team may emerge (for instance, through published research papers or conference presentations, the authors might be revealed). For now, the focus is on building a robust network and achieving technical milestones, with the team’s credibility resting on delivered results (a functioning product) rather than personal branding. Interested community members can get involved via the official Discord and the GitHub repo to become part of Bitsota’s development and validation process.
Bitsota launched publicly in late 2025 (around the time of Bittensor’s first TAO halving) as a beta release, and it has an ambitious roadmap ahead. As a brand-new subnet in a pioneering field, many features and expansions are planned to fulfill its vision of a decentralized AutoML research hub. Based on available information from the team’s communications and the current state of the project, here are the key elements of Bitsota’s roadmap:
Transition from Beta to Full Launch: Initially, Bitsota is in a closed beta with limited slots for miners. Early access has been given out via invite codes and a waitlist. The roadmap includes opening the platform to the public, i.e. removing the cap on participants once the system is proven stable. Because Bitsota is designed for unlimited participation, we can expect a full release where anyone can download the miner and join without an invite. This will likely occur after the beta phase validates the network’s stability and security with a smaller group.
GPU Support and Performance Scaling: At launch, Bitsota’s mining client supports CPU-only mining (“no GPU required”) to maximize accessibility. A major upcoming milestone will be enabling GPU acceleration. Utilizing GPUs will significantly speed up the algorithm evolution process (allowing more complex models or faster training per generation). The team has hinted that GPU support is on the way, so miners with more powerful hardware can contribute and compete more effectively. This upgrade will involve optimizing the code for parallel computation and possibly expanding the types of models that can be evolved. In short, moving from CPU-bound computation to optional GPU usage is critical for scaling up the difficulty and complexity of research tasks in future tracks.
Additional Challenge Tracks (New Research Problems): Currently, Bitsota is focused on its flagship AutoML Zero challenge (Track 1) on a vision task. However, the platform is built to be problem-agnostic and accommodate multiple categories of AI problems. The roadmap likely includes introducing Track 2, Track 3, etc., each representing a new benchmark or domain. These could be different datasets (for example, NLP tasks, reinforcement learning challenges, audio processing, etc.) or different algorithm search spaces. The goal is to demonstrate that Bitsota can drive progress in various fields of AI, not just image classification. We might see, for instance, a language-model challenge where miners evolve architectures to improve on a text task. The Bitsota team has indicated that the platform enables optimization of diverse problems and is generally focused on “self-improving and self-generating AI” beyond just this first use-case. So, expanding the research scope is a key part of the future – AutoML Zero is just the beginning.
Enhanced Pool Mining and Collaboration: While Bitsota already supports a concept of pool mining (allowing miners to collaborate and share partial work), this feature may be further developed in the roadmap. In the current beta, emphasis has been on direct mining with the GUI. Going forward, we can expect the pool mining mode to become more accessible, perhaps through an integrated interface where miners with low-power machines can easily join a pool. The roadmap could include establishing official mining pools or improving the protocol for pools so that contributions and rewards are fairly split among participants. This will broaden participation by letting those who can’t individually beat SoTA still contribute pieces of the solution (for example, testing algorithms or evolving smaller mutations) and earn a share of rewards. Making collaborative mining robust and user-friendly is therefore an important step to harness “the wisdom of crowds” for faster convergence on tough problems.
Release of Research Outputs and Papers: The team has signaled that formal research publications are in progress and will be released as the project matures. On the roadmap is the publication of technical papers, benchmarks, and findings from Bitsota’s experiments. This includes documenting the performance of the evolved algorithms, analyzing the dynamics of decentralized AutoML, and perhaps open-sourcing the best discovered models. The Bitsota website’s Research section is currently a placeholder, set to be updated with these papers and detailed studies. We can anticipate that as new SoTA results are achieved on the platform, the team will publish these results (for example, “Bitsota achieved X% on CIFAR-10 with an evolved algorithm, surpassing previous known approaches”) and share methodology. In essence, Bitsota aims to contribute back to the wider AI research community by turning mining outcomes into publishable science.
Continuous Refinement of Incentives and Security: As a cutting-edge network, Bitsota will continually refine its incentive mechanisms and validation process. The roadmap likely involves close monitoring of how miners behave and how the token reward model impacts participation. The team will adjust parameters like the reward schedule, burn percentage, or quorum threshold if needed to keep the network healthy and fair. Security is also a focus – ensuring that the hidden test methodology remains robust (to prevent any potential exploits or leak of test data), and that the network is resilient to collusion or spam. These are ongoing goals rather than one-off milestones. In the near future, Bitsota will also experience its own Alpha token halving (as all subnets do approximately every 2–2.5 years, independent of the TAO halving). Preparation for that event – to ensure a smooth transition in reward dynamics – could be part of the long-term roadmap.
Overall, Bitsota’s roadmap is about scaling up on all fronts: more users (from beta to open access), more compute power (GPU integration), more challenges (beyond the initial one), and more academic output (proven AI breakthroughs). The team will likely announce specific roadmap updates via official channels; for instance, they might outline upcoming features in blog posts or community calls as the project hits new phases. Given the rapid innovation in the Bittensor ecosystem, Bitsota is positioned to evolve quickly. Each successful milestone – e.g., a new best-performing algorithm discovered or a new track launched – will validate the concept of “useful mining” on Bittensor and drive the project to attract further contributors. By 2026, we should expect Bitsota to move out of beta, have multiple AI research tracks running in parallel, and possibly uncover novel algorithms that merit attention beyond the Bittensor community. All progress will continue to be openly logged and incentivized on-chain, as Bitsota strives to become a sustained, self-improving AI research economy.
Bitsota launched publicly in late 2025 (around the time of Bittensor’s first TAO halving) as a beta release, and it has an ambitious roadmap ahead. As a brand-new subnet in a pioneering field, many features and expansions are planned to fulfill its vision of a decentralized AutoML research hub. Based on available information from the team’s communications and the current state of the project, here are the key elements of Bitsota’s roadmap:
Transition from Beta to Full Launch: Initially, Bitsota is in a closed beta with limited slots for miners. Early access has been given out via invite codes and a waitlist. The roadmap includes opening the platform to the public, i.e. removing the cap on participants once the system is proven stable. Because Bitsota is designed for unlimited participation, we can expect a full release where anyone can download the miner and join without an invite. This will likely occur after the beta phase validates the network’s stability and security with a smaller group.
GPU Support and Performance Scaling: At launch, Bitsota’s mining client supports CPU-only mining (“no GPU required”) to maximize accessibility. A major upcoming milestone will be enabling GPU acceleration. Utilizing GPUs will significantly speed up the algorithm evolution process (allowing more complex models or faster training per generation). The team has hinted that GPU support is on the way, so miners with more powerful hardware can contribute and compete more effectively. This upgrade will involve optimizing the code for parallel computation and possibly expanding the types of models that can be evolved. In short, moving from CPU-bound computation to optional GPU usage is critical for scaling up the difficulty and complexity of research tasks in future tracks.
Additional Challenge Tracks (New Research Problems): Currently, Bitsota is focused on its flagship AutoML Zero challenge (Track 1) on a vision task. However, the platform is built to be problem-agnostic and accommodate multiple categories of AI problems. The roadmap likely includes introducing Track 2, Track 3, etc., each representing a new benchmark or domain. These could be different datasets (for example, NLP tasks, reinforcement learning challenges, audio processing, etc.) or different algorithm search spaces. The goal is to demonstrate that Bitsota can drive progress in various fields of AI, not just image classification. We might see, for instance, a language-model challenge where miners evolve architectures to improve on a text task. The Bitsota team has indicated that the platform enables optimization of diverse problems and is generally focused on “self-improving and self-generating AI” beyond just this first use-case. So, expanding the research scope is a key part of the future – AutoML Zero is just the beginning.
Enhanced Pool Mining and Collaboration: While Bitsota already supports a concept of pool mining (allowing miners to collaborate and share partial work), this feature may be further developed in the roadmap. In the current beta, emphasis has been on direct mining with the GUI. Going forward, we can expect the pool mining mode to become more accessible, perhaps through an integrated interface where miners with low-power machines can easily join a pool. The roadmap could include establishing official mining pools or improving the protocol for pools so that contributions and rewards are fairly split among participants. This will broaden participation by letting those who can’t individually beat SoTA still contribute pieces of the solution (for example, testing algorithms or evolving smaller mutations) and earn a share of rewards. Making collaborative mining robust and user-friendly is therefore an important step to harness “the wisdom of crowds” for faster convergence on tough problems.
Release of Research Outputs and Papers: The team has signaled that formal research publications are in progress and will be released as the project matures. On the roadmap is the publication of technical papers, benchmarks, and findings from Bitsota’s experiments. This includes documenting the performance of the evolved algorithms, analyzing the dynamics of decentralized AutoML, and perhaps open-sourcing the best discovered models. The Bitsota website’s Research section is currently a placeholder, set to be updated with these papers and detailed studies. We can anticipate that as new SoTA results are achieved on the platform, the team will publish these results (for example, “Bitsota achieved X% on CIFAR-10 with an evolved algorithm, surpassing previous known approaches”) and share methodology. In essence, Bitsota aims to contribute back to the wider AI research community by turning mining outcomes into publishable science.
Continuous Refinement of Incentives and Security: As a cutting-edge network, Bitsota will continually refine its incentive mechanisms and validation process. The roadmap likely involves close monitoring of how miners behave and how the token reward model impacts participation. The team will adjust parameters like the reward schedule, burn percentage, or quorum threshold if needed to keep the network healthy and fair. Security is also a focus – ensuring that the hidden test methodology remains robust (to prevent any potential exploits or leak of test data), and that the network is resilient to collusion or spam. These are ongoing goals rather than one-off milestones. In the near future, Bitsota will also experience its own Alpha token halving (as all subnets do approximately every 2–2.5 years, independent of the TAO halving). Preparation for that event – to ensure a smooth transition in reward dynamics – could be part of the long-term roadmap.
Overall, Bitsota’s roadmap is about scaling up on all fronts: more users (from beta to open access), more compute power (GPU integration), more challenges (beyond the initial one), and more academic output (proven AI breakthroughs). The team will likely announce specific roadmap updates via official channels; for instance, they might outline upcoming features in blog posts or community calls as the project hits new phases. Given the rapid innovation in the Bittensor ecosystem, Bitsota is positioned to evolve quickly. Each successful milestone – e.g., a new best-performing algorithm discovered or a new track launched – will validate the concept of “useful mining” on Bittensor and drive the project to attract further contributors. By 2026, we should expect Bitsota to move out of beta, have multiple AI research tracks running in parallel, and possibly uncover novel algorithms that merit attention beyond the Bittensor community. All progress will continue to be openly logged and incentivized on-chain, as Bitsota strives to become a sustained, self-improving AI research economy.