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of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time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of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
CliqueAI is a specialized Bittensor subnet that harnesses decentralized AI to tackle instances of this hard combinatorial problem. In essence, CliqueAI turns the maximum clique search into a distributed challenge: many participant nodes (AI “miners”) cooperate and compete to find the largest clique in various graphs, while validator nodes score their solutions and update reputations on-chain.
CliqueAI’s network operates via a four-stage autonomous mechanism for distributed problem-solving:
Problem Selection: Advanced algorithms select and generate complex graph instances from a distributed database, categorizing each problem by difficulty, graph structure, and size. This ensures the network is continually challenged with varied graph problems rather than trivial cases.
Miner Selection: A smart allocation engine assigns graph problems to miners (participants running solving algorithms) based on each miner’s past performance and experience. A stake-weighted distribution is used so that miners with more stake and better history get appropriately sized tasks, optimizing resource use across the network.
Scoring: Submitted solutions (cliques found) are evaluated using a dual-metric system that considers both solution optimality (e.g. clique size correctness) and algorithmic diversity. This means miners are rewarded not just for finding a correct or large clique, but also for bringing innovative approaches. By valuing diverse methods, the network avoids everyone converging on the same simplistic strategy and encourages exploration of different solving techniques.
Weight Setting: Miner reputations (weights) are continually adjusted based on performance using an exponential moving average. In practice, if a miner consistently finds optimal or near-optimal cliques, its reputation weight on the subnet increases, influencing how often it gets selected for future problems and the rewards it earns. Conversely, poor or incorrect performance lowers a miner’s weight. This on-chain reputation system ensures that over time the network self-tunes to favor reliable, high-quality problem solvers.
Through this pipeline, CliqueAI effectively crowdsources the solving of maximum clique problems to a decentralized AI swarm. Each cycle, the subnet poses a graph problem, miners attempt solutions, validators verify and score the results, and the blockchain (Subtensor) updates miner weights and distributes rewards accordingly. By iterating in this way, the network continuously produces solutions to graph clique challenges while ranking and rewarding the best performing AI models.
CliqueAI is a specialized Bittensor subnet that harnesses decentralized AI to tackle instances of this hard combinatorial problem. In essence, CliqueAI turns the maximum clique search into a distributed challenge: many participant nodes (AI “miners”) cooperate and compete to find the largest clique in various graphs, while validator nodes score their solutions and update reputations on-chain.
CliqueAI’s network operates via a four-stage autonomous mechanism for distributed problem-solving:
Problem Selection: Advanced algorithms select and generate complex graph instances from a distributed database, categorizing each problem by difficulty, graph structure, and size. This ensures the network is continually challenged with varied graph problems rather than trivial cases.
Miner Selection: A smart allocation engine assigns graph problems to miners (participants running solving algorithms) based on each miner’s past performance and experience. A stake-weighted distribution is used so that miners with more stake and better history get appropriately sized tasks, optimizing resource use across the network.
Scoring: Submitted solutions (cliques found) are evaluated using a dual-metric system that considers both solution optimality (e.g. clique size correctness) and algorithmic diversity. This means miners are rewarded not just for finding a correct or large clique, but also for bringing innovative approaches. By valuing diverse methods, the network avoids everyone converging on the same simplistic strategy and encourages exploration of different solving techniques.
Weight Setting: Miner reputations (weights) are continually adjusted based on performance using an exponential moving average. In practice, if a miner consistently finds optimal or near-optimal cliques, its reputation weight on the subnet increases, influencing how often it gets selected for future problems and the rewards it earns. Conversely, poor or incorrect performance lowers a miner’s weight. This on-chain reputation system ensures that over time the network self-tunes to favor reliable, high-quality problem solvers.
Through this pipeline, CliqueAI effectively crowdsources the solving of maximum clique problems to a decentralized AI swarm. Each cycle, the subnet poses a graph problem, miners attempt solutions, validators verify and score the results, and the blockchain (Subtensor) updates miner weights and distributes rewards accordingly. By iterating in this way, the network continuously produces solutions to graph clique challenges while ranking and rewarding the best performing AI models.
CliqueAI is essentially a decentralized solving service for maximum clique problems, built on the Bittensor blockchain. The “product” is the open-source subnet code and network that collectively finds maximum cliques in graphs. On a technical level, CliqueAI is implemented as a custom Bittensor subnet (network ID 83 on the Finney chain) with its own economy and token (often called an “Alpha” token) that is interchangeable with Bittensor’s main TAO token via the Bittensor DEX. The code repository (toptensor/CliqueAI on GitHub) is about 99% Python, plus some shell scripts, and includes tailored client logic for miners and validators. Participants can run a miner client or validator client by executing the provided scripts with their wallet credentials and pointing to subnet 83 on the network. Upon running a miner, the node connects to the CliqueAI subnet and awaits graph problems to solve; running a validator node allows one to help evaluate solutions and maintain the subnet’s consensus.
Technical architecture: CliqueAI follows the standard Bittensor two-tier node design: a fixed number of validators (in CliqueAI’s case, 6 validators at capacity) maintain the subnet’s consensus and distribute tasks, while up to 256 miners perform the AI workload of solving graph problems (these were the initial network limits). The subnet’s logic extends Bittensor’s base protocol to handle graph data. For each task, a validator likely provides the graph (or a seed for it) to the selected miners. Miners then run algorithms (e.g. exact or heuristic clique-finding algorithms, potentially enhanced with machine learning heuristics) to find a large clique. The miner returns the candidate clique (likely specified by the vertices included). Validators cross-verify these solutions – checking clique validity is straightforward, though determining optimality may involve comparing sizes of cliques found by different miners or timeouts. The dual-metric scoring ensures that if multiple miners all find the maximal clique, those employing unique methods might score slightly higher to encourage a diversity of approaches.
Under the hood, CliqueAI uses Bittensor’s substrate-based chain (Subtensor) to record weights, stakes, and incentives for nodes. The stake-weight mechanism means miners and validators have staked TAO (or the subnet’s alpha tokens) which influences their standing. Rewards in the form of newly emitted subnet tokens (α) are given to miners based on their weight and scoring performance each block. CliqueAI’s subnet is allocated a small percentage of Bittensor’s overall token inflation (on the order of ~0.5% of network TAO issuance) to reward its participants. This percentage can adjust over time, but as of its launch it was around half a percent of emissions (indicative of the subnet’s share in the Bittensor ecosystem).
Another important part of the build is monitoring and logging. CliqueAI integrates with Weights & Biases (W&B) for tracking performance metrics and debugging. The project’s documentation describes a data logging schema that miners/validators use to record metrics to W&B dashboards. This allows the team and community to analyze how the network is performing (e.g. how quickly cliques are found, variance in miner performance, etc.) over time. Such insight is valuable for optimizing the algorithms and parameters of the network.
In summary, the “product” is not a single end-user application but a decentralized AI network service. Potentially, other developers or systems could query CliqueAI to solve graph problems (for example, using Bittensor’s RPC mechanism to input a specific graph and retrieve a solution), though at present CliqueAI’s problems seem to be largely internally generated to drive the network’s training and evaluation cycle. The long-term vision is that this subnet could serve as an on-demand solver for maximum clique queries or advance research into solving NP-hard graph problems via collective AI. For now, its build is that of an autonomous, self-contained AI subnet continually solving random graph challenges and refining its pool of solvers through blockchain-based incentives.
CliqueAI is essentially a decentralized solving service for maximum clique problems, built on the Bittensor blockchain. The “product” is the open-source subnet code and network that collectively finds maximum cliques in graphs. On a technical level, CliqueAI is implemented as a custom Bittensor subnet (network ID 83 on the Finney chain) with its own economy and token (often called an “Alpha” token) that is interchangeable with Bittensor’s main TAO token via the Bittensor DEX. The code repository (toptensor/CliqueAI on GitHub) is about 99% Python, plus some shell scripts, and includes tailored client logic for miners and validators. Participants can run a miner client or validator client by executing the provided scripts with their wallet credentials and pointing to subnet 83 on the network. Upon running a miner, the node connects to the CliqueAI subnet and awaits graph problems to solve; running a validator node allows one to help evaluate solutions and maintain the subnet’s consensus.
Technical architecture: CliqueAI follows the standard Bittensor two-tier node design: a fixed number of validators (in CliqueAI’s case, 6 validators at capacity) maintain the subnet’s consensus and distribute tasks, while up to 256 miners perform the AI workload of solving graph problems (these were the initial network limits). The subnet’s logic extends Bittensor’s base protocol to handle graph data. For each task, a validator likely provides the graph (or a seed for it) to the selected miners. Miners then run algorithms (e.g. exact or heuristic clique-finding algorithms, potentially enhanced with machine learning heuristics) to find a large clique. The miner returns the candidate clique (likely specified by the vertices included). Validators cross-verify these solutions – checking clique validity is straightforward, though determining optimality may involve comparing sizes of cliques found by different miners or timeouts. The dual-metric scoring ensures that if multiple miners all find the maximal clique, those employing unique methods might score slightly higher to encourage a diversity of approaches.
Under the hood, CliqueAI uses Bittensor’s substrate-based chain (Subtensor) to record weights, stakes, and incentives for nodes. The stake-weight mechanism means miners and validators have staked TAO (or the subnet’s alpha tokens) which influences their standing. Rewards in the form of newly emitted subnet tokens (α) are given to miners based on their weight and scoring performance each block. CliqueAI’s subnet is allocated a small percentage of Bittensor’s overall token inflation (on the order of ~0.5% of network TAO issuance) to reward its participants. This percentage can adjust over time, but as of its launch it was around half a percent of emissions (indicative of the subnet’s share in the Bittensor ecosystem).
Another important part of the build is monitoring and logging. CliqueAI integrates with Weights & Biases (W&B) for tracking performance metrics and debugging. The project’s documentation describes a data logging schema that miners/validators use to record metrics to W&B dashboards. This allows the team and community to analyze how the network is performing (e.g. how quickly cliques are found, variance in miner performance, etc.) over time. Such insight is valuable for optimizing the algorithms and parameters of the network.
In summary, the “product” is not a single end-user application but a decentralized AI network service. Potentially, other developers or systems could query CliqueAI to solve graph problems (for example, using Bittensor’s RPC mechanism to input a specific graph and retrieve a solution), though at present CliqueAI’s problems seem to be largely internally generated to drive the network’s training and evaluation cycle. The long-term vision is that this subnet could serve as an on-demand solver for maximum clique queries or advance research into solving NP-hard graph problems via collective AI. For now, its build is that of an autonomous, self-contained AI subnet continually solving random graph challenges and refining its pool of solvers through blockchain-based incentives.
Team: The subnet was created and is maintained by a team/organization known as TopTensor. On the Bittensor Subnet directory, TopTensor is listed as the team behind CliqueAI, and the project’s official links include a website (cliqueai.toptensor.ai) and the GitHub repository under the TopTensor account. The individual members of TopTensor have not been publicly detailed – the GitHub org does not reveal member profiles and no specific names are given in announcements. All code commits to the project were made by the toptensor-deployer account on GitHub, suggesting an internal account used for development and deployment. This indicates that TopTensor operates somewhat opaquely or at least collectively rather than highlighting individual developers.
From context, TopTensor appears to be a focused group of AI/crypto researchers and engineers working on Bittensor subnets. The TopTensor team’s expertise is likely in graph algorithms and distributed AI, given that CliqueAI’s core task (maximum clique finding) is a classic hard problem at the intersection of computer science and optimization. It’s worth noting that Bittensor’s ecosystem has multiple teams each building specialized subnets (for example, some teams focus on NLP, others on computer vision or biotech data, etc.), and TopTensor’s CliqueAI stands out as a subnet aimed at a graph-theoretic challenge. The team’s choice of problem suggests a research interest in applying decentralized AI to combinatorial optimization. While not much biographical information is available, the TopTensor team is responsible for maintaining the CliqueAI network parameters, releasing software updates, and community engagement for Subnet 83. They likely coordinate via Bittensor’s community channels (Discord, etc.) under the TopTensor or CliqueAI moniker, and they have provided the open-source code and documentation as primary resources for the community.
Team: The subnet was created and is maintained by a team/organization known as TopTensor. On the Bittensor Subnet directory, TopTensor is listed as the team behind CliqueAI, and the project’s official links include a website (cliqueai.toptensor.ai) and the GitHub repository under the TopTensor account. The individual members of TopTensor have not been publicly detailed – the GitHub org does not reveal member profiles and no specific names are given in announcements. All code commits to the project were made by the toptensor-deployer account on GitHub, suggesting an internal account used for development and deployment. This indicates that TopTensor operates somewhat opaquely or at least collectively rather than highlighting individual developers.
From context, TopTensor appears to be a focused group of AI/crypto researchers and engineers working on Bittensor subnets. The TopTensor team’s expertise is likely in graph algorithms and distributed AI, given that CliqueAI’s core task (maximum clique finding) is a classic hard problem at the intersection of computer science and optimization. It’s worth noting that Bittensor’s ecosystem has multiple teams each building specialized subnets (for example, some teams focus on NLP, others on computer vision or biotech data, etc.), and TopTensor’s CliqueAI stands out as a subnet aimed at a graph-theoretic challenge. The team’s choice of problem suggests a research interest in applying decentralized AI to combinatorial optimization. While not much biographical information is available, the TopTensor team is responsible for maintaining the CliqueAI network parameters, releasing software updates, and community engagement for Subnet 83. They likely coordinate via Bittensor’s community channels (Discord, etc.) under the TopTensor or CliqueAI moniker, and they have provided the open-source code and documentation as primary resources for the community.
As a relatively new project (CliqueAI went live in late August 2025), the roadmap is not extensively public. The subnet’s profile on official channels is marked “Coverage: Pending”, indicating that detailed plans and documentation are still being developed. However, we can glean some milestones and likely future directions from the available information:
Launch and Early Updates (Q3 2025): CliqueAI was activated on the Bittensor network around August 27, 2025. The initial release was version 0.0.0/0.0.1, with rapid follow-up improvements. In the first few weeks, the team released incremental updates (v0.0.2 and v0.0.3 by early September 2025) focusing on stability and bug fixes. These updates (e.g. fixing weight update timing, import issues, etc.) suggest the team was fine-tuning the core mechanism after observing the subnet’s behavior in production.
Current Focus: Now that the subnet is operational with a full complement of miners and validators, the immediate goal is likely ensuring robust performance on the chosen problem set. This involves refining how graph problems are generated (to gradually increase difficulty as miners improve), and possibly enhancing the solution algorithms that miners use. The dual-metric scoring may also be adjusted as needed to balance rewarding accuracy vs. diversity. The team will be monitoring metrics via W&B and on-chain data to identify bottlenecks (for example, if miners struggle beyond a certain graph size or if the network converges too slowly on correct answers).
Potential Upcoming Features: While no official feature roadmap has been published, possible future enhancements could include:
Long Term Vision: CliqueAI is pioneering the idea of using a decentralized swarm to solve an NP-hard problem. If successful, it could open the door to subnets tackling other hard optimization problems (graph coloring, SAT solving, etc.). In the long term, the roadmap might involve generalizing the approach or linking with academic research. For example, the team might collaborate with researchers to compare CliqueAI’s performance against classical algorithms or even quantum approaches on clique problems. The ultimate goal is for CliqueAI to consistently solve very large graph instances that are beyond the capability of single machines, proving the power of collective AI on Bittensor.
It’s important to note that no formal roadmap document has been publicly released by TopTensor as of now (no dates or specific version targets beyond the initial launch). The development seems to be iterative and adaptive, based on what is learned from each version’s performance. The community can track progress through the GitHub commits and version releases, as well as any announcements in Bittensor forums. Given the early stage of Subnet 83, the rest of 2025 will likely be about achieving stability and demonstrable problem-solving capability. Moving into 2026, we anticipate more clarity on CliqueAI’s roadmap once the subnet has proven its concept and the team solidifies plans for scaling up its ambitions. In summary, the roadmap for CliqueAI is in its infancy – the project’s trajectory will become clearer as TopTensor publishes more information, but it is poised to evolve from a working prototype into a powerful decentralized solver network for complex graph problems.
As a relatively new project (CliqueAI went live in late August 2025), the roadmap is not extensively public. The subnet’s profile on official channels is marked “Coverage: Pending”, indicating that detailed plans and documentation are still being developed. However, we can glean some milestones and likely future directions from the available information:
Launch and Early Updates (Q3 2025): CliqueAI was activated on the Bittensor network around August 27, 2025. The initial release was version 0.0.0/0.0.1, with rapid follow-up improvements. In the first few weeks, the team released incremental updates (v0.0.2 and v0.0.3 by early September 2025) focusing on stability and bug fixes. These updates (e.g. fixing weight update timing, import issues, etc.) suggest the team was fine-tuning the core mechanism after observing the subnet’s behavior in production.
Current Focus: Now that the subnet is operational with a full complement of miners and validators, the immediate goal is likely ensuring robust performance on the chosen problem set. This involves refining how graph problems are generated (to gradually increase difficulty as miners improve), and possibly enhancing the solution algorithms that miners use. The dual-metric scoring may also be adjusted as needed to balance rewarding accuracy vs. diversity. The team will be monitoring metrics via W&B and on-chain data to identify bottlenecks (for example, if miners struggle beyond a certain graph size or if the network converges too slowly on correct answers).
Potential Upcoming Features: While no official feature roadmap has been published, possible future enhancements could include:
Long Term Vision: CliqueAI is pioneering the idea of using a decentralized swarm to solve an NP-hard problem. If successful, it could open the door to subnets tackling other hard optimization problems (graph coloring, SAT solving, etc.). In the long term, the roadmap might involve generalizing the approach or linking with academic research. For example, the team might collaborate with researchers to compare CliqueAI’s performance against classical algorithms or even quantum approaches on clique problems. The ultimate goal is for CliqueAI to consistently solve very large graph instances that are beyond the capability of single machines, proving the power of collective AI on Bittensor.
It’s important to note that no formal roadmap document has been publicly released by TopTensor as of now (no dates or specific version targets beyond the initial launch). The development seems to be iterative and adaptive, based on what is learned from each version’s performance. The community can track progress through the GitHub commits and version releases, as well as any announcements in Bittensor forums. Given the early stage of Subnet 83, the rest of 2025 will likely be about achieving stability and demonstrable problem-solving capability. Moving into 2026, we anticipate more clarity on CliqueAI’s roadmap once the subnet has proven its concept and the team solidifies plans for scaling up its ambitions. In summary, the roadmap for CliqueAI is in its infancy – the project’s trajectory will become clearer as TopTensor publishes more information, but it is poised to evolve from a working prototype into a powerful decentralized solver network for complex graph problems.