Hone is not a single AI model but a decentralized network of contributors working toward this shared AI. It leverages the Bittensor blockchain framework to crowdsource the training: no single entity owns or controls the resulting model. Instead, many independent miners (participants) around the world will run training nodes, each contributing to improving the model, and independent validators will evaluate those contributions. This open design ensures transparency and broad participation. In practical terms, the subnet provides incentives (in the form of token rewards) to those who train models that perform better on the agreed tasks, thereby aligning everyone’s efforts toward the common goal of a powerful reasoning AI. Over time, as miners feed in diverse data and strategies and validators rigorously test the AI on benchmark problems, the best techniques and models “win” more rewards, pushing the whole system’s performance upward. This dynamic is central to Bittensor’s consensus: models are continuously evaluated, ranked, and rewarded, which “ensures continuous improvement and competitiveness” in the subnet.

The ultimate objective of Hone is extremely ambitious: achieve an AI capable of human-level general intelligence on non-trivial tasks. Concretely, the team has set a milestone of reaching ~85% accuracy on ARC-AGI-2 – a level that would essentially signify superhuman performance on those abstract reasoning tasks (for context, humans are believed to score around that range). Surpassing that 85% threshold on ARC-AGI-2 is viewed as “a major leap toward genuine general intelligence.”In other words, if Hone’s decentralized training can produce a model that solves 80–85% of these formerly intractable problems, it would be a watershed moment on the path to AGI. This is a long-term vision, but it’s tangible and measurable – by focusing on a specific hard benchmark, Hone can quantitatively track progress toward AGI.

Equally important, Hone is pursuing this in a manner that adheres to open collaboration and safety. All development is open-source, and the resulting models would belong to the community rather than a private company. By engaging a global network of researchers and enthusiasts, Hone’s creators believe they can “guide open-source AGI research towards machines that can learn, reason, and plan as humans do”, closing the gap on tasks that are “easy for humans, hard for machines.” Hone “Subnet 5” is like a decentralized research lab: its purpose is to collectively build an AI that thinks more like a human, advancing the state of the art in general problem-solving. Supporters even describe Hone’s emergence as ushering in a new era of permissionless, open and incentivized artificial superintelligence – reflecting the belief that this project could accelerate AI toward genuinely human-like (and eventually super-human) intelligence, all while being accessible to the community rather than locked behind closed doors.

Hone is not a traditional consumer-facing product – it’s more of an AI platform or infrastructure build – but it does have a clear “output” and architecture. The “product” of Hone will ultimately be the AI models and services it produces: a hierarchically-trained general reasoning model that others can use. In Bittensor terms, Hone is Subnet-5 (codenamed epsilon) dedicated to Hierarchical AI Pretraining. This means the subnet provides the environment and tools for training the AI (the pretraining stage), and once sufficiently trained, that AI can be accessed or deployed. Eventually, Hone’s model is expected to be served via the Bittensor network’s API so that developers or applications can query it (for example, to solve reasoning problems or perform complex analyses) just like calling an OpenAI API – except in this case it’s open and decentralized. The Hone team has indicated that the developer experience of using the subnet’s AI will be analogous to using an OpenAI service, but provided through Bittensor’s validator API. In short, the build is creating a decentralized “AGI-as-a-service”: the community trains a powerful model and anyone can eventually utilize its intelligence via the network.

From a technical architecture perspective, Hone’s build centers on implementing the novel hierarchical learning techniques into the Bittensor framework. There are two major research components being fused in Hone’s design:

1) A World-Model Learning Module (self-supervised predictive learning): This draws from Yann LeCun’s proposed JEPA (Joint Embedding Predictive Architecture) and in particular Hierarchical JEPA (H-JEPA). In practice, this means part of Hone’s training focuses on unsupervised prediction tasks – teaching the AI to understand the world by predicting missing information. For example, the model might observe part of a pattern or scenario and try to predict the rest. This helps it develop robust internal representations (a “world model”) rather than relying purely on surface-level correlations. Hierarchical predictive learning implies the model may learn representations at multiple levels of abstraction (hints of high-level concepts and low-level details) and use those to anticipate outcomes. By incorporating H-JEPA concepts, Hone tries to ensure the AI has a solid foundation of common-sense understanding about its environment or tasks, which is crucial for generalization.

2) A Hierarchical Reasoning Module (iterative problem solving): This is inspired by the Hierarchical Reasoning Model (HRM) introduced by Guan Wang et al. in 2025. The HRM is essentially a recurrent two-module architecture that mimics how a human might think through a problem in steps. One module (often called the “H” or high-level module) acts as a planner, setting out a coarse strategy or hypothesis for the solution. Another module (“L” or low-level module) acts as a worker, fleshing out details and executing the plan. The two then interact in an outer loop: the worker’s output is fed back, and the planner decides if the solution is good or needs another refinement cycle. This loop continues, effectively allowing the model to “think for multiple steps” and refine its answer until a certain halt condition is met (the model decides it’s confident enough). This approach contrasts with a single-pass transformer that produces an answer in one go. By implementing an HRM-like architecture, Hone’s build enables iterative reasoning – the AI can approach complex tasks by breaking them down into intermediate steps, re-planning as needed, much like a human solving a puzzle through trial and error. The “hierarchical” part of the name refers both to the two-level structure of the model and to the idea that the reasoning happens at multiple time-scales: a slow deliberate process guiding a fast detailed process. This design is believed to attain significant depth of reasoning without blowing up computational requirements, as it reuses two smaller networks in a loop instead of one enormous flat model.

The combination of these two – a world-model pretraining and a hierarchical reasoning architecture – forms the crux of Hone’s technical build. In practical terms, the team will be developing custom training objectives and pipelines to train these components on Bittensor. For example, miners might train the model on self-supervised tasks (à la JEPA) using large unlabeled datasets to build general knowledge, and simultaneously or subsequently train it on reasoning tasks (à la HRM) using generated puzzle examples or past ARC tasks. The validators would need to evaluate how well a miner’s model is performing on reasoning tasks (since that’s the ultimate metric of success). This likely involves setting up a suite of benchmark tasks (possibly small ARC-style puzzles or similar abstract problems) that the miners’ models must answer. Validators will compare the answers to known solutions or use an evaluation script (for ARC, whether the output grid is correct) and score the models accordingly. In the original OpenKaito setup (which was about embedding models), validators used a contrastive InfoNCE-based metric to judge embeddings; under Hone, the evaluation might be more straightforward – e.g. percentage of test problems solved correctly – or involve an AI-based judge if solutions are not binary right/wrong. The incentive mechanism will reward models that consistently solve more of the test tasks or solve them faster/more efficiently. This ensures that the network’s collective effort goes into improving real performance on the target tasks, not just proxy metrics.

To illustrate how the “product” might be experienced: once Hone’s model becomes sufficiently capable, end-users or developers could query it for its reasoning on new problems. For instance, someone might present a new ARC-like puzzle or a logical question to the network; the Hone model (distributed across miners) would process it through its hierarchical reasoning steps and produce an answer, which could then be returned via a Validator API endpoint. The validator nodes essentially act as gateways that run the model inference and return results to users (while also making sure miners are doing the computations honestly). This is analogous to how one might use OpenAI’s GPT-4 API to solve a problem, but here the intelligence is coming from the collectively-trained Hone model on the Bittensor subnet. According to the team’s plans, using Hone’s AI “will be similar to the OpenAI … API” for embeddings or other AI services, except it’s powered by decentralized infrastructure. In summary, the build of Hone can be seen as: (1) creating a novel AI training pipeline (world-model + hierarchical reasoner) on a decentralized network, and (2) delivering the resulting general reasoning AI service to users via Bittensor. It’s an ambitious fusion of research and product: the research side is building the AGI internals, and the product side will eventually let anyone tap into that AGI capability in a permissionless way.