ViralMind

ViralMind is a decentralized platform on Solana built around a single core thesis: the most effective way to train computer-use AI is to pay people to demonstrate it. Instead of relying on synthetic data or closed corporate pipelines, ViralMind crowdsources human demonstrations of real computer tasks, converts those recordings into structured training trajectories, and uses them to develop Large Action Models (LAMs)—AI systems capable of operating software, navigating interfaces, and executing multi-step workflows exactly as a human would.

Large Action Models

The central product category is the LAM, which ViralMind distinguishes from conventional language models by its ability to act rather than just respond. Where a chatbot generates text, a LAM navigates a browser, fills a form, adjusts settings, or completes a transaction. ViralMind's flagship model, VM-1, is trained on millions of demonstrated data points and is offered in two tiers: small open-source models designed as drop-in replacements for OCR modules in existing automation stacks, and a foundation LAM accessible via API for gaming, enterprise workflow automation, and streaming applications. VM-1 uses a hybrid architecture combining large language model reasoning with structured action modeling and reinforcement feedback loops, optimized for low-latency multi-step execution.

The Training Gym

The primary user-facing product is the Training Gym—an environment where contributors record themselves performing computer tasks that the platform needs training coverage on. The recording system captures three components per demonstration: a video of the screen (recording.mp4), a structured event log of all user interactions in JSON Lines format (input_log.jsonl), and a metadata file containing task configuration and reward parameters (meta.json).

The event log is granular. It records mouse movements, mousedown and mouseup events, scrolls, key presses and releases, and accessibility tree snapshots that capture the underlying UI structure—bounding boxes, element roles, states—alongside the visual stream. This dual capture of visual and semantic interface data is what enables the models to generalize across software environments rather than memorizing pixel-level patterns.

Quality matters for compensation. AI-powered data quality agents review submissions and score them against accuracy and completeness criteria. Demonstrations that show deliberate, unambiguous actions, complete full task workflows without errors, and maintain consistent pacing score higher. Reward amounts, measured in $VIRAL tokens, are set per task at a maximum possible payout; final earnings depend on the quality score the submission receives.

The Forge

Alongside the public Training Gym, ViralMind offers The Forge—an enterprise-tier tool that lets businesses commission training data for their specific workflows. An enterprise client defines the tasks they want automated, funds a Training Pool using $VIRAL or USDC, and ViralMind routes those tasks to contributors in the Training Gym. The resulting model can be deployed via the VM-1 API or as a private, self-hosted system. The Forge makes ViralMind relevant to organizations that need proprietary AI automation but cannot or choose not to build internal data pipelines.

Data Marketplace

Training datasets generated on the platform can be bought and sold through a data marketplace, giving contributors a secondary avenue to monetize high-quality demonstration collections. This also allows developers building or fine-tuning their own models to purchase domain-specific action data without replicating the recording infrastructure.

$VIRAL Token

The $VIRAL token is the economic backbone of the platform and operates on Solana. Its fixed maximum supply is 1,000,000,000 tokens. At the time of the platform's early operation, circulating supply sat near 965,888,531 with approximately 3,000 holders.

Token utility is threefold. First, it is the primary reward instrument: contributors earn $VIRAL when they submit demonstrations that pass quality evaluation. Second, participants stake $VIRAL to enter competitions hosted on the platform, with a 5-10% protocol fee on those stakes directed to the treasury. Third, holding a minimum balance grants free access to certain competition tiers. $VIRAL is also used to fund Training Pools in The Forge, tying enterprise demand directly to token utility.

Competition Layer

ViralMind layers a competitive dimension onto the training process—users can participate in challenges where AI agents are tested, broken, or improved through adversarial interaction. This gamification is not ornamental. The founding team's own credibility is built partly on winning the JailbreakMe competition on Solana, a challenge where projects deployed AI assistants with safety restrictions and contestants attempted to circumvent them through adversarial prompting and reverse engineering. ViralMind's founders won two of the largest prize pools in that competition, totaling over $120,000 USD.

Team

ViralMind was founded by three engineers with overlapping AI and infrastructure backgrounds.

Dillon Dupont (Co-Founder, CTO) studied at MIT and worked as an AI engineer at Microsoft on large-scale model deployment and automation. He created the GPT-4V-Act framework and contributed to Omniparser v2, a tool for parsing UI elements from visual input—directly relevant to the accessibility-tree-based capture system underlying the Training Gym.

Jaxon Heitz (Co-Founder, CEO) studied at the University of Hawaii Manoa and brings a business and finance background to the project. He oversees tokenomics design and the $VIRAL incentive system.

Morgan Dean (Co-Founder, CIO) holds a degree in Cognitive Systems from the University of British Columbia and is responsible for infrastructure: scaling the training pipeline, deploying decentralized compute, and optimizing efficiency across concurrent training instances. Dean engineered the AI simulation used in the JailbreakMe challenge.

Technical Stack

The platform is built on PyTorch and TensorFlow for model training, WebRTC and Puppeteer for UI interaction and AI-driven browser automation, and FastAPI with GraphQL for API communication. Solana handles $VIRAL token transactions and Training Pool settlement. Deployment options include ViralMind-hosted inference, local model deployment, and API access.

Market Context

ViralMind's approach sits at the intersection of two trends: the rapid growth of agentic AI and decentralized data economies. The computer-use agent sector is positioned to grow substantially as enterprises seek to automate knowledge work that does not fit narrow software automation. ViralMind's open participation model attempts to solve the cold-start problem of agent training data—capturing diverse human behavior across software environments—while aligning incentives between data contributors, token holders, and enterprise customers through a shared token economy on Solana.