Conviction

AGI is increasingly an operational challenge of bringing tasks under distribution using reasoning models with tool access. The winners won't just have better models—they'll have better operational systems for deploying AI at scale.

Years are required to train novices in electrician, HVAC, and solar installation roles. The opportunity: build a full-stack services company combining minimally trained workers with AI guidance systems, AR hardware, and remote expert oversight. Don't sell software to trade companies—become the trade company.

Agents trained on games and simulations lack real-world enterprise complexity. Build multi-agent simulations reflecting actual office communication (Slack, video calls, file sharing) for training domain-generalist work agents.

Large enterprises lack unified reasoning systems and integration hubs for AI actions. Build a live enterprise ontology mapping plus centralized action registry for AI execution across legacy systems.

Healthcare faces repetitive, time-sensitive supply and sample movement between departments. Deploy robots that generate revenue while collecting deployment data for continuous improvement. Pick a specific vertical, deploy robots that create immediate value, and use the data to build an insurmountable advantage.

Drug development has single-digit clinical trial success rates despite improved molecular design. Use models like single-cell perturbation datasets to "pull forward" clinical risk to pre-clinical settings.

Consumer no-code tools lack enterprise features like access control, integrations, and compliance. Build an enterprise-specific application builder with business-oriented templates and security features.

50% of CMDB data is unreliable, contributing to over 60% of IT incidents caused by misconfigurations. Build agentless discovery, graph construction, and AI reasoning for robust enterprise IT visibility.

Nvidia dominance in AI chip market limits competition and supply options. Build alternative chip solutions including TPUs, reconfigurable architectures, and latency-optimized inference hardware.

Generated 3D models lack precision needed for manufacturing and construction applications. Combine generative models with validation and simulation for production-ready 3D assets.

Manual metadata collection is painful and incomplete; legacy catalogs poorly support unstructured data. Build AI-driven classification, policy application, quality detection, and lineage understanding for modern data stacks.

Fragmented industries like HOAs, BPOs, and accounting firms are slow to adopt AI tools. Acquire and operate underinvested businesses directly rather than selling into them.