Investment firm focused on companies turning advanced research into infrastructure, products and durable returns. We combine capital discipline, technical diligence and a long-duration view of innovation cycles in artificial general intelligence.
AGI companies funded since 2021 across the globe.
Initial checks and selective follow-ons in frontier architectures.
Weighted return on investment across the entire portfolio.
Total capital raised by our startups in subsequent rounds (Series A+).
Angel investor with a core conviction: artificial general intelligence emerges at the intersection of deep reinforcement learning, neuro-symbolic reasoning and internal world models — and the startups mastering this convergence will define the next computational paradigm.
The focus is on teams that move beyond simple LLM scaling to explore the composability of latent representations, causal attention mechanisms, hierarchical sub-goal generation and continual learning without catastrophic forgetting. AGI will not be a single model, but an ecosystem of specialized modules orchestrated by a recursive meta-learning loop.
Beyond funding, technical support is provided to founders: multi-agent system architecture, evaluation strategies on ARC-AGI-type reasoning benchmarks, intrinsic reward design for autonomous exploration, and navigating regulatory challenges related to general-capability systems.
AGI emerges from the convergence of compute scaling, persistent episodic memory architectures and counterfactual causal reasoning systems.
Direct involvement in architectural decisions: computational graph topologies, model distillation strategies and multi-task training curricula design.
Startup evaluation based on reproducible metrics: out-of-distribution generalization capability, adversarial robustness and performance on zero-shot compositional reasoning tasks.
Priority to teams integrating iterative RLHF, formal verification of safety properties, mechanistic interpretability and corrigibility protocols from the earliest development phases.
Transformer scaling laws, combined with breakthroughs in chain-of-thought prompting and Monte Carlo tree search-style reasoning, point to a clear trajectory toward systems with general cognitive capability.
Investing at the earliest stage, where fundamental innovations in knowledge representation and world model learning are still at proof-of-concept stage.
The trajectory toward ASI (Artificial Superintelligence) demands investments now in computational governance frameworks, safe shutdown protocols and value alignment for recursively self-improving optimization systems.
Hybrid architectures combining Bayesian probabilistic inference with differentiable symbolic reasoning modules. Focus on program synthesis, gradient-based constraint solving and hierarchical concept abstraction via primitive library learning (DreamCoder, neural λ-calculus).
Multi-agent systems with hierarchical planning through chain-of-thought task decomposition, dynamic tool use and sparse-attention working memory. Exploring cognitive scaffolding architectures where agents build and revise their own decision heuristics through meta-learning.
Research in feature decomposition via sparse autoencoders, induction circuit mapping in transformers, mesa-optimizer detection through internal gradient analysis and automated red-teaming protocols for systems with unanticipated emergent capabilities.
Architectures learning compressed representations of the physical world through self-supervised prediction. Video prediction models, intuitive physics models, mental simulation via latent diffusion and imagination planning in structured latent spaces (Yann LeCun's JEPA, Genie).
Infrastructure for frontier model training: tensor/pipeline/expert parallelism, mixed-precision quantization (FP8/INT4), mixture-of-experts architecture with conditional routing, ML compilers (XLA, Triton) and low-latency communication fabrics for multi-datacenter clusters.
Unified cross-modal encoding systems via contrastive learning (CLIP, ImageBind) extended to proprioception, haptics and vestibular flow. The symbol grounding problem hypothesis resolved through embodied AI and active interaction with the physical environment via sim-to-real transfer.
The most important criterion: exceptional founders.
A non-incremental technical advance on the path toward AGI.
A credible roadmap toward systems with general cognitive capability.
We invest early in teams that combine technical edge, commercial ambition and execution discipline. Infrastructure, models, agents, safety and production tooling: if the technology creates durable advantage, it is relevant to us.