Why feedback speed matters in machine-learning systems
How research cycles, observability and the cost of being wrong shape model development.
Research / Library
Questions before
conclusions.
Our work starts with hard questions about feedback, infrastructure and utility. This library sets out the research themes guiding our investment platform.
Research library
Read note ↘
Compute economics beyond the headline GPU count
A full-stack view of utilisation, power, networking, cooling and orchestration.
From token issuance to durable operational utility
Why useful tokenisation begins with rights, workflows and infrastructure—not the token itself.
Designing for the regime the model has not seen
Tail behaviour, adaptive controls and the limits of historical validation.
01 / Quant systems · Research note
Why feedback speed matters in machine-learning systems
A model does not become useful because it is complex. It becomes useful when the system around it can tell us, quickly and honestly, where it works and where it fails.
Quantitative markets offer unusually fast feedback, but speed alone can mislead. Labels can be noisy, execution can change the result, and a model may learn a temporary market condition rather than a durable relationship. The research advantage comes from shortening the loop between hypothesis, controlled deployment, observation and revision while protecting that loop from false confidence.
We therefore view data lineage, experiment tracking, simulation quality and live observability as one research product. A faster loop is valuable only when it preserves the information needed to distinguish insight from accident.
02 / AI infrastructure · Investment brief
Compute economics beyond the headline GPU count
Installed accelerators are not the same as productive compute. The investable system includes every constraint between power delivery and useful workload completion.
Utilisation can be lost through network bottlenecks, storage design, cooling limits, workload scheduling or software friction. Power cost matters, but so do availability, density, reliability and the ability to serve a changing mix of training and inference demand.
This shifts diligence from a simple capacity question to a systems question: what work can the infrastructure complete, at what total cost, with what constraints, and how will that position change as hardware and workloads evolve?
03 / Token systems · Investment thesis
From token issuance to durable operational utility
The token is a representation layer. The investment question is whether the system underneath it creates a better way to own, access, coordinate or settle something valuable.
Durable token systems require more than issuance technology. They need clear rights, credible governance, identity and compliance pathways, lifecycle operations, custody and interoperability. Without these, programmability remains a demonstration rather than an operating advantage.
Our foundry thesis starts with real workflows and measurable coordination costs. Tokenisation becomes interesting when it can remove friction, expand useful access or enable products that could not operate effectively on existing rails.
04 / Risk systems · Research outline
Designing for the regime the model has not seen
Historical success is evidence, not a guarantee. Robust systems make room for structural change, dependency shifts and outcomes outside the training distribution.
The practical response is not to predict every shock. It is to make assumptions visible, monitor when they weaken, limit the damage of correlated errors and preserve the ability to intervene. Research and control systems should evolve together.