Research library
Our essays explore the choices people face when they build and use AI. Most appear on Integrated Intelligence, Brightbeam’s Substack; others are published elsewhere, for example by the Agentic AI Foundation. You’ll also find a recorded technical introduction to CHAP below, and the Medtech playbook we published with Irish Medtech Skillnet and Irish Medtech.
An operator knows when the procedure needs a closer look. This essay proposes a way to retain part of that experience as agent memory, with its source, limits and human review attached.
Written notation lets many players follow one score. The article applies the idea to people and agents: it walks through CHAP’s workspace, Core methods and profiles, and shows the CHAP coordinator running as an ordinary MCP server.
A fluent answer can hide the limits of a model’s knowledge. The essay examines uncertainty and what a system needs to keep track of its sources and distinguish a finding from a guess.
Connecting an agent to enterprise tools gives it the means to act. The next decision is which actions it may take alone, which need approval and which should be blocked.
An agent can mistake instructions in a document for instructions it should obey. The essay makes the case for restricting access and reviewing work before actions become irreversible.
Five diagrams explain the Composable Digital Intelligence paper: where different architectures have strengths, how they fit together and who answers for the result.
Language models are capable in some areas and surprisingly weak in others. Philip Black examines what those differences tell us about intelligence and the role of specialised components and human experts.
If you need to show what an agent knew at the time, its memory needs a record of its own. The essay argues for a store controlled by the workflow, with retention rules and an audit trail.
The model is one part of a production agent. This essay examines the surrounding memory, tools and controls, including rules that need to be enforced in code.
A colour-gamut analogy explores the limits of learning from existing material, and how new senses and feedback from the world might extend them.
Faster decoding can make a model practical on hardware an organisation owns. The essay explains the mechanism and why the gain depends on the model, workload and serving conditions.
For organisations at scale or under regulation, model choice includes decisions about tokenisation, precision and adaptation. The essay examines when those choices justify the engineering effort.
An introduction to the Collaborative Human-Agent Protocol and its record of collaboration between people and AI agents.
Where generative AI can be applied across Medtech, with a roadmap from first experiments to operational capability. Built with six Medtech companies in Ireland, and free to read online.
No results. Try another theme or a different search.
