Skip to content

World I

The Machine

How language models work, how agents are built and governed, and what it takes to run them in production.

AI & Machine Learning

Language models, agents, evaluation and the AI Systems collection. Open the shelf →

Written by Rajesh Ranjan Mahapatra

Rooms, series and atlases

Essays

176 essays, newest first where the site knows the date.

Research papers

40 papers from the research programmes.

A reliability model for compound AIA first-principles reliability model for compound AI systems that separates route exposure, shared failure causes and containment, with an e
Active inference as a design hypothesis for agentsA first-principles test of when expected free energy improves artificial agents, how it differs from reward-based control, and where the bio
Agent plus operational twinA first-principles architecture for letting an AI agent rehearse actions in a stateful operational twin while authoritative systems retain c
An evidence grammar for artificial mindsA first-principles method for keeping behavioural, report, mechanistic, provenance and replication evidence distinct while composing them in
Aristotle for artificial intelligence
Benchmarks, capabilities and construct validityA first-principles method for deciding whether an AI benchmark score measures the intended capability, a shortcut, or the interface to the t
Causal attribution of consequential tool calls
Cognitive integrity and retained human intelligenceA practical framework for deciding which human capabilities must remain exercised when AI delegation is faster, based on failure detection,
Collective intelligence across people and agentsWhy communication sometimes creates knowledge and sometimes creates herding, with a correlation-aware protocol and an executable truth-seeki
Control theory for guardrailsWhy runtime AI guardrails need measured state, bounded delay, sufficient authority and a verified fallback, with an executable guardrail sta
Creativity, analogy and human-AI co-creationA first-principles account of when AI-assisted recombination becomes useful novelty, why analogy must preserve causal relations, and how to
Extended, distributed and coupled intelligenceA causal test for deciding when a tool, team or AI system becomes part of a cognitive mechanism rather than merely assisting it.
Failure injection as the primary evaluationWhy consequential compound AI systems should be released against typed fault propagation, containment, recovery and proof, with an executabl
From model to system: scaffolding, memory, tools and agentsA first-principles method for deciding whether an apparent agent capability belongs to model weights, runtime, context, orchestration, tools
Human-AI joint cognitive systemsA first-principles guide to allocating perception, interpretation, proposal, authority and recovery across human–AI systems, with an executa
Integrated information and the causal structure of artificial systemsWhy integrated information theory evaluates a physical cause-effect structure rather than an algorithm, and how to build a bounded causal do
Interventions on artificial minds
Machine introspection: evidence and contaminationA first-principles test for whether a language model's self-report comes from privileged access to its hidden state or from prompt, output,
Model-independent containment and property placementA first-principles account of which agent safety properties can be enforced outside the model, where to place them, and how to test a typed
Moral patients, moral agents and the basis of statusWhy the capacity to be harmed and the capacity to be held responsible require separate evidence, thresholds and governance decisions.
Neural codes, synapses and plasticityA first-principles account of why neural codes only become explanatory when activity, downstream readout and plasticity are tested together.
Normal accidents and graceful recovery in agentic systemsA first-principles account of why agentic failures cascade, how causal coupling makes some accidents unavoidable, and how to design firebrea
Perception as inference, action and correctionWhy perception is a closed-loop hypothesis test in which expectations propose causes, precision-weighted evidence corrects them, and action
Reasoning, planning and tool use under uncertaintyA causal test for deciding when generated reasoning genuinely guides tool-using action, when it merely narrates a choice, and how to instrum
Reasoning traces, chain of thought and hidden computationA causal method for deciding when a model's articulated rationale tracks the computation that produced its answer, with an executable TRACE
Reflexive self-models without infinite regressA first-principles architecture for reflexive control: compress an agent's allocation traces into a bounded causal state, feed that state ba
Retention, recurrence and machine stateA first-principles account of how caches, recurrent states and state-space dynamics preserve the past, with an executable temporal-preservat
Retrieval as inference, sufficiency and entitlement
Self-models, symbols and editable identityWhen self-models improve control, how symbols make identity editable, and why recursive self-description can become confabulation.
States, contents and neural correlatesWhy neural correlates of consciousness become interpretable only when global state, specific content and the evidence bridge are experimenta
The artificial neuron, backpropagation and credit assignmentA first-principles explanation of how backpropagation assigns path-conditioned credit, why gradients fail across depth, and how to audit a h
The end-to-end argument for agent protocolsWhy sound agent protocols cannot carry the full responsibility for intent, present authority, real-world effects and recovery, with an execu
The human intelligence stackHuman intelligence is a task-specific feasible set created by coupled bodily, cognitive, social, linguistic and cultural constraints.
Tokenisation and the worlds it makes visibleHow token choice redistributes context, statistical evidence and computational difficulty before a transformer begins to reason, with an exe
Training a language model end to endA first-principles account of how data, objective, optimiser, scale and measurement jointly create a language model, with a reproducible min
Training provenance and evaluation validity
Variational inference and the ELBOHow the evidence lower bound turns intractable Bayesian posterior inference into optimisation, why the projection creates approximation debt
When multi-agent systems are justifiedA first-principles test for deciding when several agents create real value, when one agent with tools is enough, and how to measure the coor
Why biological intelligence has stakesA first-principles account of how embodiment, viability control, interoception and valence make outcomes matter to biological intelligence,
World models that survive interventionWhy predictive accuracy is not enough for a world model, and how to test whether learned dynamics preserve causal structure when actions, me

Insight Editions

16 study editions with covers, PDFs and EPUBs. All editions →

The other worlds

World IIThe Path World IIIThe Frontier World IVThe Build World VThe Mind World VIThe Source World VIIThe Number World VIIIThe Mystery World IXThe Wisdom World XThe Library