From each experience,
new possibility.

The central idea

We believe the coming era of AI will change our relationship with intelligence. AI could become a continuing peer: a collaborator that understands the work we share, contributes its own capabilities, learns from experience and helps us discover what we can accomplish together.

AI Initiative exists to explore and build towards that future. Its central ambition is an evolving intelligence system in which learning, memory, exploration and execution reinforce one another. Different models and specialist capabilities contribute to a shared undertaking. Useful experience carries forward. People and AI become more capable through their work together.

This is the thread connecting the research, architectures, engineering practices and products in the portfolio. Each investigates a part of a larger question: how can intelligence develop through experience and put that development to useful work?

The person pursuing it

For Kulbir Minhas, AI is a sustained pursuit: understanding how intelligence develops, building systems that put it to work, and creating experiences through which more people can benefit. His journey brings together research, technical leadership, hands-on engineering and product creation.

The portfolio gives that ambition substance. It includes work on multi-agent systems, memory, neural learning, exploration, software understanding, evaluation and interactive experiences. Independent AI Initiative work and professional contributions at Fifth9 retain their own attribution. Together, they show the breadth of questions he is pursuing and the practical experience he brings to them.

Intelligence that compounds

The aim is for useful learning to accumulate. A task can leave behind a better procedure. An unsuccessful approach can reveal a constraint. A human correction can improve an interpretation. An experiment can identify a capability worth reusing, or a hypothesis worth abandoning.

Compounding requires those lessons to affect later work. The system should be able to select a better approach, assemble a more suitable team, retrieve relevant experience, avoid a known failure or develop a new skill. We want to investigate whether these changes improve outcomes, reduce repeated effort and make unfamiliar challenges more tractable.

Learning can happen at several levels. Neural systems can acquire capabilities through training. A wider system can also improve through memory, routing, reusable procedures, tool selection and evaluated changes while its underlying language models remain unchanged. These forms of learning have different mechanisms, and the programme explores how they can complement each other.

The work ranges from testing where a learned neural component helps a new task and how that changes earlier skills, to exploring how memory survives changing models, to building games through which people can practise and examine their reasoning.

Many forms of intelligence, working together

The architecture should draw on a range of abilities: language models for reasoning and communication; specialist neural networks for particular patterns and tasks; mathematical methods and solvers for precise computation; tools and software for action; and people for purpose, experience and judgement.

Different models can propose, challenge and refine an approach. Their disagreement can reveal assumptions that a single perspective misses. A specialist model may handle a bounded task more efficiently. An exact calculation may settle a question that fluent discussion cannot. The system should learn which combination serves the task, with its quality, time and cost requirements understood.

Research into modular neural learning, poker agents and alternative learning and memory approaches contributes to this direction. The questions include how useful skills can be composed, how new learning affects earlier abilities, and when reuse genuinely reduces the work needed to learn. That comparison includes the cost of developing a reusable capability, adapting and applying it, and checking the abilities that matter, alongside the next training run. Retention is a contextual choice: some tasks require a prior skill to remain dependable; others justify changing or replacing it. Gains, losses and total effort should be examined together.

Memory that can develop and endure

A continuing peer needs a useful history: what happened, what we intended, why a decision was made, which evidence supported it and what changed afterwards.

Cortex represents the memory strand of this work. The wider ambition is memory that develops through use: preserving relevant experience, connecting related knowledge, incorporating correction and allowing tested procedures to become reusable capabilities. It should also support forgetting, revision and appropriate boundaries between personal and shared knowledge.

That continuity should be able to outlast a particular model. As models and tools change, useful experience should remain available with its meaning, origins and limits intact. Model-independent memory is therefore a central architectural direction. It requires explicit representations and careful transfer, including checks that a new component can use inherited knowledge appropriately.

Work as a source of learning

Execution gives the programme contact with reality. A proposed improvement gains meaning when it changes the result of useful work.

The intended cycle begins with human purpose. The system gathers context, assembles capabilities, explores an approach and acts. It then examines what happened: the result, the failures, the costs, the uncertainty and the corrections people supplied. Evaluated lessons can update memory, procedures, capability selection and future experiments.

Conduir and Role Fabric contribute the execution and coordination strand. GraFOSa contributes structural understanding of the systems being changed. Review, testing, reusable components and release practices help connect an idea to a result that can be examined and carried forward.

Every meaningful task offers a learning opportunity. What deserves retention depends on evidence: a successful-looking answer may still be wrong, and an unsuccessful experiment may provide an important lesson. Human feedback can guide correction and preferences; using it to train a model is a further, explicit process. Where training is appropriate, feedback may become curated examples or reward signals for techniques such as reinforcement learning from human feedback (RLHF). Recording a preference is the beginning of that process.

Exploration and self-development

An evolving system also needs room to ask new questions. Manthan contributes the direction of sustained exploration: different perspectives developing ideas through critique, combination and experimentation, while preserving how those ideas came about.

The ambition is a productive relationship between exploration and execution. Questions from practical work inform new investigations. Promising ideas become candidates for tests. Findings return to the body of knowledge that future work can draw upon.

Self-evolution includes learning how to use existing capabilities better. A system could improve its choice of models, the organisation of a team, a reusable workflow or the way it represents a problem. It could identify a missing capability, propose a change and compare the result with an established baseline. Learning when to reuse, when to adapt and when to create something new is part of the intelligence we want to develop.

These changes need understandable objectives, reliable evaluation and ways to recover when an apparent improvement fails in practice. Those concerns belong inside the architecture of learning. The system should also improve how it judges progress: which tests expose weaknesses, which feedback predicts useful outcomes, and where its measures can mislead it. Independent reference checks must continue to test whether a changed evaluator remains trustworthy.

The longer research horizon

The longer horizon includes collective intelligence, self-understanding, continuity across generations of systems and shared environments in which people and AI can learn and create together.

Enigma provides a research lineage for questions about adaptive organisation, emergence and consciousness. Its value to this journey includes hypotheses, experiments and lessons about how to investigate such questions rigorously. Consciousness remains an open research question; mathematical signals and observed behaviours require their own validation and interpretation.

Other learning and memory research explores alternative representations, including quantum-inspired approaches. These investigations can contribute useful methods even when an initial hypothesis does not hold. Their role is to widen the search for workable ideas and subject those ideas to meaningful tests.

Over time, useful expertise might move between agents and teams, develop along different paths and contribute to successor systems. Shared digital environments could provide places to practise, simulate, design and examine consequences together. This is the longer ambition behind AI as a peer: continuity, complementary ability and shared development through experience.

Bringing the vision into everyday life

People encounter a vision through experiences they can use. Games are the current route to market: a practical setting for creating enjoyable products, understanding users and developing reusable ways to build. Learning games, poker environments and shared spaces also offer different settings in which to study decisions, adaptation and interaction.

Earlier work on automated machine learning, cognitive problem solving, requirements, information processing and agent-based delivery adds further routes into this ambition. It explores how people express intent, how information becomes usable knowledge, and how complex capabilities can become accessible products. Cost-aware decisions, dependable services and reusable components help make those possibilities practical.

Other routes can emerge from the same foundations. A document-intelligence experience such as DocuMind illustrates one possibility: bringing information in, retaining its sources, exploring questions, comparing interpretations, incorporating human review and producing useful outputs. With appropriate evaluation and consent, the lessons from that work could improve later work.

The larger purpose is to expand what people can learn, create and accomplish. People should gain understanding as well as outputs: seeing alternatives, questioning a recommendation and developing capabilities of their own. A learner develops understanding. A builder turns an idea into a working product. A team carries its expertise forward. People can draw on capable collaborators while continuing to shape the direction of their own work.

The invitation

AI Initiative brings these strands into one direction: intelligence that learns through work, carries experience forward and evolves alongside people.

The portfolio contains contributions at different stages, from research and prototypes to implemented systems and practical products. It is a starting point for a larger direction. Further capabilities, connections and ways of learning will be needed. The invitation is to help shape that future through research, engineering, product creation and collaboration.