Dastan Aitzhanov

I build AI systems and write about the governance questions that appear when they begin to act.

I think about AI the way an owner does, not only the way an engineer does: who authorized it, who pays, who owns the result, who bears the downside, and whether it can be stopped.

CurrentlyBuilding and testing a personal AI system around durable private context, actions that stay inside explicit authority, and work that can move between a person's own machines and shared compute. I publish the parts that hold up.

Current inquiry

As AI systems move from answering questions to exercising delegated authority, how should that authority be granted, bounded, verified, and revoked?

The recurring question is no longer only what a model can produce. It is what the surrounding system should permit it to see, decide, and do.

  1. A smaller model with better structure

    Independent investigation · Evaluation and technical article

    In 35 head-to-head enterprise retrieval tests, Llama 3.1 8B with a knowledge graph beat Llama 3.3 70B on its own: a 55% higher core score, 13x lower input-token cost, 3x lower latency, and a verifiable tool-call chain.

  2. Local AI and intelligent scaffolding

    Independent investigation · Systems experiment and technical article

    A hands-on experiment in running a smaller local model inside explicit agent boundaries, observable workflows, and independent controls.

  3. Industrial asset monitoring with IoT and machine learning

    Public technical work · Co-authored two-part technical work at AWS

    Co-authored two-part AWS architecture for connecting industrial telemetry, asset models, monitoring, machine-learning training, and scheduled inference.

  4. Human review for industrial predictive maintenance

    Public technical work · Co-authored technical work at AWS

    Co-authored AWS technical work connecting anomaly detection, human review, dataset improvement, and model retraining for industrial equipment.

All selected work

Research threads

Four questions connect the experiments, essays, and shorter notes.

  1. Delegated authority

    What changes when an AI system moves from producing information to taking action?

  2. Personal AI

    What should remain near the person when AI becomes part of ordinary computing?

  3. Distributed governance

    Where does policy attach when computation crosses independently owned machines?

  4. Ownership and incentives

    Who decides, who carries the risk, and who remains responsible?

  1. Govern the boundary

    Model-based rules assume capability is scarce. As capable models become small and cheap, the durable place to attach policy is the layer that decides what an AI system may do.

  2. Capacity is not consent

    Distributed compute is usually framed as a problem of finding available machines. The harder question is who is allowed to decide what those machines do.

  3. Probabilistic Intelligence, Deterministic Control

    Generative models are for reasoning that is still open. Repeated decisions belong to a narrower model, and repeated actions belong to ordinary software.

All writing

  1. An agent can own the mistake. It cannot own the cost.

    Accountability means bearing the downside. An agent that spends money it cannot repay needs a limit set before it acts, not an apology after.

  2. Autonomy should scale with reversibility

    The useful question is not whether an AI agent is autonomous, but how much autonomy makes sense for a particular action.

  3. Clarification is part of intelligence

    A system that confidently completes the wrong task is not more intelligent than one that knows when it needs to ask.

  4. Novelty belongs to models. Repetition belongs to software.

    Use a generative model to find a path the first time. Once the path repeats, call it. Do not pay a model to rediscover it.

All ideas

I want to hear from people working out how AI systems should be governed once they act: the builders setting their limits and the teams accountable for what they do. Email me at dastan.aitzhanov@gmail.com.