The online learning methods the agent learns, supported by a multi-agent prompt-recursion framework. These methods were motivated by the traditional strategies an agent uses to decide whether to explore or exploit.

Note

  • Computational Methods

    Gaby AI does not fine-tune large language models. Computational demand is expected to remain at the scale of basic numerical methods—comparable to ordinary least squares (OLS) regression or stochastic gradient descent (SGD)—rather than the high FLOPs required for large-scale model training.

Pros of RL / Temporal Isolation

The benefits per tech department:

  • Software development: decoupling the knowledge base from model training avoids the instability and inconsistency inherent to current AI systems. By isolating Gaby’s temporal and long-term memory from the underlying base models, the application gains scalability and resilience, reducing susceptibility to fluctuations in evolving AI technologies.
  • Database management: as the application is still in its early stages, it needs a strong, robust framework for database management. Once this framework is optimised, the efficiency of the existing data-cleaning ETL pipeline can be improved too.

Cons of RL / Temporal Isolation

  • Database vulnerability: if the database workflow and architecture are not properly defined, the system is at high risk of structural failure.
  • LLM personalisation: requires more time invested in the data analytics stages when drawing features to fine-tune models.

Cognitive models, shelved ​

This section documents my trial and error implementing Bayesian belief networks and reinforcement learning algorithms to mimic the decision models of data analysts, engineers and scientists.

While my objective was to trial strategies for improving an AI's linguistics and behaviour, this may be delayed or dropped for Gaby AI, since his objective is to make decisions based on:

  • the user's domain (approximated from interactions and the dataset), e.g. cafe sales, stock markets..., and
  • machine learning life cycles.

Human-like inference may not be needed, so it is waived (and there are only 24 hours in a day).