Agentic General Behaviour

Those who understand the LLM architecture, know about the very mechanical nature of these machines. For others they seem to depict a perceived intelligence. This intelligence however is unlike any witnessed before. It is able to solve exams on any topic, have an answer for everything that is known, excel on our highest IQ tests and even perform on benchmarks consisting of tasks at par with human level reasoning.

Yet LLMs lack behavioural similarity, a sense of self-awareness and ability for high fidelity interaction with the environment.

Sub-systems to the rescue, to go beyond context and single cycle execution:

• Memory - to go beyond context length limitation.
• Cron - to schedule for tasks, self introspect, recursive calling etc.

Behavioural attributes of a human that a combination of above sub-systems will allow:

• Dream: a culmination of our long-term memory and short-term events
• Self-learn: fine-tune and context engineer itself
• Stimulus: agentic loop driven by not just prompts but a sense of purpose

Through this the models can achieve perceived general behaviour and be more than just simple input-process-output machines.