State and Statefulness of a system:
State of a system simply means form (or configuration of all active variables) of a system at an instance. Statefulness means preserving essence of that state across multiple instances. Any intelligent entity must be stateful to operate in the temporal dimension of the our universe. However, how much statefulness is bare minimum to beat human level intelligence? I don't know. However, let's find out how much stateful is a human brain & what's the state of the art AI model architectures capable of achieving:
Statefulness of the Human Brain:
Let's quickly glance over how does bran solves memory, so we have a state of the art reference of how it good memory looks like.
- What types of memory do human brains have? Our brains have short-term sensory memory, episodic memory, long term seemingly infinite memory, ...
Short-Term memory: Episodic memory: Long-Term memory:
- How does a brain forms and forget these things though?
Forgetting Memory Mechanism is just as important as forming it.
- How does the inference loop of the brain looks like?
[ REST IS UNDER PROGRESS ]
I will be exploring the following:
- Hebbian Memory & its emergent properties: Fast changing matrix, offloadable, sort of decoupled from the residual stream... gets you for free significant activation sparsity during inference how does forgetting works in hebbian memory training stability & imeading SGD long run agentic task performance - here the COOM like benchmark
- Deep Seek's n gram lookup memory: Model can be trained on a lot larger corpus of data and can achieve very high retrieval accuracy and performance because it's able to load information in its residual stream - dynamically within an inference instead of a deterministic prompt engineering system deciding what's best in the given situation - the model gets to decide.
- Other forms of short and long term statefullness (architectural level ideas, and not RAG)
Philosophical Question:
What % of human effort should currently be going into architectural level advancements vs scaffolding level advancements (agents, orchestration, RAG, etc...)