where does “emotion” fit within machine intelligence?

If emotion in the human brain through the limbic system, especially the amygdala and hippocampus, is concerned with memory, then what role might “emotion” play in machine intelligence if memory in transformer-based LLMs equates to immutable vector representations augmented by MCP pipelines?

Further, if most other biologic brain regions are interconnected through a combination of motor movement and interactions with the external world, then how can machine intelligence, when relegated to the parsing of existing “knowledge” as large corpora by reasoning through prompting necessitating ever increasing hardware expenditure through the scaling hypothesis, be effective when applied to generalized automation if it does not have the same access to external stimuli/percepts as natural intelligence given reinforcement strategies have proven to be unreliable in unforeseen circumstances?

Does that mean machine intelligence should only automate decision making in specific scenarios when concerned with reasoning over known outcomes until an energy-based heuristic and experiential intelligence is internalized as probabilistic latent representations, or is a new paradigm necessary such as Yann LeCun et al.’s JEPA approach?

Propping up fundamentally limited architecture with unsustainable compute costs and token-consumption does not solve the underlying realities associated with brute-force autoregressive scaling and will not work forever without net new computing paradigms like massive advancements in quantum.

Seems like joint-embedding predictive architectures plus some form of search over potential future outcomes as a validation harness for common sense world models is indeed the future.

further reading

or watching