The episode explores achieving Artificial General Intelligence (AGI) by integrating neuroscience and machine learning concepts. It critiques current large language models (LLMs) for their data dependency and lack of human-like learning. The Free Energy Principle (FEP) suggests biological systems minimize surprise by aligning internal models with sensory inputs. This concept is mirrored in the Surprise Minimizing Reinforcement Learning (SMiRL) algorithm in machine learning. Liquid Time Constant Networks (LTCs), with their online learning and spiking neural network capabilities, are proposed as a potential AGI framework. The novel spike timing dependent plasticity (STDP) algorithm is highlighted for its neuron-level surprise minimization. The proposed AGI architecture, combining LTCs and surprise minimization, aims to surpass LLMs by continuously learning and improving. The importance of routines and real-time interaction in developing intelligence is emphasized, contrasting with LLMs' reliance on extensive data. An experiment is suggested to train AGI on human problem-solving processes to solve ARC-AGI puzzles. The focus on surprise minimization aims to create more efficient, human-like AI systems, challenging the current data-driven AI paradigm.