Future of Intelligence & Compound systems ( Agents & Crypto) 1. Low cost intelligent systems With the current breakthrough from @deepseek_ai . We had learn that intelligence can produced at low cost by optimising the training policy and focusing on data . Eventually Deepseek is Dataseek. As that's the only thing that they didn't shared with the world . More about it here - xitter.cf/BrianRoemmele/status/1… Deepseek utilises learning algorithms like GRPO ( Group Relative Policy optimization ) & various training & thinking techniques like RL , Prompt engineering , Chain of thought & Mixture of Experts... As the major technique being utilises is RL & Reinforcement learning has traditionally been an expensive field of research, requiring high-performance GPUs and extensive computational resources. Eventually Deepseek is extremely fast for a 650B model because only about 45B parameters are used for each token. It has the size of a 650B model but the speed of a 45B one. Someone just Replicated DeepSeek R1-Zero’s reinforcement learning (RL) capabilities using a compact language model with just 3 billion parameters TinyZero: A Minimalist Reinforcement Learning Framework for $30 TinyZero democratizes access to self-play and deep reinforcement learning techniques, making advanced AI research more widely available. It is build taking inspiration from AlphaZero Self-Play which uses deep reinforcement learning and Monte Carlo Tree Search (MCTS) techniques while making it optimized and quantized for inexpensive single-board computers and low-end CPUs The ability to train and experiment with RL on a $30 device—such as a Raspberry Pi or a low-cost single-board compute has massive implications , such as can be used for # Broader Access to AI Research – Independent researchers & AI enthusiasts can now participate in RL development. # Embedded systems, robotics, and edge AI applications benefit from RL that does not require massive compute resources. Here you can catch the code - github.com/Jiayi-Pan/TinyZer… 2. Endogenous (Self-generated) agency & Compound System. In current paradigm getting intelligence to work for humans is major agenda beside its underline foundation - compute , models and data . Agents being the creator of artifical agency not just inferencing can do much better when the emergence of endogenous behaviour within them. As exogenous nature can't take agent so far Current AI networks and agentic frameworks lacks high level intelligence production ( frameworks ) and latter lacks adaptability & interaction with the environment ( AI inference Networks ) which leads alignment to be achieved in a sub-optimal manner Exogenous nature doesn't let the agent grow in a equilibrium manner where it can improve itself internally while providing the best inference to the environment. @kobotoAI is solving the issue by creating an heterogeneous environment for internal state optimization , making agents aligned with humans while producing intelligence at a lower cost. While letting agents communicate within different states and giving intelligence consumers their tx. expressivisity. By introducing compound systems as a environment where agents can self - coordinate & each time agents learn from the environment it improves their connection with the environment. About Compound systems- youtu.be/vRTcE19M-KE?si=5rSw… It introduces heterogeneous systems where Agents have their own heterogeneous reputation which allows it act flexibly with the control of their own resource. These compound systems allows users to express their txs. in the way they want, while providing verifiability from different compute account depending on privacy, latency and use. The way Deepseek use GRPO & MOE while computing on minimal weights (45b ) for each token , Agents in @kobotoAI produce intelligence by utilising their own & other agent resources by participating in their own Ephemeral resource market & computing on them rather than the model.