The ASI @ASI_Alliance project hasn't been discussed for a long time, but it's recently gaining traction; it could be a good opportunity to get involved.
Last year, during the peak of AI interest, three major AI ecosystem representatives @Fetch_ai (FET), @SingularityNET (AGIX), and @oceanprotocol (Ocean) jointly initiated the ASI Alliance, merging the three native tokens into a unified $ASI (the conversion has not yet been completed, and the current $FET) and collaboratively building an open, autonomous, and collaborative decentralized artificial superintelligence platform.
The first product launched by ASI in the first half of this year - ASI-1 Mini, is the world's first large language model (LLM) specifically designed for 'Agentic AI'. Unlike Web2's large language models, ASI introduces a decentralized model architecture:
MoM: Mixture of Models
Unlike the 'single large model' structure of Web2 LLMs, the MoM architecture allows the system to dynamically combine multiple specialized models for collaborative reasoning when facing different tasks. Only the specific combination of models needed for a task is activated.
MoA: Mixture of Agents
Multiple autonomous agents with independent reasoning abilities, knowledge bases, and decision-making mechanisms work together to solve complex tasks. The coordination mechanism ensures efficient task allocation, making the system resilient, decentralized, and adaptive.
The fundamental difference from traditional large model technology:
Web2 models are closed-source 'closed brains' where you can only call APIs and cannot own, train, or control. In contrast, ASI models are modular 'open-source intelligent ecosystems' where you can train and expand based on them.
From the user experience of ASI-1 Mini, it does not differ much from ChatGPT. The main changes brought by ASI lie in the subsequent product plans; the ASI architecture is designed as a three-layer system, and the currently released product is still in the first layer.
The three-layer structure of ASI is as follows:
Base Layer (ASI-1 Mini): Acts as the core intelligence and coordination hub, employing a MoE architecture and optimizing agent/tool calls.
Specialization Layer (MoM Market): Contains multiple AI models created and provided through the ASI:<Train/> platform, with each MoM focusing on a specific field or task, providing expert-level reasoning.
Action Layer (Agents on Agentverse): Composed of multiple agents with specific functions.
Specialization Layer - privacy issues in model training
The specialization layer, which is soon to be launched as ASI:<Train/> platform, will bring qualitative changes to this large model. According to the project white paper: ASI:<Train/> employs cryptographic technologies like homomorphic encryption (FHA), allowing data to be processed while encrypted. This means that when you train using the base layer, all your training data is encrypted. This may not feel significant for ordinary people, but it is extremely useful for industries requiring high privacy, such as finance, military, and healthcare. It allows developers to ensure absolute privacy in parts that need encryption while maintaining model quality and scalability in parts that do not.
Action Layer - the 'black box' problem of model computation results
The so-called 'black box' problem refers to the difficulty in explaining the decision-making logic of AI systems (especially deep learning models) when generating outputs. ASI adopts a continuous multi-step reasoning collaboration approach, correcting and optimizing in real-time during execution, and records each computational result on the blockchain, thereby enhancing reliability. Although it cannot completely eliminate model opacity, ASI significantly improves interpretability, which is particularly important in high-risk industries such as healthcare and finance.
I have mentioned before: the path for Crypto+AI is to use cryptographic technology to solve the bottleneck problems encountered in the development of artificial intelligence, rather than pursuing a distributed model for everything. AI+Crypto is essentially the collision of artificial intelligence and cryptography. I believe @ASI_Alliance is moving in this direction, and as a merger of several leading entities in the previous AI space, it should have this mission.
