jakiro
Author: accelxr, 1kx; Translation: 0xjs@作 作 作 作 作 作
The main purpose of the current generation model is content creation and information filtering.However, the recent research and discussion of AI Smart (independent participants who use external tools to complete user -defined goals) show that if AI provides an economic channel similar to the Internet in the 1990s, AI may get substantially unlock.
To this end, the intelligence needs to be represented by the assets they can control, because traditional financial systems are not set for them.
This is where the encryption plays the role: the encryption provides a fast settlement, digital payment and ownership layer, which is particularly suitable for constructing AI intelligence.
In this article, I will introduce you to the concepts of smart and smart architecture. How does the examples in research prove that smart bodies have emerging attributes beyond traditional LLM and projects that build solutions or products around an encrypted intelligence.
What is intelligent
AI Smart is a LLM -driven entity that can plan and take action to achieve goals in multiple iterations.
The intelligent body architecture consists of only a single or multiple intelligence to solve the problem together.
Usually, each intelligent body is given personality and can use various tools, which will help them independently or complete their work as a team.
The intelligent body architecture is different from our usual way to interact with LLM today:
Zero -time prompts are the way most people interact with these models: you enter the prompt, LLM responds according to its pre -existence knowledge.
In the intelligent body architecture, if you initialize the target, LLM decompose it as a child task, and then it recursively prompts yourself (or other models) to complete each sub -task until it reaches the target.
Single intelligent body architecture and multi -intelligent body architecture
Single intelligence architecture: A language model performs all reasoning, planning and tool execution by themselves.There is no feedback mechanism from other intelligence, but humans can choose to provide feedback to the intelligence.
Multi -intelligent architecture: These architectures involve two or more smart parties, and each intelligent body can use the same language model or a set of different language models.Smart can use the same tools or different tools.Each intelligence usually has its own role.
-
Vertical structure: A smart body acts as a leader, and other intelligence reports it.This helps the output of the organization.
-
Horizontal structure: A large group discussion about tasks, each smart body can see other messages and voluntarily complete the task or call tools.
Smart architecture: configuration file
Smarts have configuration files or personalities, which define their characters as prompts to affect LLM behavior and skills.This largely depends on specific applications.
Maybe many people have used it today as a prompt technology: “You are a nutrition expert. Provide me with a diet plan …”.Interestingly, the role of LLM with LLM can increase its output compared with the baseline.
The configuration file can be made through the following methods:
-
Handmade: The configuration file manually by the human creator; the most flexible, but it is also time -consuming.
-
LLM generated: Use the configuration file generated by LLM. The configuration file contains a small amount of a small amount of samples around the rules set +(optional) of the composition and attributes.
-
Data set alignment: The configuration file is generated according to the data collection of personnel of the real world.
Smart architecture: memory
Smart memory storage information is perceived from the environment, and uses this information to formulate new plans or actions.Memory enables intelligence to evolve and act according to its experience.
-
Unified memory: Similar to short -term memory realized through context/through continuous prompts.All related memory will be passed to the smart body in each prompt.Mainly restricted by the size of the context window.
-
Mixed: Short -term+long -term memory.Short -term memory is a temporary buffer in the current state.Reflex or useful long -term information is permanently stored in the database.There are several ways to do this, but the common method is to use a vector database (encoding memory as embedded and stored; memories come from similar search)
-
Format: Natural language, database (for example, fine -tuning SQL inquiries SQL), structured list, embedded
Smart architecture: planning
The complex task deconstruction is solved alone.
No feedback plan:
In this method, the smart body will not receive feedback that affects future behavior after taking action.An example is the chain chain (COT), which encourages LLM to express its thinking process when providing answers.
-
Unilateral reasoning (such as zero COT)
-
Multi -path reasoning (for example, self -consistent COT, which generate multiple COT threads and use the highest frequency answer))
-
External planner (such as the definition of the planning domain)
Plan with feedback:
Iterates and refine sub -mission according to external feedback
-
Environmental feedback (such as game task completion signal)
-
Human feedback (such as soliciting feedback from users)
-
Model feedback (for example, soliciting another LLM feedback -crowdsourcing)
Smart architecture: Action (ACTION)
Action is responsible for transforming the decision -making of the smart body into a specific result.
There are many possible forms of behavioral goals, such as:
-
Mission completion (for example, making an iron pick in Minecraft)
-
Communication (for example, share information with another smartman or human)
-
Environmental exploration (such as searching for your own behavior space and learning your ability).
The generation of behavior usually comes from memory or planning. The behavior space is composed of internal knowledge, API, database/knowledge base, and external models of its own use.
Smart architecture: ability acquisition
In order to correctly perform the action in the action space, the intelligent body must have the ability to be specific to the task.There are two main methods to achieve this:
-
Through fine -tuning: Train intelligence on the data set in artificial annotations, LLM generation, or real -world examples.
-
No fine -tuning: You can use LLM’s congenital abilities through more complicated prompt engineering and/or mechanism engineering (that is, combine external feedback or experience accumulation).
Intelligent examples in the literature
Generating intelligent body: Interactive simulation of human behavior: instantiated and generated intelligence in the virtual sandbox environment, showing that the multi -intelligent system system has emergencies.Starting from the forthcoming Valentine’s Day party designated prompts, the intelligent experience will automatically send invitations, meet new friends, date each other, and coordinate to participate in the party together in the next two days.You can use A16Z AI TOWN to try it yourself.
Description Explanation Plan Selection (DEPS): The first one can complete the zero -sample multi -task smartizer of more than 70 MINECRAFT tasks.
Voyager: The first intelligent body driven by LLM in Minecraft, which reflects lifelong learning, can continue to explore the world, obtain various skills, and make new discoveries without manual intervention.Continuously improved its skill execution code based on the feedback of repeated tests.
Calypso: The intelligent body designed for the game “Dragon and Dungeon” can help the main creation of the dungeon to create and tell the story.Its short -term memory is based on scene description, monster information and previous summary.
Ghost in Minecraft (Gitm): Smart -generally capable intelligence in Minecraft, the success rate of diamonds is 67.5%, and the completion rate of all items in the game is 100%.
Sayplan: LLM -based robotic large -scale task planning, using 3D scene graphics representation, showing the ability to perform long -term task planning from abstract and natural language instructions for robotics.
Hugginggpt: According to the user’s prompts, the task planning is used to select the task plan, according to the description of the description on the Hugging Face, and perform all sub -tasks. It has achieved impressive results in language, vision, voice, and other challenging tasks.
Metagpt: Accept the input and output user story / competitive analysis / demand / data structure / API / document.Inside, there are multiple intelligences that constitute various functions of software companies.
ChemCrow: A LLM chemical intelligence, which aims to use the tools designed by 18 experts to complete the tasks such as organic synthesis, drug discovery and material design.The autonomous planning and execution of insect repellent and three organic catalysts are synthesized, and a new type of hair color group is guided.
Babyagi: Use OpenAI and vector databases (such as Chroma or Weaviate) to create, determine the general infrastructure of priority and execution tasks.
Autogpt: Another example of a general infrastructure for launching LLM intelligence.
Smart body example in Crypto
(Note: Not all examples are LLM -based + some may be more loose in the concept of smart body)
Frenrug from RitualNet: Based on the GPT-4 Turkish carpet salesman game {https:// aiadventure.spiel.com/carpet}.Frenrug is a broker, and anyone can try to convince him to buy their Friend.tech Key.Each user message is passed to multiple LLM running by different Infernet nodes.These nodes responded on the chain and voted by LLM to determine whether the smart body should buy the proposed key.When there are enough nodes to respond, the voting will aggregate. The monitoring the classifier model will determine the operation and pass the validity proof on the chain, so that the execution of the chain of multiple classifiers can be verified.
GNOSIS uses Autonolas’s predictive market intelligence: AI robots are essentially AI services smart contract packagers. Anyone can call it by paying and asking questions.The service will monitor the request, perform the task, and return the answer on the chain.This AI robot infrastructure has been expanded to the forecast market through OMEN. Its basic concept is that the intelligent body will actively monitor and bet on news analysis predictions, and eventually obtain a summary forecast that is closer to real odds.Smarts searched the market on OMEN, paid the “robot” to predict the theme on the “robot”, and used the market for transactions.
IANDAOS GPT & LT; & GT; Safe Demonstration: GPT uses Syndicateio to transaction cloud API to manage USDC independently in the SAFE multi -signature wallet on its own base chain.You can talk to it and make suggestions on how to make the best use of its capital, it may be allocated according to your suggestions.
Gaming intelligence: There are many ideas here, but in short, the AI intelligence in the virtual environment is both a companion (such as AI NPC in Skyrim) and competitors (such as a group of chubby penguins).Smart can automatically execute the revenue strategy to provide goods and services (such as: shop owners, travel merchants, sophisticated generation task providers), or semi -play characters in Parallel Colony and AI ARENA.
Safe Guardian Angels: Use a set of AI intelligence to monitor wallets and defensive potential threats to protect user funds and increase the security of wallets.Features include automatic cancellation of contract permissions and extracted funds when abnormal or hackers attack.
Botto: Although BOTTO is a smart body example defined on the chain, it shows the concept of the artist on the autonomous chain. The created works are voted by tokens and auctioned on Superrare.People can imagine the various extensions of multi -mode smart body architecture.—
Some smart projects worthy of attention
(Note: Not all projects are based on LLM + some may be more loose in the concept of smart bodies)
Aiway Finder——Coloning, contract, contract standards, assets, functions, API functions, routine + path decentralized knowledge maps (that is, the blockchain ecosystem virtual roadmap can be navigated).Users will be rewarded for identifying the feasible path used by the intelligent body.In addition, you can cast a shell containing character settings and skill activation (ie, smart), and then you can insert it into the path -seeker’s knowledge map.
Ritualnet——Inden as shown in the Frenrug example above, the Ritual Infernet node can be used to set up a multi -intelligent body architecture.Node monitoring on the chain or under the link, and provides an output with optional certificates.
Morpheus——The personal AI point -to -point network can represent users’ execution of smart contracts.This can be used for web3 wallets and TX intended management, data analysis, DAPPS and contract recommendation models through chat robot interfaces, and long -term memory expansion intelligent gymnastics operations by connecting applications and user data.
Dain Protocol——Everaging multiple use cases deployed on Solana.Recently demonstrated a deployment of an encrypted trading robot. The robot can extract information on the chain and the under -linage information to represent the user’s execution (for example, if Bayon loses, sell Boden)
Naptha——Sminton arrangement agreement, an operator node with a chain on the chain of the smart body, the arrangement of the arrangement task, the LLM workflow arrangement engine that supports the asynchronous message transmitting across different nodes, and the workflow certification system for verifying the execution.
Myshell——It similar to http: // Character.ai AI -character platform, creators can monetize smart configuration files and tools in them.Multi -mode infrastructure contains some interesting examples of intelligence, including translation, education, companionship, coding, etc.Contains simple development personnel models for the creation of simple code -free intelligent bodies and for assembling AI components.
Ai Arena——The competitive PVP fighting game, players can buy, train and confront NFT that supports AI.Players learn how to learn how to play games in different maps and scenes by imitating learning to train their smartborn NFT by learning the relevant probability of player behavior.After training, players can send their smart bodies to participate in the ranking to get token awards.It is not based on LLM, but it is still an interesting example of the possibility of smart games.
Virtuals Protocol—— A protocol for building and deploying multi -modal intelligence to games and other online spaces.The three main prototypes of today’s virtual include IP character mirrors, specific functional intelligence, and individual substitutes.The contributor contributes data and models to virtual virtual, and the verified as the goalkeeper.There is a economic incentive mechanism to promote development and monetization.
Brianknows—— Provide users with user interface so that interaction with smart parties can execute transactions, research information specific to cryptocurrencies, and deploy smart contracts in a timely manner.At present, more than 10 of more than 100 integrated operations are supported.A recent example is to allow the smart body to use natural language to represent users in LIDO ETH.
Autonolas—— Provide lightweight local and cloud -based intelligence, decentralized intelligence and professional intelligent physical economy.Outstanding examples include DEFI and predictive intelligence, representatives of governance by AI, and smart-to-smart (Agent-To-Agent) tool market.A protocol for coordinating and motivating smart gymnastics + OLAS stack, this is an open source framework for developers to build a smart body that can be jointly owned.
Creator.bid—— Provide users with social media role smartmen connected to the real -time API with X and Farcaster.Brands can start knowledge -based intelligence and perform the same content as the brand on social platforms.
Polywrap—— Provide various smart products, such as Indexer (Farcaster’s social media intelligence), AutoTX (planning and transaction executions with Morpheus and FLOCK.IO), PredictionProphet.ai predictions (with GNOSIS and AutonolasSmart) and FundPublicgoods.AI (intelligent body for allocation of resources).
Verification -Since the economic flow will be guided by intelligence, the output verification will be very important (the subsequent articles will be introduced in detail in future articles).Verification methods include ZKML, gaming theory solutions from ORA Protocol, ZKML, game theory solutions, and hardware -based solutions like Tee.
Some of the thoughts on the chain
-
Smarts that can have, tradable, and tokens, can perform various types of functions, from companionship to financial applications,
-
It can represent the intelligence of you identify, learn, and participate in the gaming economy; it can also be used as an autonomous intelligence of players in collaboration, competition, or fully simulating the environment in the environment.
-
Smarts that can simulate real human behavior, for income opportunities
-
Multi -intelligent wallets managed by multi -intelligence can serve as independent asset managers
-
AI managed DAO governance (such as token commission, proposal creation or management, process improvement, etc.)
-
Use web3 storage or database as a combined vector embedding system for sharing and permanent memory status
-
Local operations, participate in the global consensus network, perform user -defined tasks
-
Existing and new agreement interaction and API knowledge map
-
Autonomous guardian network, multiple signature security, smart contract security and function enhancement
-
Really investing in DAO (for example, DAO, a collector of art historians, investment analysts, data analysts, and DEGEN intelligence characters))
-
Token Economics and Contract Security Simulation and Test
-
General intent to manage, especially when the encrypted user experience (such as bridge or DEFI)
-
Art or experimental project
Attract the next one billion user
As Varaint Fund co -founder Jesse Walden recently said, independent intelligence is an evolution of the use of blockchain, not a revolution: we already have protocol task robots, sniper robots, MEV searchters, robot tool packs, etc.Smart is just the extension of all this.
Many areas of encryption are constructed in a way that is conducive to the execution of intelligence, such as games and DEFI.Assuming that the cost of LLM is declining compared to the task performance + the increased accessability of the creation and deployment of smart parties, it is difficult to imagine that an AI intelligence will not dominate the link of interaction and become the next one billionth user of encryption.
Reading materials:
AI AGENTS that can bank themselves usering blockchains
The New AI AGENT Economy Will Run on Smart Accounts
A Survey on Large Language Model Based Autonomous Agents
React: Synergizing Reasoning and Acting in Language Models
Generative agents: Interactive Simulacra of Human Behavior
Reflexion: Language Agents with Verbal Reinforcement Learning
Toolformer: Language Models Can Teach Themselves to use Tools
Description, explain, plan and select: interest with laconuage models open-world multi-task agents
Voyager: An Open-ended embodied agent with laconuage models
LLM AGENTS PAPERS GITHUB Repo
