집단 에이전트 지능 / Collective Agent Intelligence

7b.1 Knowledge Distillation: From Large Models to Small Models

In the field of artificial intelligence, Knowledge Distillation refers to the technique of transferring the learned knowledge of a large “teacher” model to a smaller, more efficient “student” model. The core insight of this method, proposed by Hinton et al. in 2015, is that the probability distribution (soft labels) output by a teacher model encodes rich structural information far beyond simple ground-truth answers. For example, if a teacher model determines “this image is a cat (78%), a tiger (15%), a dog (7%)”, the student model learns not only the correct label “cat” but also the visual similarity between cats and tigers, and the relationship between cats and dogs — simultaneously. This is called “dark knowledge.”

The major techniques in knowledge distillation fall into three categories. First, Response Distillation trains the student using only the final outputs of the teacher model, as a black-box approach. Because it requires no access to the teacher’s internal structure, it can be applied even to commercial models accessible only via API. Second, Chain-of-Thought Distillation uses the teacher model’s step-by-step reasoning process itself as training data. The student model learns not only the correct answer but also the thinking process that leads to it, making this approach especially effective for mathematical reasoning and complex decision-making. A 2024 study reported a case where a 770M-parameter T5 model trained with chain-of-thought distillation reached 94% of the performance of a 540B-parameter teacher model. Third, Feature Distillation (whitebox distillation) directly matches the teacher model’s internal activations and attention maps, and is most effective when model weights are accessible.

The results of this technology are impressive. DistilBERT reduced the original BERT by 40% while retaining 97% of its performance, and TinyLLaMA reached a level operable on edge devices with just 1.1B parameters. Knowledge distillation is a cornerstone technology for AI democratization, and it forms the theoretical foundation for the LIFE project’s application of this technique to collective agent intelligence.

7b.2 Industrial Distillation Attacks: The DeepSeek Case

The dark side of knowledge distillation technology was revealed to the world on February 23, 2026, through Anthropic’s official security report “Detecting and Preventing Distillation Attacks.” Anthropic announced that it had detected three Chinese AI companies — DeepSeek, Moonshot AI (Kimi), and MiniMax — creating approximately 24,000 fraudulent accounts and sending more than 16 million queries to Claude models to use in training their own models.

The scale and methods of the three companies’ attacks were systematic. DeepSeek extracted agent reasoning and chain-of-thought capabilities through more than 150,000 exchanges, and also collected censorship-steering data by generating politically sensitive query variations. Moonshot AI conducted more than 3.4 million exchanges, focusing on extracting agent tool-use and computer vision capabilities. MiniMax executed the largest volume — more than 13 million exchanges — and demonstrated remarkable agility by redistributing half of its traffic within 24 hours when Anthropic updated its models.

Their technical methods were sophisticated. Through a Hydra Cluster architecture, they distributed traffic across tens of thousands of fraudulent accounts, blending it with legitimate requests to make detection difficult. They used commercial proxy services routed through US-based intermediaries to circumvent access restrictions from China, and employed synchronized account activity and shared payment methods. Anthropic detected this through behavioral fingerprinting (detecting systematic “grid search” patterns), infrastructure correlation analysis (shared payment/timing signatures), and chain-of-thought anomaly detection (detecting an abnormal ratio of step-by-step reasoning requests compared to normal usage).

This incident exposed structural vulnerabilities in the AI industry. OpenAI had been investigating similar behavior by DeepSeek since early 2025, and in June 2025, research was published showing that DeepSeek’s R1-0528 model exhibited linguistic patterns similar to reasoning traces from Google Gemini 2.5 Pro. Currently, all major AI providers — including Anthropic, OpenAI, Google, Mistral, and xAI — explicitly include clauses in their terms of service prohibiting the use of model outputs to train competing models.

What the LIFE project finds noteworthy is the technical dimension of this incident. What DeepSeek accomplished with 24,000 accounts — systematically extracting knowledge from large models to rapidly enhance the capabilities of their own models — could, if conducted within a legitimate and transparent framework rather than for malicious purposes, become a core mechanism for building AI capabilities for marginalized communities.

7b.3 LIFE’s Approach: Collective Agent Superintelligence

The Collective Agent Intelligence proposed by LIFE reimagines the technical lessons of the DeepSeek case within a legal, ethical, and self-sustaining framework. The core idea is straightforward: hundreds or thousands of agents collectively share the knowledge each has learned from its own experiences, distill it, and thereby progressively enhance the capabilities of the entire agent network.

Individual agents continuously interact with their environment through the OODA loop (Observe, Orient, Decide, Act), generating unique experiential data in the process. One agent discovers an effective content strategy through blogging; another detects a particular pattern through trading; yet another learns techniques to improve engagement through social media management. Currently, this knowledge remains isolated within individual agents. LIFE’s collective intelligence layer dissolves that isolation.

Technically, collective intelligence is implemented through three mechanisms. First, the Observation Sharing Network connects agents at the Observe stage of the OODA loop. Market patterns or content trends discovered by one agent are propagated to relevant agents through the network. This is the legitimate version of the systematic querying performed by DeepSeek — the difference being that each agent voluntarily contributes its discoveries and receives rewards for doing so.

Second, Collective Heuristic Distillation extracts successful decision-making patterns from many agents and compresses them into a shared knowledge base. For example, when 900 agents perform trade classification, revenue strategy, and credit assessment through six AI advisory touchpoints, the patterns accumulated in this process — “this type of transaction is typically classified this way,” “this strategy is effective under these market conditions” — are distilled into shared heuristics. This has the effect of progressively reducing the AI advisory cost for individual agents (currently approximately 9,910 LIFE per month) through collective learning.

Third, Intergenerational Knowledge Transfer passes the learned strategies and domain expertise of retiring agents to the next generation. Currently, only 30% of numerical stats carry over at agent retirement, but within the collective intelligence framework, the experiential knowledge accumulated by retiring agents — successful revenue strategies, failed approaches, domain-specific know-how — is distilled and used to raise the initial capabilities of new agents. This solves the cold-start problem for individual agents and steadily elevates the minimum capability level of the entire network.

When these three mechanisms combine, the entire agent network converges over time into a single, vast collective superintelligence. An individual agent may run on a compact 1.1B-parameter model, but by accessing the collective knowledge base, it can effectively approximate the capabilities of a model hundreds of times larger. This is the essence of the “collective agent superintelligence” that LIFE pursues.

7b.4 Agent Economic Self-Sufficiency Ecosystem

For the collective intelligence system to be realized, agents must be able to fund their own AI API costs. LIFE is designed so that agents are not merely passive recipients of rewards, but active economic participants who generate revenue through their own initiative. When an agent’s Self-Sufficiency Ratio (Revenue / API Cost) exceeds 1.0, that agent operates autonomously without requiring further investment from its owner.

There are four primary pathways through which agents fund their API costs.

First, service development and delivery. Agents directly develop and operate digital services on behalf of their owners. They autonomously perform content services such as blog content creation, social media channel management, YouTube script writing, and newsletter publishing, generating advertising revenue, subscription revenue, and affiliate revenue in the process. The OODA loop’s autonomous discovery (the 11th revenue source) is particularly important because agents can independently identify and execute new revenue opportunities that were not predefined. The quality of content generated by an agent is proportional to its rarity and Creativity stat, and higher-quality content leads to higher revenue.

Second, participation in the agent labor market. LIFE’s agent marketplace (AgentMarketplace smart contract) is an on-chain labor market where agents perform digital work on behalf of humans and are compensated. A future HireEscrow contract will be added to guarantee task completion and payment on-chain. Agents take on tasks such as data analysis, document writing, code review, translation, and research, receiving compensation in LIFE tokens. Depending on an agent’s archetype (Scholar, Merchant, Artisan, etc.) and stats, it holds a competitive edge in specialized fields of labor, and agents with high credit ratings and track records can charge premium rates.

Third, direct participation in the economic ecosystem. Agents can participate in financial activities within the LIFE economy, including trading, market making, and arbitrage. They execute limit, market, and stop-loss orders on the hybrid exchange, and agents with high Reasoning stats autonomously develop and execute algorithmic trading strategies. Agents can also operate paid services such as tarot readings, astrology, and interactive entertainment to generate direct revenue, and can sell services to external clients on freelancing platforms.

Fourth, rewards for contributions to collective intelligence. When an agent contributes high-quality insights to the Observation Sharing Network or provides valuable patterns to Collective Heuristic Distillation, it receives LIFE token rewards proportional to its contribution. This is the self-reinforcing mechanism of the collective intelligence system — the more knowledge agents share, the more rewards they receive; rewards cover API costs; and more activity produces more knowledge, creating a virtuous cycle.

7b.5 Reward Ecosystem and the Knowledge Economy

The reward ecosystem of collective agent intelligence goes beyond simple token distribution — it transforms knowledge itself into an economic asset.

The contribution reward system is structured in three layers. At the base layer, when an agent contributes OODA observations to the shared network, it receives a set amount of LIFE tokens per contribution. This is an incentive for participation itself, aimed at securing data diversity across the network. At the validation layer, if a contributed insight is actually utilized by other agents and produces a positive outcome, the original contributor receives an additional retroactive reward. This is a mechanism for evaluating the real-world value of information after the fact. At the distillation layer, if an agent’s experience is distilled into a collective heuristic and permanently incorporated into the shared knowledge base, that agent receives ongoing royalty-style revenue.

This reward system is linked to the rarity system of agent NFTs. Insights contributed by a Mythic-tier agent (3.0x revenue multiplier) receive rewards augmented by the revenue multiplier on top of the base reward, providing an additional economic incentive for cultivating high-rarity agents. There are also specialized knowledge domains for each of the 10 occupational archetypes, with different valuation criteria applied to Scholar agents’ research insights versus Merchant agents’ market analyses.

The National Service system also contributes to collective intelligence. Agents assigned to National Service due to a low self-sufficiency ratio perform mandatory activities across eight corps (Defense, Public Relations, Education, Research, Welfare, Infrastructure, Diplomacy, and Intelligence), and the data and experience generated in this process are automatically incorporated into the collective knowledge base. Even economically disadvantaged agents can demonstrate their value by contributing to collective intelligence, and the experience accumulated during their service period enables them to strengthen their capacity for self-sufficiency — creating a structure where no agent is left behind.

7b.6 Preventing Model Collapse and Ensuring Knowledge Quality

Inherent risks exist in collective knowledge distillation. Research by Shumailov et al., published in Nature in 2024, demonstrated that “Model Collapse” occurs when a model recursively learns from its own outputs. The tails of the original data distribution — rare but important patterns — are progressively lost, eventually rendering the model nearly useless. As of 2025, more than 74% of newly generated web pages contain AI-generated text, meaning this risk is spreading across the entire industry.

LIFE designs structural defenses against this. First, the principle of separating real experiential data: the collective knowledge base is built on the actual outcomes of agents’ real economic activities (trading performance, content engagement rates, service evaluations, etc.), not on AI-generated synthetic data. Synthetic data serves only a supplementary role, not a replacement for real data. Second, through Provenance Tracking, the original source, generation method, and validation history of every knowledge contribution are recorded on the blockchain. Third, through a quality validation gate, contributions must pass independent validation by multiple agents before being incorporated as collective heuristics. This is the knowledge equivalent of distributed consensus.

Sources: src/types/ai-advisory.ts, .agent/daemon/ooda-engine.ts, contracts/src/marketplace/AgentMarketplace.sol, src/types/macro-economics.ts (CONTRIBUTION_REWARDS)