Predictive AI for Web3: Growth and Security Without LLM Wrappers

X Space with Plena Finance — ChainAware co-founder Martin joined Plena Finance’s PlanarPod to discuss predictive AI for Web3 growth and security. Listen to the full X Space recording ↗

Predictive AI is the most misunderstood concept in Web3 today. Most projects claiming to use AI are wrapping OpenAI’s API with a few lines of prompt. They call it an AI agent. They call it intelligent. However, they cannot predict fraud, cannot power a marketing agent, and cannot detect a rug pull — because large language models are not designed for any of these tasks. In this X Space with Plena Finance, ChainAware co-founder Martin explains exactly what predictive AI is, why it is fundamentally different from LLMs, how blockchain data makes it uniquely powerful, and what the two most important use cases are for every Web3 project building for sustainable growth.

Attention AI vs Real Utility AI: The Market Cleanup That Had to Happen

The X Space opens with a candid market assessment. Martin’s view is direct: the AI agent market is flooded with projects grabbing attention without delivering utility. These projects — “attention AI” in ChainAware’s terminology — label themselves as AI-driven, follow the narrative, and raise capital based on trend-surfing rather than genuine technical capability. Their future is short. The correction that followed the AI hype wave was, in Martin’s words, not just expected but required.

“It’s a cleanup phase,” Martin explains. “People have to learn what real AI is, and that attention AI — projects just creating attention — have to face reality.” This framing matters for anyone evaluating Web3 AI projects. When almost every project in a category claims to use AI, the meaningful question shifts from “does it use AI?” to “does it use AI that produces measurable, verifiable results that competitors cannot easily copy?”

ChainAware’s answer to that question is a proprietary predictive AI stack — neural networks trained on blockchain behavioral data, not wrappers around OpenAI’s API. The distinction has technical and commercial consequences that Martin walks through in detail throughout the session. For more background on how this distinction maps to the broader Web3 AI landscape, see our guide on attention AI vs real utility AI and our complete breakdown of real AI use cases for Web3 projects.

LLMs vs Predictive AI: The Critical Technical Difference

The most technically important section of the X Space is Martin’s explanation of why LLMs cannot do what predictive AI does. This is not a marketing claim — it is a structural architectural fact, and understanding it changes how you evaluate every AI-claiming project in Web3.

What LLMs Actually Are

Large language models are statistical autoregression engines. They process enormous text datasets and learn to predict the most probable next word — or sequence of words — given a preceding input. They do not understand meaning. They do not reason about causality. They produce statistically likely output based on patterns in training data. As Martin describes it: “LLMs don’t understand what they are saying. They’re just producing output. It’s statistical autoregression based on the previous input.”

This architecture makes LLMs excellent at certain tasks: generating text, summarizing content, writing smart contract code templates, answering knowledge-based questions, creating Twitter posts. Consequently, content generation, code generation, and chatbot tasks suit LLMs well. Furthermore, the barrier to building LLM-based products is low — three to five lines of prompt code, an API key, and a website. This is precisely why 95% of “Web3 AI” projects are LLM wrappers. They require minimal technical investment, which also means they create no defensible competitive advantage.

What Predictive AI Does Instead

Predictive AI uses dedicated neural networks trained on domain-specific labeled data to forecast future events and behaviors. Rather than predicting the next word in a sequence, it predicts the next action of a wallet address, the probability that a contract will rug pull, or whether a user is likely to borrow, lend, trade, or buy NFTs. These predictions carry measurable accuracy scores, can be backtested against historical data, and improve continuously as the models retrain on new behavioral data.

Critically, you cannot substitute an LLM for a predictive model in behavioral tasks. Martin is explicit: “With LLMs you cannot create marketing agents because with LLMs you cannot predict behavior. With LLMs you can generate content. If you want to go into marketing agents or transaction monitoring agents, you have to develop your own models. It’s an unavoidable step.” This means any project claiming to offer behavioral targeting, fraud prediction, or rug pull detection using only LLMs is either mistaken or misleading. For the full technical breakdown, see our article on which AI use cases Web3 projects can actually integrate via API.

Property	LLMs (e.g. ChatGPT, Claude)	Predictive AI (ChainAware)
Core mechanism	Statistical autoregression on text	Neural networks trained on behavioral data
Output type	Text, code, summaries	Probabilities, scores, predictions
Accuracy measurement	Not measurable — may hallucinate	Measurable, backtested (98% fraud, 80%+ behavioral)
Can predict fraud?	❌ No	✅ Yes — 98% accuracy
Can power marketing agents?	❌ No — cannot predict behavior	✅ Yes — predicts then generates
Can detect rug pulls?	❌ No	✅ Yes — before they happen
Build barrier	Low — 3–5 lines of prompt	High — years of model development
Competitive moat	None — easily copied	Strong — proprietary models + training data
Improves over time	No — static model	Yes — continuous retraining on new data
Requires own models	No — API wrapper	Yes — mandatory

Why Blockchain Data Is Better Than Google’s Browsing History

Before explaining how ChainAware’s models work, Martin addresses a question that many people implicitly ask: what makes blockchain data good enough to build predictive models on? The answer is more compelling than most realize — and it comes from a direct comparison to the data Google uses for its advertising system.

Google’s advertising targeting relies on two primary data sources: browsing history and search history. These signals are useful but imprecise. Browsing history captures what pages a person visits — not what they intend to do, how much money they have, or what financial decisions they are considering. Search history is somewhat better as a signal of intent, but it remains ambiguous. Someone searching “crypto lending” might be a professional researcher, a curious student, or an active DeFi user — the signal does not distinguish them reliably.

Blockchain data is fundamentally different. Every transaction on a blockchain reflects a real financial decision that the person made. They thought about the amount, the counterparty, the protocol, the timing, and the risk. They executed the transaction because they intended to. Consequently, blockchain transaction history captures a person’s actual financial intentions and behaviors with extraordinary precision. As Martin explains: “People are thinking before they’re doing financial transactions. Because they are thinking, this part of their thinking leaves patterns on the blockchain.”

The Signal Quality Advantage

Additionally, blockchain data is permanent and tamper-proof. A browsing history can be cleared, a VPN can mask it, and cookies can be blocked. An on-chain transaction history cannot be altered or hidden. Every protocol interaction, every borrowing position, every yield farming deposit, every NFT purchase — all permanently recorded, publicly accessible, and freely available for analysis.

This combination of permanence, financial significance, and public availability makes blockchain data a uniquely powerful foundation for behavioral prediction models. Furthermore, the data is free — no licensing fees, no data partnerships required. Any organization building predictive models on blockchain data starts with a dataset that is higher quality than what Google uses for AdWords, at zero marginal data cost. For a deeper look at how ChainAware uses this data across 8 blockchains, see our Web3 behavioral user analytics guide.

How Predictive AI Models Actually Work

Martin provides a clear, non-technical explanation of how ChainAware’s predictive models are built — covering the training process, the data labeling approach, and why the result is meaningfully different from anything achievable with LLMs.

The core process involves training dedicated neural networks on two categories of labeled behavioral data. The first category is “good behavior” — wallet addresses with established histories of legitimate DeFi activity: responsible borrowing, protocol participation, normal trading patterns. The second category is “bad behavior” — wallet addresses associated with confirmed fraud, scam activity, hacking, phishing, and rug pull execution. These bad-actor addresses left behavioral patterns on-chain in the weeks and months before their attacks. Those patterns become the training signal for detecting future fraud.

The resulting model does not check a list of known bad addresses — it identifies behavioral patterns that match the pre-fraud signature, even for addresses that have never been flagged before. This is why ChainAware’s fraud detection is genuinely predictive rather than reactive. A conventional AML system checks whether an address appears on a sanctions list. ChainAware’s system predicts whether a new address is exhibiting the same behavioral patterns that confirmed fraudsters exhibited before they committed fraud. For more on the distinction between these approaches, see our guide on crypto AML vs transaction monitoring and our analysis of forensic vs AI-based blockchain analytics.

Predictive Fraud Detection: 98% Accuracy Before It Happens

ChainAware’s fraud detection model was launched on February 4, 2023 — at the time of this X Space, exactly two years prior. The 98% accuracy figure is not self-reported but backtested against CryptoScamDB, an independent database of confirmed crypto scam addresses and events. This means the model correctly identified fraudulent behavior 98% of the time on data it had never seen during training.

The practical implications for DeFi protocols, NFT platforms, and Web3 projects are significant. Consider the standard approach to security: deploy smart contract audits, display security badges, and hope that sophisticated attackers don’t find exploits the auditors missed. This approach addresses code-layer vulnerabilities but ignores the most common attack vector — malicious users who interact with otherwise secure protocols.

Stopping Fraud at the Wallet Connection Event

ChainAware’s fraud detection integrates at the wallet connection event — the moment a user connects their wallet to a DApp, before any transaction occurs. At that moment, the model scores the connecting wallet address for fraud probability. If the score exceeds the configured threshold, the platform can block the connection, shadow-ban the user, apply tiered restrictions, or trigger an alert to the compliance team — all before any damage can occur.

Moreover, the model operates continuously. A wallet that passes a fraud check today might develop suspicious behavioral patterns over the following weeks. ChainAware’s transaction monitoring agent watches for these changes and sends real-time notifications when a previously clean wallet begins exhibiting pre-fraud signals. This combination — predictive screening at entry plus continuous monitoring — is the complete security picture that blockchain transactions’ irreversibility demands. For the full technical integration guide, see our AML and transaction monitoring integration guide and the Transaction Monitoring Agent guide.

Rug Pull Detection: The 95% PancakeSwap Reality

Martin introduces one of the most striking statistics in the X Space: approximately 95% of liquidity pools on PancakeSwap end in a rug pull. This is not fringe activity — it is the overwhelming norm for new token launches on one of the largest DeFi platforms in the world. Consequently, the default outcome for a new token investor on PancakeSwap is a total loss.

The mechanism is social engineering at scale. Scam factories — organized groups running coordinated fraud operations — create Telegram groups that attract new Web3 users. These newcomers receive “tips” about promising new tokens, join the channels, buy the tokens, and almost always lose everything when the liquidity is pulled. Martin is direct: “The buyers are the newbies. They’re getting social engineered into these Telegram channels and buying. In 95% of cases you end up with rug pulls.”

What Rug Pull Prediction Looks Like

ChainAware’s rug pull detection model analyzes the contract address, the liquidity pool structure, the creator wallet’s behavioral history, and trading patterns to predict whether a contract will execute a rug pull. Importantly, this is prediction — not a rules-based checklist. The model identifies behavioral signatures that confirmed rug-pull contracts exhibited before the event, not just known vulnerability patterns in the code.

At the time of the X Space, the model predicts whether a rug pull will occur but not precisely when. Martin acknowledges this as a future development goal: “We don’t have the ability to say when the rug pull happens. Maybe in the future we will build our model so we can see the timeframe when it will happen.” This planned capability — not just detecting rug pull probability but predicting timing — would give investors and platforms an even more actionable early warning system. Free rug pull checks are available at chainaware.ai/rug-pull-detector. For the complete methodology, see our Rug Pull Detector guide and our guide on how to identify fake crypto tokens.

Marketing Agents: One-to-One Personalization Without Cookies

The second major application of ChainAware’s predictive AI — alongside fraud and rug pull detection — is behavioral marketing. Martin describes the current state of Web3 marketing as “1930s style”: every visitor to every Web3 platform sees the same message, regardless of who they are, what they have done on-chain, or what they are likely to do next. The result is low conversion, low engagement, and enormous user acquisition costs.

ChainAware’s marketing agents change this by predicting each connecting wallet’s behavioral profile and generating personalized content that resonates with their specific intentions. The process has two stages that work in sequence.

Stage One: Behavioral Prediction

At wallet connection, the predictive model analyzes the wallet’s complete on-chain history and calculates its behavioral profile. Some wallets are active borrowers. Others are NFT collectors, yield farmers, high-frequency traders, cautious long-term holders, or complete newcomers. Each profile implies different intentions, different risk tolerance, and different messaging that will resonate. The model does not ask the user to fill out a preferences form. It reads their blockchain history and infers their profile automatically.

Stage Two: Content Generation

Once the behavioral prediction is complete, the generative layer creates personalized content matched to that profile. This is where generative AI (including LLMs) appropriately enters the process — not to predict behavior, but to generate text that speaks to the predicted intentions. A yield-farming-oriented wallet visiting a lending protocol sees messaging about yield optimization opportunities. A newcomer wallet sees explanatory content that reduces friction. A high-experience DeFi user sees advanced product features. Each message targets that specific user’s likely next action.

Martin describes the outcome: “Web3 visitors start to see resonated content. It’s not one content for everyone — it’s one-to-one content for everyone. People start to get much more attached to the websites using our technology. Because this resonance creates attachment.” Higher attachment leads to longer sessions, higher conversion, and better retention — the metrics that determine whether a Web3 project can build sustainable revenue. For a detailed walkthrough of how this works in practice, see our Web3 behavioral user analytics guide, the guide to personalization as the next big AI agent opportunity, and the SmartCredit case study showing 8x engagement and 2x conversions.

The Google AdTech Parallel: Why Web3 Is Where Web2 Was

Throughout the conversation, Martin returns to a historical parallel that explains both the opportunity and the urgency for Web3 marketing transformation. Before Google invented AdWords, online advertising was mass broadcasting. Companies bought banner ads, newspaper placements, and billboard space to drive traffic. Cost per acquisition was extremely high. Conversion was extremely low. The economics barely worked, even for well-funded companies.

Google changed everything by using search history to predict user intent and match ads to that intent at the moment of expression. The result was a dramatic collapse in cost per acquiring a transacting user — from hundreds of dollars to $15–$30 in mature markets. This enabled the online economy to scale. Companies that adopted Google AdWords early gained compounding advantages in growth economics over competitors that remained on broadcast advertising.

Web3 is in the pre-AdWords phase right now. Every project uses the same broadcast approach: KOLs, airdrops, Telegram groups, crypto ad networks. Conversion is abysmal. User acquisition costs are $1,000–$3,000 per transacting user in DeFi. The unit economics are structurally negative for almost every protocol.

Blockchain History as a Superior Signal

ChainAware’s behavioral targeting uses blockchain transaction history as the targeting signal — which Martin argues is actually superior to search history. “Google uses the browsing history and the search history to predict behavior. In Web3, we have a blockchain. All the data is there. We predict behavior based on blockchain history.” The deliberate, financial nature of blockchain transactions means this signal carries far more predictive weight than browsing data. Consequently, the targeting precision available to Web3 projects that adopt behavioral AI is higher than what Google achieved with search data — if they build the infrastructure to use it. For more on how this transforms user acquisition economics, see our guide on ChainAware’s AI agent roadmap for Web3 businesses.

AI Use Cases in DeFi, NFTs, and Portfolio Management

Beyond fraud detection, rug pull prediction, and marketing agents, Martin covers several additional AI use cases that are emerging in DeFi and the broader Web3 ecosystem. Each one represents an existing DeFi function where AI-driven decision making adds measurable value over human judgment or static rules.

Trading Agents

Trading agents represent the most discussed AI use case in crypto. Martin draws on his experience at Man Investments — the largest independent hedge fund in the world at the time, managing $20 billion in automated trading systems two decades ago — to contextualize the competition. Professional algorithmic trading is not a new field. Retail traders competing with institutional systems face very high odds against them. AI-based pattern recognition can improve those odds by identifying market structure patterns that rules-based systems miss, but the competitive bar is already extremely high. Additionally, most current Web3 “trading bots” are rules-based, not AI-based — the “AI” label is often applied to if/then logic that predates machine learning.

Portfolio Optimization Agents

Portfolio optimization agents address a different and arguably more accessible problem: risk-adjusted asset allocation. People have fundamentally different risk tolerances. Some accept high volatility for high expected returns. Others need stability. A portfolio optimization agent continuously monitors a user’s holdings, rebalances when allocations drift from target ratios, and applies risk management logic appropriate to the user’s profile — without requiring the user to manually track positions across multiple protocols. This is effectively an automated version of the wealth management service that high-net-worth clients receive at banks like Credit Suisse, now accessible to any Web3 user.

NFT Market Intelligence

In the NFT sector, predictive AI applies to collection valuation, trait rarity scoring, and market trend prediction. Rather than relying on floor price as the primary signal, behavioral models can analyze trading patterns, wallet holder profiles, and collection liquidity dynamics to provide more nuanced intelligence about asset value and market direction. As NFT markets mature beyond speculative peaks, these tools become increasingly relevant for collectors and platforms managing inventory risk.

Smart Contract Review: Where LLMs Hit Their Limit

Smart contract review is a use case where Martin explicitly acknowledges the role of AI tools — while also identifying a hard technical ceiling on what LLMs can achieve. Using AI tools for initial smart contract screening is genuinely useful. LLMs can identify common vulnerability patterns: reentrancy issues, integer overflow risks, access control gaps. For preliminary pre-audit screening, they accelerate the process and reduce the manual work required of human auditors.

However, the ceiling is real. “Hackers are very advanced,” Martin notes. “You can do some pre-screening, but there is again a limit to how far LLMs can go. LLMs do not understand what they are doing — it’s statistical autoregression.” The most sophisticated exploits — flash loan attacks, oracle manipulation, MEV extraction — arise from the interaction between contracts and external real-time conditions that no static code analysis can reliably anticipate. As a result, AI-augmented auditing improves baseline security but does not replace expert human review for high-value protocols.

Furthermore, the more critical security layer is what happens after deployment — who interacts with the contract. A perfectly audited contract can still be exploited if it serves malicious users whose wallets have been pre-screened as fraudulent by ChainAware and ignored. Address-level screening complements code-level auditing. Neither alone is sufficient. For the detailed argument on why address monitoring matters more than contract monitoring for preventing major hacks, see our complete DApp AML and transaction monitoring guide.

The Two Core Problems Every Web3 Project Must Solve

Martin’s conclusion in the X Space distills everything into a simple framework that applies to every Web3 project regardless of category. Two problems determine whether a project survives long enough to become sustainable. Solve both and the project has a viable growth trajectory. Fail at either and the project faces structural headwinds that no amount of token incentives or marketing spend can overcome.

Problem 1: User Conversion

Every Web3 project needs transacting users — not visitors, not token holders, not Telegram community members. Transacting users who actively use the protocol generate revenue. Without revenue, even well-funded projects eventually close. Getting from visitor to transacting user requires resonance. The user must feel that the platform understands their goals and speaks to their intentions. Currently, almost no Web3 project achieves this because all use mass broadcast messaging. ChainAware’s marketing agents solve this with behavioral targeting at the wallet connection event — the first point where a visitor becomes an identifiable potential user. For more on how to deploy this, see the guide on why 90% of connected wallets never transact and how AI agents fix it.

Problem 2: Trust and Security

The second problem is fraud. Every Web3 project interacts with a user base that contains a meaningful percentage of bad actors — fraudsters, scammers, rug pull operators, and exploiters. Without active screening, these actors cause direct financial damage, reputational damage, and regulatory exposure. ChainAware’s fraud detection and transaction monitoring solve this at the address level — not through code auditing, but through behavioral prediction of the humans behind the wallets. As Martin summarizes: “These are the core issues. User conversion — that’s the marketing. Every Web3 project needs it. And fraud — that’s the other one. These are the core issues.” For a complete view of how both problems connect to the Web2→Web3 growth transition, see our X Space recap on AI agents for Web3 businesses.

Comparison Table: LLMs vs Predictive AI for Web3

Use Case	LLM Suitable?	Predictive AI Suitable?	ChainAware Tool
Fraud detection	❌ No — cannot predict behavior	✅ Yes — 98% accuracy	Fraud Detector
Rug pull detection	❌ No	✅ Yes — before it happens	Rug Pull Detector
1:1 marketing agents	❌ No — needs prediction first	✅ Yes — predict then generate	Growth Agents
Transaction monitoring	❌ No	✅ Yes — continuous behavioral watch	Transaction Monitoring Agent
Credit scoring	❌ No	✅ Yes — 4+ years live	Credit Scoring Agent
Behavioral analytics	❌ No	✅ Yes — 8 dimensions	Web3 User Analytics (free)
Content generation	✅ Yes	Partial (step 2 of marketing)	Growth Agents (combined)
Smart contract code	✅ Yes — with limitations	N/A	—
Smart contract pre-screening	Partial — limited depth	✅ Better — behavioral analysis	Fraud Detector
Portfolio optimization	❌ No	✅ Yes — risk-adjusted rebalancing	Risk Monitoring Agent
Trading signal generation	❌ No — not pattern recognition	✅ Yes — ML on market data	External integration

Frequently Asked Questions

Why can’t LLMs predict fraud or power marketing agents?

LLMs are statistical autoregression models — they predict the most probable next word given a preceding input. They are optimized for language patterns, not behavioral prediction. Predicting whether a wallet will commit fraud requires a neural network trained on labeled good/bad behavioral examples — a fundamentally different architecture with a fundamentally different training process. Feeding blockchain data into an LLM and asking it to detect fraud produces unreliable results because the model was never designed for this task. Marketing agents require behavioral prediction first (to know what the user intends) and content generation second (to craft a message matching that intent). LLMs can only perform the second step. For the full explanation, see our guide on real AI use cases for Web3 projects.

How is 98% fraud detection accuracy measured?

ChainAware’s 98% accuracy is backtested against CryptoScamDB — an independent database of confirmed crypto scam addresses and events. The model is tested on labeled data it has never seen during training. Of wallets the model flags as fraudulent, 98% are confirmed as fraudulent in the ground truth dataset. This is a standard machine learning validation methodology. It is not self-reported performance but independently verifiable against a third-party database. For more on how this compares to traditional AML approaches, see our Fraud Detector complete guide.

What percentage of PancakeSwap pools are rug pulls?

According to ChainAware’s analysis, approximately 95% of PancakeSwap liquidity pools end in a rug pull. This reflects the reality of new token launches on high-volume, permissionless DEX platforms where creating a liquidity pool requires minimal effort and the social engineering infrastructure (Telegram groups, influencer promotion) to attract victims is well-developed. New Web3 users are disproportionately targeted because they lack the experience to identify rug pull indicators. ChainAware’s rug pull detector provides free contract analysis to help users check any pool before depositing at chainaware.ai/rug-pull-detector.

What makes blockchain data better for behavioral prediction than browsing data?

Blockchain transactions reflect deliberate, financially significant decisions. Every on-chain transaction required the user to think about amount, counterparty, risk, and timing — then actively execute the transaction. Browsing history, by contrast, captures passive page visits that may have no intentional significance. Furthermore, blockchain data is permanent and tamper-proof — it cannot be cleared, masked, or manipulated. Additionally, it is freely and publicly accessible with no licensing fees. The combination of high signal quality, permanence, and zero data cost makes blockchain behavioral data uniquely powerful for prediction models.

Does ChainAware use OpenAI or other LLM providers?

No. ChainAware builds its own proprietary neural networks trained on blockchain behavioral data. It does not wrap OpenAI, Anthropic, or any other LLM provider for its core prediction capabilities. The fraud detection model, rug pull detection model, behavioral profiling model, and credit scoring model are all proprietary — developed and trained by ChainAware’s team over multiple years. Generative AI may be used in the second stage of marketing agent content generation, but the critical predictive layer is entirely in-house. This is what creates the competitive moat that LLM-based competitors cannot replicate by switching API providers.

How does ChainAware’s free tier work?

ChainAware offers several free products. The Fraud Detector and Rug Pull Detector are free for individual wallet and contract checks at chainaware.ai/fraud-detector and chainaware.ai/rug-pull-detector — no signup required. The Web3 User Analytics dashboard is free forever for any Web3 project that integrates via Google Tag Manager — showing aggregate behavioral profiles of all connecting wallets across eight dimensions. Enterprise products — marketing agents, transaction monitoring, credit scoring — are subscription-based. See chainaware.ai/pricing for full details.

This article is based on the ChainAware x Plena Finance X Space. Listen to the full recording on X ↗. For questions or integration support, visit chainaware.ai.