A beginner’s guide to AI

CoinDesk published its pre-conference primer on artificial intelligence on May 4, 2026, timing the piece to land exactly three weeks before Consensus Miami opens its doors on May 26. The article covers terminology from "agentic AI" to "generative models," pitched at crypto-native readers who sense the convergence but lack the vocabulary. That convergence is real, accelerating, and worth understanding on its own terms, not through the lens of Silicon Valley hype cycles.

The Vocabulary Problem

The AI field suffers from a branding affliction. Every six months, a new prefix arrives: generative, agentic, multimodal, foundational. Each term carries genuine technical meaning, but the marketing apparatus strips that meaning within weeks. Generative AI, the category that includes large language models like GPT-4, Claude, and Gemini, refers to systems that produce new outputs (text, images, code, audio) rather than merely classifying inputs. The technology dates to the transformer architecture paper published by Google researchers in 2017, but commercial deployment only began in earnest with ChatGPT's November 2022 launch.

Agentic AI is the 2025-2026 buzzword. It describes systems that can plan, execute multi-step tasks, use external tools, and operate with minimal human oversight. The distinction matters: a generative model answers questions; an agentic system books your flight, files your taxes, and rebalances your portfolio. Anthropic, OpenAI, Google DeepMind, and at least a dozen well-funded startups are racing to ship agentic products. Anthropic's Claude computer-use capability launched in late 2024. OpenAI's Operator followed in January 2025. By early 2026, autonomous agent frameworks from LangChain, CrewAI, and AutoGen had accumulated over 200,000 GitHub stars collectively.

The vocabulary matters for Consensus attendees because the crypto-AI intersection is no longer speculative. It is a $12 billion market segment by some estimates, spanning decentralized compute networks (Render, Akash, io.net), AI token projects (FET, AGIX merged into ASI Alliance in mid-2024), and on-chain agent protocols (Virtuals, AI16Z's ELIZA framework). Understanding what these projects actually do requires understanding what AI actually is.

Generative Models and Their Limits

A large language model is, at its core, a next-token predictor trained on vast corpora of text. GPT-4, released in March 2023, was trained on an estimated 13 trillion tokens. Meta's Llama 3.1 405B, released in July 2024, trained on 15 trillion tokens. The cost of training a frontier model now exceeds $100 million by most credible estimates, with some analysts placing GPT-5's training budget north of $500 million.

These models excel at pattern completion, translation, summarization, and code generation. They fail at tasks requiring true reasoning over novel domains, reliable arithmetic, and consistent factual recall. The "hallucination" problem, where models generate plausible but false statements, remains unsolved as of mid-2026. Mitigation strategies exist (retrieval-augmented generation, chain-of-thought prompting, tool use), but no architecture has eliminated the fundamental issue.

For Bitcoin-adjacent applications, the limitations are critical. A model that hallucinates transaction histories or fabricates block data is worse than useless; it is dangerous. The crypto community learned this early. Projects like Kaito and Arkham Intelligence use AI for on-chain analysis but layer extensive verification systems on top. Trust minimization, the same principle that makes Bitcoin's consensus mechanism valuable, applies equally to AI outputs. You verify. You do not trust.

Agentic Systems and Autonomy

The agentic paradigm shifts AI from reactive to proactive. An agent receives a goal, decomposes it into subtasks, selects tools, executes actions, observes results, and iterates. The architecture typically involves a planning module, a memory system (short-term and long-term), and tool integrations (web browsing, code execution, API calls, file manipulation).

This has obvious implications for financial automation. An AI agent that can read market data, analyze on-chain flows, execute trades, and manage risk represents something genuinely new. Several protocols are building exactly this. Spectral Finance (acquired by Circle in 2024 for an undisclosed sum) built credit-scoring agents for DeFi lending. Autonolas launched an on-chain agent framework with over 1,600 active agents by Q1 2026. The ASI Alliance (formed from the merger of Fetch.ai, SingularityNET, and Ocean Protocol) claims a combined market cap exceeding $5 billion and positions itself as the decentralized alternative to OpenAI's closed ecosystem.

The bull case: AI agents operating on open, permissionless rails (Bitcoin, Ethereum, Solana) create a parallel financial system where autonomous software transacts without human intermediaries. Machine-to-machine payments become trivial. Lightning Network micropayments fuel API calls between agents. The internet of value becomes the internet of autonomous value.

The bear case: agent autonomy without robust alignment guarantees creates systemic risk. A trading agent with a misspecified objective function can drain a treasury in seconds. The March 2025 incident where an AI agent on a Solana-based protocol executed $4 million in unintended trades before being halted illustrates the danger. Autonomy without accountability is not freedom; it is chaos.

The Decentralized Compute Thesis

Training and running AI models requires enormous computational resources. Nvidia shipped an estimated 3.76 million H100 GPUs in 2024, each costing between $25,000 and $40,000. The supply constraint is real: hyperscalers (Microsoft, Google, Amazon, Meta) have committed over $200 billion in combined capex for 2025-2026, largely for AI data centers.

Decentralized compute networks argue that idle GPUs worldwide represent an untapped supply. Render Network processes GPU-intensive rendering tasks across a distributed node network. Akash Network offers a decentralized cloud marketplace. io.net aggregates GPU supply from data centers, crypto miners, and individual providers, claiming over 500,000 GPUs in its network as of early 2026.

The skeptic's view, articulated by researchers at Stanford's HAI institute, is that latency, bandwidth, and coordination overhead make decentralized training of frontier models impractical. You cannot train a model requiring thousands of tightly synchronized GPUs across nodes scattered worldwide with variable network conditions. The rebuttal from decentralized compute advocates is that inference (running trained models) has different requirements than training, and that inference demand vastly exceeds training demand. A model is trained once; it is queried billions of times.

This mirrors Bitcoin's own history. Early critics said distributed consensus was too slow and wasteful to compete with centralized payment processors. They were technically correct about throughput. They were strategically wrong about what mattered. The value of censorship resistance, permissionlessness, and trust minimization justified the efficiency tradeoff. The same logic may apply to decentralized inference: slightly higher latency in exchange for no single point of censorship or failure.

AI and Monetary Sovereignty

Here is where the story connects most directly to what Bitcoin represents. Artificial intelligence is not neutral infrastructure. It is shaped by whoever controls the training data, the compute, and the deployment policies. OpenAI's content policies determine what 200 million weekly users can and cannot generate. Google's Gemini refuses certain financial queries. Anthropic's Claude operates within usage policies set by a single San Francisco company.

When AI systems increasingly mediate financial decisions, credit scoring, insurance underwriting, tax optimization, investment advice, the entity controlling the AI effectively controls the economic agency of its users. This is a monetary sovereignty question dressed in technological clothing. If your AI agent cannot execute a transaction because a policy filter flags it, your economic freedom is constrained by a corporate content moderation team, not by law, not by market forces, but by an opaque neural network's classification boundary.

Bitcoin's value proposition has always been the separation of money from state. The next decade's fight may be the separation of economic agency from platform. Open-source AI models (Llama, Mistral, DeepSeek) running on decentralized compute, transacting over Bitcoin and Lightning, represent a coherent vision of sovereign economic activity. The user controls the model, the compute, and the money. No single entity can censor, deplatform, or surveil the entire stack.

This is not a utopian fantasy. It is an engineering roadmap with identifiable milestones. DeepSeek's R1 model, released in January 2025, demonstrated that a Chinese lab with reportedly $6 million in training costs could produce reasoning capabilities competitive with models costing 100x more. The cost curve is collapsing. Running a capable local model on consumer hardware (an M4 Mac with 128GB RAM handles 70B-parameter models comfortably) is already practical in 2026.

What to Watch

Consensus Miami 2026, running May 26-28, will feature at least twelve AI-focused panels based on the preliminary agenda. Three developments deserve attention over the next six months.

First, the Bitcoin-native AI agent standard. Several teams, including those building on Lightning (LangChain's Bitcoin plugin, Fedi's federated AI module), are working toward a standard for agent-to-agent payments over Lightning. If a credible standard emerges by Q4 2026, it could position Bitcoin as the default settlement layer for machine-to-machine commerce, a TAM that dwarfs human-to-human payments.

Second, regulatory clarity on AI agents in financial services. The SEC's March 2026 concept release on "autonomous digital advisors" signals that the agency views AI agents executing trades as potentially falling under investment advisor regulations. The EU's AI Act, fully enforceable from August 2026, imposes transparency and audit requirements on high-risk AI systems, a category that explicitly includes credit scoring and financial decision-making. Jurisdictional arbitrage will intensify.

Third, the open-source versus closed-source compute war. Meta continues releasing Llama models openly. OpenAI remains closed. The market will vote. If open models reach parity with closed models (DeepSeek's trajectory suggests this is plausible by late 2026 or early 2027), the decentralized compute thesis becomes dramatically stronger. You cannot build a censorship-resistant AI stack on top of a model controlled by a single company's API terms of service.

The convergence of AI and Bitcoin is not a marketing narrative invented for conference panels. It is a structural inevitability. Both technologies address the same fundamental question: who controls the systems that govern economic life? The centralized answer is familiar: governments, corporations, platforms. The decentralized answer is still being built. Consensus Miami will not resolve the question, but it will reveal which teams are building seriously and which are merely tokenizing the hype.

---

Source: CoinDesk