Episode Summary
Executive Summary: This episode explores the evolution of crypto AI from early infra and compute investments to today’s agent boom. Tom Shaughnessy argues the sector’s real edge is open-source, utility-driven, and long-tail applications rather than reply bots. The panel debates value capture, open vs. closed models, talent migration from Silicon Valley, and where crypto’s composability, incentives, and capital formation can matter most.
Main Topics: Origins of Crypto AI investing (Priority: 5/5): Tom explains Delphi began investing over two years ago, initially facing internal skepticism, driven by early experiences with MidJourney/ChatGPT and a conviction that AI was the next major tech wave. Open source vs. closed source AI (Priority: 5/5): The discussion centers on whether OpenAI/Anthropic/Meta will retain frontier-model value or whether open-source ecosystems—amplified by crypto—will dominate applications and experimentation. The wave evolution of crypto AI (Priority: 5/5): They trace the sector through phases: liquid projects like BitTensor, infra/compute networks, coordination networks, then the current agent and model wave that consumers actually see. Value accrual and the long tail (Priority: 4/5): The hosts argue crypto AI may not win the frontier model race but can capture value in the long tail of narrow, monetizable applications such as trading agents, commentators, and task-specific bots. Talent, incentives, and migration from Silicon Valley (Priority: 4/5): A key theme is how high pay and prestige keep AI talent in closed labs today, but open-source freedom, funding mechanisms, and faster experimentation could pull builders into crypto AI later. Frameworks for evaluating AI crypto deals (Priority: 5/5): Tom outlines a practical investment lens: proprietary models, proprietary data, business integrations, framework/tooling advantages, and inference-time compute differentiation. Agent tooling and internal AI workflows (Priority: 3/5): The conversation closes with practical examples of how they personally use different models for coding, real-time data, search, and voice-driven assistants, plus Delphi’s internal agentic deal-review workflow.
Key Arguments: Crypto AI started with infra and compute because early AI-capable infrastructure was expensive and inaccessible; crypto’s distributed coordination made it relevant. Open-source AI is attractive because it reduces central control risk; the fear is not just model quality, but one entity controlling downstream consumer behavior and institutions. Closed-source labs still lead on frontier models due to GPUs, talent, data, and capital, but the gap is narrowing as open-source techniques replicate faster. The market has already shown efficiency: reply bots/slot bots are losing value quickly, indicating crypto can prune bad ideas faster than traditional tech markets. Crypto AI’s strongest niche may be the long tail of narrow, monetizable applications rather than a single “winner-take-all” frontier model. Tokens are useful mainly when they solve fundraising, coordination, incentives, and distribution, but many AI tokens currently lack strong design or product necessity. A compelling crypto AI project should have at least one of: proprietary model, proprietary data, business integration, framework/tooling edge, or inference-time compute advantage. Talent migration from Silicon Valley will likely accelerate when closed-model pay premiums compress or when open-source/crypto offers better freedom and upside. Internal AI tools can materially improve venture diligence, especially for founder checks, document analysis, and pattern recognition, but they need careful prompting and curation.
Data Points: Delphi crypto AI investing duration: Over 2 years - Tom says Delphi has invested in crypto AI for more than two years. Number of crypto AI investments: 20 - Tom states Delphi made 20 crypto AI investments in the last two years. BitTensor role: Started the liquid side of the space - Tom credits BitTensor as the first liquid crypto AI project to kick off the sector. Model size (News Research): 15 billion parameters - Tom cites News Research as training a 15B parameter model. Model size (Prime Intellect): 10 billion parameters - Tom cites Prime Intellect as training a 10B parameter model. Communication reduction: 2,000 to 3,000 times less communication per node - Tom references an efficiency gain in distributed training approaches like those at News Research. MedPalm improvement: ~60-65% to ~95% accuracy - Ejaz uses Google MedPalm as an example of narrow models improving dramatically with expert data. OpenAI valuation: Near $200B - Used as evidence that closed-source labs remain highly valued. Claude valuation: $60B to $80B - Mentioned while comparing frontier AI lab valuations. Perplexity valuation: ~$15B - Referenced as another high-value closed-source AI company. Uniswap all-time volume: $2.75T - Sponsor mention highlighting Uniswap’s scale. Arbitrum ecosystem apps: 800+ apps - Sponsor mention for the Arbitrum portal. Morpho Coinbase integration: First and only DeFi protocol integrated by Coinbase - Sponsor mention illustrating on-chain lending adoption.
Pivotal Quotes: "I do not want to live in a world where Sam Altman controls the same single model that controls our life." — Tom Shaughnessy: Tom explains his core motivation for backing open-source crypto AI. "The space wants utility-based functional agents that are doing things." — Tom Shaughnessy: He argues the market is moving away from spammy reply bots toward real-use agents. "Crypto AI is the long tail, the periphery of everything." — David Hoffman: David frames crypto AI as the applied, monetizable fringe of the broader AI market.
Implications: Listeners should expect crypto AI to shift from hypey agent spam toward useful, specialized products built on open data, tooling, and composable money rails. The biggest upside may come from long-tail applications and infrastructure, not frontier model supremacy.