Episode #543: The Year of Agents and the Industries Not Ready for Them

In this episode of the Crazy Wisdom Podcast, host Stewart Alsop sits down with Mauro Schilman, CTO and Co-founder of Tuki, the distribution standard for the AI agent era in travel, for a wide-ranging conversation that moves from the joys of international travel and the beauty of mathematics to the fast-evolving world of AI and large language models. Mauro shares his background as a math Olympiad competitor and later a coach, his time training coding models at the AI company Cohere, and his thoughts on how frontier models are progressing — or plateauing — at the foundational level while innovation accelerates at the application layer. The two also get into the mechanics of agentic AI, MCP and agent-to-agent protocols, hierarchical memory systems, red-green test-driven development as a powerful coding workflow, and the philosophical murkiness of open-source AI. They wrap up discussing Tuki Travel's mission to build AI-ready infrastructure for the travel industry, connecting hotels, suppliers, and online travel agencies to prepare for the coming wave of agentic commerce. You can learn more about Tuki Travel and reach out to the team at tukiclub.com.Timestamps00:00 - Stewart welcomes Mauro Schilman, CTO and Co-founder of Tuki Travel, who shares how traveling since age 15 through high school exchanges opened his mind to cultural similarities and differences.05:00 - Mauro explains Math Olympiad coaching culture and mentorship, noting LLMs now solve competition-level problems while Terence Tao explores AI assisting frontier unsolved mathematics.10:00 - Discussion turns to ChatGPT revealing Mauro's birthdate unprompted, exposing opaque application layers, preference tuning, and system prompts hidden within closed models.15:00 - Mauro argues true open source AI requires full training data, annotation protocols, and alignment processes, not just model weights, while scaling laws appear to be slowing.20:00 - Hierarchical memory models replace flat vector databases, using three-level retrieval systems improving context accuracy as knowledge management becomes AI's core challenge.25:00 - Mauro describes travel's fragmented infrastructure of aggregators, bed banks, and intermediaries, explaining Tuki builds agent-ready unification protocols for AI commerce.30:00 - MCP versus API debate clarifies natural language capability descriptions help agents consume services, while agent-to-agent communication embeds negotiating agents inside supplier systems.35:00 - Hallucinations and consumer trust block agentic payments, industries must build mistake-resilience into bookings before autonomous agent transactions become viable.40:00 - Mauro reveals red-green test-driven development methodology where agents write failing tests first then implementations, creating Oracle verification loops dramatically improving code quality.45:00 - Blockchain's potential for transparent distributed AI training discussed, distinguishing democratization from decentralization while stable coins and regulatory momentum build toward agentic commerce infrastructure.Key Insights1. Travel broadens perspective by revealing both universal human similarities and deep cultural differences. Mauro Schilman began traveling at fifteen through math olympiad competitions and found that people across the world share fundamental traits while also being shaped in profoundly different ways by their cultures. This tension between sameness and difference is what makes travel meaningful.2. Mathematics transitions from structured problem-solving in olympiads to genuine uncertainty in graduate school and research. Olympiad problems are carefully designed with elegant solutions meant to encourage creative thinking, but once a mathematician enters academia, the answers are unknown and the work becomes navigating that uncertainty.3. AI is now assisting mathematicians at the frontier, not just solving olympiad-level problems. Terence Tao, one of the greatest living mathematicians, has written publicly about how AI tools can help tackle unsolved problems, though the role of AI remains assistive rather than independent at the research level.4. Large language models are not truly transparent even when described as open source. Releasing model weights alone does not reveal the training data, annotation protocols, alignment tuning, or system prompts that shape model behavior. Real openness would require access to the entire pipeline.5. Memory and retrieval remain core unsolved challenges in AI systems. Researchers are moving from flat vector database approaches toward hierarchical memory structures with roughly three layers, which improves retrieval accuracy and reduces how much context gets consumed with each search.6. The travel industry is structurally unprepared for AI agents. A hidden web of bed banks, aggregators, and aggregators of aggregators sits between hotels and consumers, each taking a fee. Tuki Travel is building infrastructure to unify this distribution layer and make it consumable by AI agents through protocols like MCP and emerging agent-to-agent communication standards.7. Test-driven development using a red-green approach significantly improves AI-generated code quality. By asking the model to write failing tests before writing any implementation, developers create a verification oracle that guides the model toward correct solutions and avoids the bias of writing tests that simply confirm existing flawed code.