Our GPT-5 first impressions (one thing you missed)

This episode reveals that while GPT-5 offers only incremental gains, the real story is the architectural shift toward multi-agent systems, forcing an industry-wide pivot from model performance to a competitive battle over the agentic application layer.

First Impressions of GPT-5: Practical Use Cases and Incremental Gains

• The discussion begins with a real-world test of GPT-5’s capabilities, demonstrating its immediate value for complex, non-technical tasks. One of the speakers recounts using the model to navigate Portugal's notoriously difficult financial services website to file invoice receipts, a task his vacationing accountant usually handles.
• The model successfully provided a clear summary and a precise, step-by-step guide for the convoluted process.
• This highlights the model's strength in information synthesis and presentation, making complex bureaucratic procedures accessible to a layperson.
• The speaker notes, "I find docs so difficult to troll, but you ask GPT like questions and just the way in which the information is presented is so user friendly."
• Strategic Implication: While not a leap toward AGI, this level of practical utility in navigating real-world systems demonstrates a tangible increase in value, suggesting a market for AI assistants that can manage complex administrative and logistical tasks.

The Architectural Shift: From Mixture of Experts to Mixture of Models

• Richard provides a technical breakdown of GPT-5's underlying architecture, framing it as a significant evolution from its predecessor, GPT-4. He explains that the new model moves beyond a Mixture of Experts (MoE)—an architecture where different specialized sub-networks within a single model handle parts of a query—to a Mixture of Models (MoM).
• In this new MoM paradigm, a "router" sits in front of multiple, entirely separate models. When a query is received, the router intelligently selects and directs it to the most appropriate and efficient underlying model.
• This design is more cost-effective, as simple queries can be routed to smaller, cheaper models instead of unnecessarily engaging the most powerful one.
• Richard analyzes this shift as a foundational step toward more complex agentic systems: "Now we've moved from mixture of experts to like mixture of models... this is like a clear step towards a more kind of like multi-agent system I would say."
• Actionable Insight for Researchers: This architectural evolution from monolithic models to a routed system of specialized models is a critical trend. Researchers should investigate how this modular approach can be leveraged for decentralized agent networks and more efficient on-chain AI computation.

The Commoditization of Foundation Models and the Pivot to Agents

• The conversation pivots to the broader market implications of these incremental improvements. Richard expresses a growing indifference to new model releases, arguing that the performance gap between frontier models is shrinking, leading to their commoditization.
• With diminishing returns on model performance, the speakers agree that major AI labs like OpenAI must find new ways to compete and monetize.
• The clear strategic direction is moving up the stack from providing foundational models to building applications, specifically AI agents that perform tasks for users.
• Strategic Consideration for Investors: As foundation model providers encroach on the application layer, standalone agent-based startups face immense pressure. Investors should critically evaluate whether these startups have a defensible moat, such as proprietary data, unique distribution channels, or a specialized focus that larger players are unlikely to replicate.

Evaluating Early Agent Capabilities: A Reality Check

• Despite the industry's pivot toward agents, the current capabilities of these systems are critically examined and found wanting. One speaker shares a frustrating experience with OpenAI's "Agent Mode."
• He assigned it a relatively simple task: to analyze a list of speakers from an AI conference and identify those based in Europe.
• After more than 24 hours, the agent had still not completed the task, highlighting significant issues with speed and reliability.
• Actionable Insight: The hype around autonomous agents currently outpaces their real-world performance. Investors and researchers should maintain a healthy skepticism and focus on solutions that address the current limitations of speed, cost, and reliability, as these are the primary bottlenecks to widespread adoption.

The Platform Wars: Data Integration and Competitive Moats

• The discussion explores the strategic importance of data access, focusing on GPT-5's announced integrations with Google products like Gmail and Calendar. This raises questions about the long-term viability of building on a competitor's platform.
• The speakers debate whether Google will eventually block OpenAI's access to its ecosystem to favor its own model, Gemini, creating significant "platform risk."
• This potential conflict mirrors the historical platform wars, where access to user data and key software integrations becomes a primary competitive weapon.
• Strategic Consideration: The risk of de-platforming is a major threat for any AI application built on proprietary ecosystems. This underscores the strategic value of decentralized data and compute protocols in the Crypto AI space, as they offer a potential hedge against centralized control and censorship.

From General Hosting to "MCP-as-a-Service": A Startup's Pivot

• One of the speakers details his startup's recent pivot, offering a powerful lesson for builders in the Crypto AI space. After meeting with the Vercel CTO, his team shifted its strategy around MCP (Multi-agent Communication Protocol), a standard for enabling different AI agents to communicate.
• Their initial product, a general-purpose MCP hosting service, failed to gain traction because potential users didn't know how to derive value from the raw infrastructure.
• The startup pivoted to "MCP-as-a-Service," offering a specific, productized solution designed to improve the onboarding experience for developer-focused companies. This targeted approach solves a clear business need.
• Actionable Insight for Builders: This case study demonstrates that building general-purpose infrastructure is insufficient for finding product-market fit. Success in the current market requires creating opinionated, productized solutions that solve a specific, high-value problem for a well-defined customer.

The Importance of a Rigorous Experimental Framework

• Drawing from his own experiences, Richard reflects on the decision to pause his "$1 OATH" project. He attributes the pause to a lack of strategic clarity and a proper methodology for testing their assumptions.
• He emphasizes the mistake of launching without a clear hypothesis. "We didn't even know what we were trying to test," he admits, which made it impossible to iterate effectively based on user feedback.
• His team is now implementing a more rigorous experimental framework inspired by "The Lean Startup" to validate or invalidate hypotheses before committing significant resources.
• Strategic Takeaway: For early-stage Crypto AI ventures, technical innovation must be paired with a disciplined, hypothesis-driven approach to product development. Defining and testing assumptions is more critical than simply building and launching a product.

Conclusion

The AI race is shifting from foundational model supremacy to a battle for agentic applications and user integration. Investors and researchers must now focus on which companies can solve specific problems and secure data access, as this will define the next wave of value creation in the AI ecosystem.