The Future of AI Systems: EP99.04-PREVIEW

Chris discusses the competitive landscape of AI models, speculating on which company will lead by the end of 2025, particularly questioning whether Google will have the best model with its Gemini 2.5 Pro, which has shown improvements in coding capabilities. He expresses skepticism about other companies like XAI and Anthropic being able to release competitive models soon, while noting that OpenAI's chances of leading are surprisingly low at 3%.

The conversation highlights a shift among major AI providers towards tooling enhancements rather than groundbreaking new models, suggesting a plateau in model development. Chris mentions instability issues with Gemini 2.5 Pro, citing mixed user experiences and community feedback about performance regressions. Users are increasingly reliant on Gemini 2.5, expecting reliability despite its experimental label, leading to frustrations on platforms like X when issues arise.

The importance of context in AI is emphasized, particularly in handling large data sets and multiple tool calls. The discussion touches on Anthropic's advancements in context understanding and questions whether their models are fundamentally more intelligent than GPT models. Upcoming models are expected to improve tool calling and reasoning capabilities, with the ability to self-correct when encountering issues.

Observability in AI processes is discussed, with multi-step corrections allowing users to monitor and intervene in ongoing tasks, enhancing the reliability of AI outputs. The conversation underscores the significance of methodical thinking and effective tool use in advancing AI technology. Trust in AI systems is a critical issue, especially regarding their ability to perform tasks with significant real-world consequences. Concerns are raised about the risks of unmonitored AI decision-making in high-stakes scenarios, emphasizing the necessity of human oversight.

The evolution of AI protocols, particularly in agent thinking and tool calling, is reflected upon, with a sense of déjà vu regarding recycled AI ideas. The conversation acknowledges that while agentic applications have existed for some time, they are now being utilized more effectively, with companies like Anthropic exploring native tool calling integration. The potential of AI agents equipped with various tools is discussed, emphasizing the need for clarity in tool selection to avoid confusion when multiple tools serve similar functions.

The need for a structured methodology in agent operations is highlighted to ensure consistency and reproducibility in outcomes. Users expect reliable performance from AI, whether for complex tasks or simple comparisons. The conversation questions whether models can be trained to select the right tools effectively and speculates on the capabilities of future models like GPT-5, which may integrate tool-calling functionalities.

Concerns are raised about the inefficiency of agents expending excessive resources on simple tasks, underscoring the need for better planning. The speaker expresses frustration with the inconsistency of AI systems and emphasizes reproducibility as a key issue. The concept of a "skills button" is introduced, allowing users to specify their intentions, though its effectiveness is questioned. Clustering tools based on task requirements is suggested to enhance AI performance.

The conversation centers on the future of AI systems, highlighting the importance of trained skills over random Model Control Protocols (MCPs) for enhancing productivity. Anticipation builds around upcoming AI announcements and their practical applications, envisioning a future where users can command AI to perform tasks seamlessly. The discussion predicts a shift in user interfaces, with AI applications becoming the primary starting point for online interactions.

Critique is directed at OpenAI, suggesting that the company is reacting to the popularity of MCPs rather than leading in innovation. The speaker argues that OpenAI lacks groundbreaking developments and compares its current state to a time when it hinted at revolutionary technology. The conversation also touches on the evolution of AI models, suggesting that advancements may not be as exponential as previously believed.

The discussion highlights the importance of strong protocols and effective AI models for the widespread adoption of MCP servers. Accessing an MCP server provides a list of tools that must be managed carefully to avoid conflicts. The conversation speculates on future models and their potential for seamless tool integration while raising concerns about whether models should focus on tool integration or inference.

The challenges of transitioning between various AI platforms, particularly with systems like Gemini, are noted, along with the potential for Managed Cloud Platforms (MCPs) to host valuable private data. Concerns about a monopolistic marketplace for MCPs are raised, though there is optimism that ease of deployment may mitigate this risk. The discussion encourages companies with proprietary data to consider monetizing it through MCPs, as there is significant profit potential in providing effective tools.

Automation of repetitive tasks is identified as a valuable application of MCPs, with organizations encouraged to build agents that automate these functions. The potential for agent-to-agent protocols to evolve into a service model for SaaS applications is discussed, raising concerns about monopolization and control. The conversation also addresses the lack of corporate solutions for AI usage, emphasizing the need for organizations to control data entry and exit points to prevent leaks.

Critiques of specific AI products question their relevance and effectiveness, with skepticism about AI tools that offer minimal value. The ease of developing similar tools by competent developers in a short time is noted, leading to an overall skepticism about the future of AI in these applications. The conversation concludes with anticipation for upcoming announcements from major players like Google and OpenAI, humorously acknowledging the competitive nature of these events.