A Reality Check on AI Diffusion

In this episode, I speak with Janusz Marecki, AI Partner at Ahrens, and the conversation cuts through the optimism that often surrounds AI. Janusz sets out a clear position. Current AI systems have fundamental shortcomings. They approximate answers rather than produce deterministic outcomes. Every token carries a probability of error, and those errors compound as outputs grow. This is not a marginal issue. It is embedded in how AI systems work. He also challenges a widely held assumption. AI systems are not continually learning from experience or new data. Once training stops, their knowledge is fixed. Attempts to update them through external data can introduce contradictions rather than resolve gaps. This creates the conditions for hallucination and, consequently, undermines trust. Enterprises are adapting through agentic systems. By breaking tasks into smaller components and orchestrating multiple models, organisations are improving performance. This comes at a cost because AI systems become more specialised and less general. The promise of broad applicability begins to narrow. Human oversight remains central. In mission critical environments, outputs need to be verified, either by people or by deterministic systems. Without this, reliability is insufficient. With it, the efficiency gains begin to erode. The implications extend to the workforce. Entry level programming roles are already being automated. Yet these roles are essential for developing future expertise. Removing them risks creating a structural gap in capability over time. Janusz draws a sharp distinction. There is real value in today’s AI applications. That is not in question. But expectations around artificial general intelligence are misplaced. Achieving AGI will require fundamentally new advances, not incremental improvements to existing models. This is a conversation grounded in reality rather than projection. Chapters 00:00 The Shortcomings of Current AI Systems 02:32 Reality Check on AI Adoption 05:57 Understanding Errors in AI Outputs 11:02 The Role of Agentic AI Systems 14:49 Humans in the Loop: Necessity or Burden? 17:49 The Future of Programming Jobs 19:14 AI Bubble vs. AGI Bubble 21:18 Defining Artificial General Intelligence 22:42 The Role of an AI Partner 25:12 Investments in Innovative AI Startups