Broken Ladders & Curved Spaces: The Unseen Dynamics of AI

How AI Breaks the Career Ladder: Automation and Coordination Failure in Knowledge-Intensive Firms by Claudio Zucca

AI automation creates an intertemporal tradeoff for knowledge-intensive firms: cutting junior workers to exploit cheap automation boosts short-term profits but depletes the career pipeline needed to develop future senior talent. Individual f irms can resolve this tension through lateral hiring from competitors, but the industry as a whole cannot—lateral markets merely redistribute talent from a collectively shrinking pool. We develop a dynamic model with nested CES production calibrated to law firms, investment banking, and management consulting, treating the industry as a closed system where senior professionals must be developed internally through multi-year apprenticeships. When all f irms optimize automation individually, they collectively eliminate 92–95% of the junior talent pipeline. Dynamic simulations show firms maintaining career ladders sacrifice 16–32% of potential automation gains in the short run but ultimately achieve 60–112% higher output when partner scarcity becomes the binding constraint. Crossover periods occur 8–15 years after the automation shock (at absolute periods t=14 for investment banking, t=16 for law firms, and t=20 for management consulting). Although automation initially boosts output by 50%, the subsequent destruction of the career pipeline depletes the senior professionals who are production complements to automation, causing output to fall below the pre-shock baseline despite continued robot usage. This creates a coordination failure combining Becker’s training externality with common pool resource dynamics: individual firms can free-ride on competitors’ training investments through lateral hiring, but when all firms automate simultaneously, lateral markets cannot expand aggregate capacity. Solutions require industry wide coordination

Hidden Signals in Text Embeddings: Recovering Skill Hierarchies with Inductive Hyperbolic Representations by Thiago Brant

Many areas of learning have a hidden order: some skills and ideas need to come before others. In this work, we ask whether that order can be inferred from the subtle signals already embedded in text embeddings, even when the model is not explicitly told which concept should point to which. We train a self-supervised system that learns from text descriptions, places concepts in a curved space well suited for hierarchies, and uses only a very small hint to keep the overall direction consistent. When tested on entirely new skills, the method recovers prerequisite direction better than a comparable standard baseline. Control experiments also show that the gains come from meaningful semantic information in the text. Overall, the results suggest that text representations contain hidden clues about learning order, and that hyperbolic geometry helps bring that structure out.