Page 209 of 506 ยท Back to latest writing
DSPy argues that language-model systems should be written as modular programs and optimized by compilers, not assembled from fragile hand-written prompt strings. The approach replaces manual prompt tweaking with explicit objectives, evaluation data, and systematic search.
May marked a shift from fascination with foundation models toward the systems engineering that makes AI useful. As models commoditize, advantage is moving to harnesses, memory, failure handling, safe execution, and redesigned operating workflows.
Tom Freston co-founded MTV and built it into a global cultural brand by backing offbeat creative talent over conventional corporate management. His later path offers lessons on leading creative teams, absorbing abrupt failure, and finding value in a nonlinear career.
AI agents become dependable when connected to real tools and data, constrained by permissions, and continuously evaluated, shifting automation from isolated generation toward trustworthy execution with less administrative friction.
This paper argues that generative AI may not flatten performance differences in knowledge work. It may widen them, especially when star employees have the judgment, autonomy, and reputation to turn AI into portable bargaining power.
This paper argues that AI compute is not a simple commodity input. Chip architecture, energy use, software, training workloads, inference workloads, and policy all change what one more unit of compute really means.
An enterprise AI cost paper proposes a dedicated operating model because generative AI spend is driven by usage, ownership, governance, and workflow design. It connects unit economics to accountability for where, why, and by whom inference is consumed.
A Strategy and Leadership paper argues that CEOs lead AI transformation through generalist, expert, or disruptive operating models, each with different risks. The framework helps leaders choose where expertise, authority, and organizational disruption should sit.