Many critiques of AI remain rooted in the pre-agentic era. It is past time to update your priors!
Lanier writes that “attempting to open the black box by making a system spit out otherwise unnecessary items like scripts, sketches, or intentions will involve building another black box to interpret the first—an infinite regress.” He suggests a system should “trace the most unique and influential contributors when a big model provides a valuable output.” But this critique is rooted in the pre-agentic era (the New Yorker piece that forms the basis of this discussion is from 2023). Agentic LLM systems leverage so many contributions, often from many different LLMs, generative models, and organizing systems, that a list of provenances — covering the generated final output as well as the interstitial outputs that were relevant but discarded, or used only temporarily as scaffolding — would result in a compendium of references that is uselessly long. This is one of the biggest problems with so many recent critiques of AI systems: they are far less relevant since the agentic revolution that happened at the end of 2025. While chatbots are still important for many retail consumers, they are elemental compared to the complex agentic systems that are replacing work tasks. The value of an LLM-based system now has less to do with a legible textual output, and more with the organization of computation, planning, and action that contribute to the completion of a set of tasks. Perhaps you can still reveal human contributions in these complex systems, but it would be difficult to represent without its own chain of infinite interpretative regress.
