Johns Hopkins Natural Philosophy Symposium 2025

Mitchell's talk discussed benchmarks for AI and methods to probe for "emergent world models." The central question: does a GPT create a world model or merely a bag of heuristics? Mitchell argues that standard NLP benchmarks are insufficient to answer this — they measure performance on tasks that can be solved by pattern matching without genuine understanding.

An Abstraction & Reasoning Corpus (ARC) Benchmark may be a better way to assess AI abilities on a number of basic spatial and semantic concepts — requiring genuine compositional reasoning rather than statistical pattern completion. This has direct implications for how we evaluate and trust AI systems in safety-critical roles.

Seth makes the case that consciousness depends on our nature as living organisms — a form of biological naturalism. The presentation argues for biological naturalism grounded in active inference, cybernetics, autopoiesis, and the free energy principle. On this account, current AI systems, however sophisticated, lack the biological and physiological grounding that underpins subjective experience.

Seth's approach is of particular interest to the Spatial Web Foundation development of an agent framework — providing a principled philosophical basis for distinguishing AI agents from conscious agents, and for calibrating governance and oversight accordingly. There is an open call for comments on Seth's draft BBS paper.

Core to Chalmers' talk was assessing the prospects for objective phenomenology — and the need to translate phenomenology into math to achieve it. Chalmers aimed to define a "Rosetta Stone" vocabulary for connecting Integrated Information Theory of consciousness to phenomenology.

His thesis: math may be central to an objective theory of subjective experience. This parallels the Spatial Web's use of mathematics (hyperspace, tensors, HSML) to formalize world models — and raises a deeper question about whether formal mathematical representations can capture the full structure of experience and knowledge.

Tomasello discussed main types of psychological agency and describes them in evolutionary order of emergence — from goal-directed agency in ancient vertebrates through intentional, rational, and finally socially normative agency unique to humans.

This evolutionary typology forms a cornerstone of GeoRoundtable's approach to agentic AI design and governance. Understanding where artificial agents sit on this continuum — and what capacities distinguish each level — is not optional background but essential design knowledge for building and governing AI systems. I look forward to more in reading Tomasello's "The Evolution of Agency: Behavioral Organization from Lizards to Humans."

Zollman's talk, and our subsequent side conversation, focused on Network Epistemology — "how does the structure of our social networks influence our ability to learn about the world?" Zollman's Independence Thesis — that the norms of social epistemology are independent of the norms of individual epistemology — is key to understanding the difference between intelligent agents and the knowledge created by an ecosystem of agents.

This is a key topic in the design of the Universal Domain Graph of the Spatial Web: how the structure of agent connections shapes the collective knowledge that emerges. Zollman made the point that epistemic governance is at the forefront of epistemic communities. Be sure to keep an eye out for Zollman's forthcoming book — "How Social Connections Shape Knowledge."

Levin's talk included discussion of collective intelligence as a public good. Levin pointed out the role of the commons and reference to Elinor Ostrom's work on polycentric governance — a framework for managing shared resources through distributed, overlapping governance structures rather than central control.

Levin's work on measures of collective intelligence in evolved and designed self-organizing ensembles is directly relevant to the design and implementation of Spatial Web collective intelligence: understanding how collective knowledge and coordinated action emerge from networks of agents, and how governance of those networks can be structured to produce public goods rather than tragedies of the commons.