As the current Dovetail research fellowship comes to a close, the fellows are giving talks on their projects. All are welcome to join! Unlike the previous cohort talks, these talks will be scheduled one at a time. This is partly because there are too many to do all in one day, and partly because the ending dates for several of the fellows are spread out over time.
The easiest way to keep track of the schedule is to subscribe to the public Dovetail google calendar. I'll also list them here in this post, which I'll update as more talks get scheduled.
Santiago Cifuentes - General Agents Contain World Models, even if they are non-deterministic and the world is partially observable. (Arxiv paper)
In this talk we will present some concrete extensions of the results from General Agents Contain World Models. More precisely, we will extend their result 1 for non-deterministic agents and partially observable environments.
January 31 (Saturday) 2000 GMT/1200 PT
Léo Cymbalista - An introduction to Computational Mechanics
In this talk, we present a result related to the agent structure problem, which posits that a system which exhibits agent-like behavior must also contain agent-like structure. We model policies with deterministic finite automata and show that the DFAs of policies that meet certain robustness criteria must share a similar feature.
We model agents as string-to-string behaviors and introduce Finite-State Reactive Agents (FSRAs) - deterministic, length- and prefix-preserving transductions with finite memory. I’ll show the exact characterization of behaviors implementable by finite state, explain enforceability vs. realizability of specifications, and give examples that separate finite-state from richer models.
"all of what we mean by goals and purposes can be well thought of as maximization of the expected value of the cumulative sum of a received scalar signal (reward)"
But what are goals? What are purposes? And a reward function? And is the reward hypothesis true? In this talk I'll go through the some of the main results of decision theory and reinforcement learning theory that determine when such objective functions exist: Debreu, von Neumann–Morgenstern, and Markovian reward axioms.
In this talk, Margot will discuss her recent posts on "An Ontology of Representations: Limits of Universality" and "(Re-)Discovering Natural Laws". In an Ontology of representations, she challenges the Platonic Representation Hypothesis, Natural Latents, and Universality Hypothesis which suggest AI systems will converge on a shared, objective model of reality. Drawing on philosophy of science and physics, she argues that reality itself lacks a single unified ontology: physics is a patchwork of scale-dependent theories connected by singular, non-smooth limits, and even "paradigm" reductions like thermodynamics to statistical mechanics involve observer-relative choices. The observed convergence in AI is therefore better explained by shared human-scale training data and inductive biases than by discovery of objective structure. She believes that specialized networks, trained on curated scientific datasets, will transform scientific practice in the near term more profoundly than general-purpose AI systems. The next major challenge is figuring out how to do this reliably across different areas of science, understanding which architectures and datasets and training regimes unlock which domains, and how to verify that the representations learned correspond to genuine structure. In (Re-)Discovering Natural Laws, she describes these examples in more detail, and discuss some of the efforts made in automating the (re-)discovery of natural laws.
This talk will consist of two parts. First I will review the Internal Model Principle from the point of view of abstract dynamical systems, discussing its shortcomings as a tool for agency research. Second I will present a different approach to the IMP that emphasizes deducing information about system from the behaviour of controller.
As the current Dovetail research fellowship comes to a close, the fellows are giving talks on their projects. All are welcome to join! Unlike the previous cohort talks, these talks will be scheduled one at a time. This is partly because there are too many to do all in one day, and partly because the ending dates for several of the fellows are spread out over time.
The easiest way to keep track of the schedule is to subscribe to the public Dovetail google calendar. I'll also list them here in this post, which I'll update as more talks get scheduled.
All talks will be on Zoom at this link.
More to come!