Research Workshops

This workshop brought together researchers with machine learning backgrounds to work on long-term AI safety problems that can be modeled in current machine learning systems and frameworks, for instance those described in “Concrete Problems in AI Safety” and “Alignment for Advanced Machine Learning Systems”.

This workshop was funded in part by a grant from the Artificial Intelligence Journal.

This two-day weekend workshop brought together researchers with interests in long-term theoretical AI safety research. The workshop covered the context and content of current AI safety research agendas and projects (with a focus on MIRI’s Agent Foundations technical agenda). It was geared for researchers who have technical backgrounds and who have not previously worked extensively with MIRI.

This small three-day workshop brought together researchers with machine learning backgrounds to work on long-term AI safety problems that can be modeled in current machine learning systems and frameworks, for instance those described in “Concrete Problems in AI Safety” and “Alignment for Advanced Machine Learning Systems”.

Topics included zero-shot learning using shared embeddings, the differences between quantilization and regularization, generative adversarial networks and Goodhart’s Law, and mathematical formalizations of conservative concept learning.

Participants at this three-day workshop — most of them veterans of past workshops — worked on a variety of problems related to MIRI’s Agent Foundations technical agenda.

Topics included safe exploration in rich domains, the difference between predicting a human and predicting HCH, and decision theories resulting from other decision theories self-modifying.

This small three-day workshop brought together researchers with machine learning backgrounds to work on long-term AI safety problems that can be modeled in current machine learning systems and frameworks, for instance those described in “Concrete Problems in AI Safety” and “Alignment for Advanced Machine Learning Systems”.

Topics included concept learning with different ontologies, problems for Task AGI, censored representations, and conservative concepts.

This three-day workshop brought together researchers with machine learning backgrounds to work on long-term AI safety problems that can be modeled in current machine learning systems and frameworks, for instance those described in “Concrete Problems in AI Safety” and “Alignment for Advanced Machine Learning Systems”.

Topics included learning human-interpretable and causal models of the environment; engineering cost functions based on impact measures to disincentivize side effects; designing robust metrics for the quality of a purported explanation of a plan; and developing a formal model of Goodhart’s Law which yields mild optimization.

Participants at this workshop — all of them veterans of past workshops — worked on a variety of problems related to MIRI’s Agent Foundations technical agenda, with a focus on decision theory and the formal construction of logical counterfactuals.

The Colloquium Series on Robust and Beneficial AI included a series of workshops to facilitate conversations and collaborations between people interested in a number of different approaches to the technical challenges associated with AI robustness and reliability.

The fourth workshop of CSRBAI focused on the topics of designing agents that behave well in their environments, without ignoring the effects of the agent’s own actions on the environment or on other agents within the environment.

The Colloquium Series on Robust and Beneficial AI included a series of workshops to facilitate conversations and collaborations between people interested in a number of different approaches to the technical challenges associated with AI robustness and reliability.

The third workshop of CSRBAI focused on the topic of preference specification for highly capable AI systems, in which the perennial problem of wanting code to “do what I mean, not what I said” becomes increasingly challenging.

The Colloquium Series on Robust and Beneficial AI included a series of workshops to facilitate conversations and collaborations between people interested in a number of different approaches to the technical challenges associated with AI robustness and reliability.

The second workshop of CSRBAI focused on the topic of robustness and error-tolerance in AI systems, and how to ensure that when AI system fail, they fail gracefully and detectably.

The Colloquium Series on Robust and Beneficial AI included a series of workshops to facilitate conversations and collaborations between people interested in a number of different approaches to the technical challenges associated with AI robustness and reliability.

The first workshop of CSRBAI focused on the topic of transparency in AI systems, and how we can increase transparency while maintaining capabilities.

Participants worked on questions of self-reference in type theory and automated theorem provers, with the goal of studying systems that model themselves.

This was the sixth in a series of introductory workshops, where MIRI brought together researchers with different backgrounds, discussed open problems in one of the technical agenda topics, and began projects and collaborations in that area.

The topic of this workshop was decision theory, and projects begun at the workshop are discussed in the following post: Proof Length and Logical Counterfactuals Revisited

This was the fifth in a series of introductory workshops, where MIRI brought together researchers with different backgrounds, discussed open problems in one of the technical agenda topics, and began projects and collaborations in that area.

The topic of this workshop was logical uncertainty, and projects begun at the workshop are discussed in the following post: What’s logical coherence for anyway?

This was the fourth in a series of introductory workshops, where MIRI brought together researchers with different backgrounds, discussed open problems in one of the technical agenda topics, and began projects and collaborations in that area.

The topic of this workshop was Vingean reflection, and projects begun at the workshop are discussed in the following posts:

• Vingean Reflection: Open Problems
• Relating Modal Polymorphism to PA with soundness
• Waterfall Truth Predicates

This was the third in a series of introductory workshops, where MIRI brought together researchers with different backgrounds, discussed open problems in one of the technical agenda topics, and began projects and collaborations in that area.

The topic of this workshop was decision theory, and projects begun at the workshop are discussed in the following post: Fixed point theorem in the finite and infinite case

This was the second in a series of introductory workshops, where MIRI brought together researchers with different backgrounds, discussed open problems in one of the technical agenda topics, and began projects and collaborations in that area.

The topic of this workshop was logical uncertainty, and projects begun at the workshop are discussed in the following posts:

• Stationary algorithmic probability
• A tractable, interpretable formulation of approximate conditioning for pairwise-specified probability distributions over truth values

This was the first in a series of introductory workshops, where MIRI brought together researchers with different backgrounds, discussed open problems in one of the technical agenda topics, and began projects and collaborations in that area.

The topic of this workshop was decision theory, and projects begun at the workshop are discussed in the following posts:

• Optimal and Causal Counterfactual Worlds
• An Informal Conjecture on Proof Length and Logical Counterfactuals
• Provability Counterfactuals vs Three Axioms of Galles and Pearl
• A Counterexample to an Informal Conjecture on Proof Length and Logical Counterfactuals

Participants at this workshop — all of them veterans of past workshops — worked on a variety of problems related to Friendly AI. The first tech report from this workshop is available here.

Participants at this workshop focused on the Löbian obstacle, probabilistic logic, and the intersection of logic and probability more generally. The results of this workshop are described here. See photos from the workshop here.

Participants at this workshop investigated problems related to reflective agents, probabilistic logic, and priors over logical statements / the logical omniscience problem. Some results from this workshop were developed further at the December 2013 workshop and described here.

This workshop focused on a variety of open problems related to normative decision theory. Participants brainstormed “well-posed problems” in the area, built on LaVictoire et al.’s Löbian cooperation work, made some progress on formalizing updateless decision theory, and formulated additional toy problems such as the Ultimate Newcomb’s Problem.

These results are still being written up in various forms.

This workshop focused on a variety of issues related to the Löbian obstacle for self-modifying systems, and to Demski’s earlier work on logical prior probability. The primary result was a proof that attempting to create a probability distribution which performs scientific induction on Π₁ statements, converging to probability 1 for the true versions of such statements, can create zero limiting probabilities assigned to true Π₂ statements. This result is still being written up, but it has been discussed briefly in a blog post by Demski. Other bits of progress were developed at further workshops and described here.

This three-week workshop addressed multiple open research problems simultaneously. First, participants found an improved version of the reflection principle discovered in the previous workshop, though this progress is still being written up. Second, participants improved upon earlier work by LaVictoire, resulting in the paper “Robust Cooperation in the Prisoner’s Dilemma: Program Equilibrium via Provability Logic.” Third, participants improved upon Benja Fallenstein’s parametric polymorphism approach to tackling the Löbian obstacle for self-modifying systems.

This workshop pursued one line of attack on the Löbian obstacle for self-modifying systems. The primary result of this workshop was a non-constructive “loophole” in Tarski’s undefinability of truth (via a fixed point theorem), which was later written up in draft form as “Definability of Truth in Probabilistic Logic” (see discussions here, here, and here).