Machine Intelligence Research Institute

History

In 2000, Eliezer Yudkowsky , Brian and Sabine Atkins founded the Singularity Institute for Artificial Intelligence to “help humanity prepare for the moment when machine intelligence exceeded human intelligence”. ^{[9] [10] [11]} In early 2005, SIAI relocated from Atlanta, Georgia to Silicon Valley. From 2006 to 2012, the Institute collaborated with Singularity University to produce the Singularity Summit , a science and technology conference. Speakers included Steven Pinker , Peter Norvig , Stephen Wolfram , John Tooby , James Randi , and Douglas Hofstadter .^{[12] [13] [14]}

In mid-2012, the Institute spun off a new organization called the Center for Applied Rationality , which focuses on cognitive science to improve people’s effectiveness in their daily lives. ^{[15] [16] [17]} Having previously shortened its name to “Singularity Institute”, in January 2013 SIAI changed its name to “Machine Intelligence Research Institute” in order to avoid confusion with Singularity University. MIRI gave control of the Singularity Summit to Singularity University and shifted its focus towards research in mathematics and theoretical computer science. ^[18]

In mid-2014, Nick Bostrom’s book Superintelligence: Paths, Dangers, Strategies helped spark public discussion about AI’s long-run social impact, and endorsements of Bill Gates and Elon Musk . ^{[19] [20] [21] [22]} Stephen Hawking and AI pioneer Stuart Russell co-authored at Huffington Post article citing the work of MIRI and other organizations in the area:

The short-term impact of AI depends on who controls it, the long-term impact depends on whether it can be controlled at all. […] Although we are facing the prospect of a greater or lesser incidence of this problem, we believe that it is possible that the Cambridge Center for Existential Risk , the Future of Humanity Institute , the Machine Intelligence Research Institute, and the Future of Life Institute . ^[7]

In early 2015, MIRI’s research is cited in a document that is more important than that of AI that is robust and beneficial. ^[23] Bostrom, Russell, Bart Selman , Francesca Rossi , Thomas Dietterich , Manuela M. Veloso , and researchers at MIRI. ^[8] [24]

In 2016, MIRI published a paper, Logical induction . A positive, unpublished review of this paper led the Open Philanthropy Project to award MIRI $ 3.75 million. ^[25] [26]

Research

Forecasting

MIRI studies strategic issues related to AI, such as: What can we predict about future AI technology? How can we improve our forecasting ability? Which interventions available today appear to be the most beneficial, given what we know? ^[27]

Beginning in 2014, MIRI has been working with the AI ​​Impacts project. Impacts studies historical instances of discontinuous technological change, and has developed new measures of the relative computational power of humans and computer hardware. ^[28] [29]

MIRI researchers ‘interest in discontinuous AI progress stems from I. J. Good ‘ s argument that will be able to finally advanced systems in software engineering, leading to a feedback loops of capable AI systems: ^{[4] [30] [23] [ 22]}

Let an ultraintelligent machine be defined as a machine that can surpass all the intellectual activities of any man but clever. Since the design of machines is one of those intellectual activities, an ultraintelligent machine could design even better machines; There would be an explosion of intelligence, and the intelligence of man would be far behind. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control. ^[31]

Writers like Bostrom uses the term superintelligence in place of Good’s ultraintelligence . ^[19] Following Vernor Vinge , Good’s idea of intelligence has come to be associated with the idea of ​​a ” singularity “. ^{[32] [33] [34]} Bostrom and researchers at MIRI have expressed skepticism about the views of singularity advocates like Ray Kurzweil’s superintelligence is “just around the corner”. MIRI researchers advocated early safety work as a precautionary measure, while arguing that predictions of AI progress have not proven reliable. ^{[35] [22] [19]}

Eliezer Yudkowsky, MIRI’s co-founder and senior researcher, is frequently cited for his writing on the long-term social impact of progress in AI. Russell and Norvig’s Artificial Intelligence: A Modern Approach , the standard textbook in the field of AI, summarizes Yudkowsky’s thesis:

If ultraintelligent machines are a possibility, we would make sure that we design their predecessors in such a way that they design themselves to treat us well. […] Yudkowsky (2008) ^[30] goes into more detail about how to design a Friendly AI . He asserts that friendliness should be designed, but that the designers should recognize that their own designs may be flawed, and that the robot will learn and evolve over time. Thus the problem is one of mechanism design-to define a mechanism for evolving the system of checks and balances, and to give the systems utility functions that will remain friendly in the face of such changes. ^[33]

Yudkowsky writes on the importance of friendly artificial intelligence in smarter-than-human systems. ^[36] This informal goal is reflected in MIRI’s recent publications as the requirement that Al systems be “aligned with human interests”. ^[4] Following Bostrom and Steve Omohundro , MIRI researchers believe that the intelligent world is not a problem, but it is a question of obstacles, or threats if they are not specifically designed to promote their operators’ goals. ^{[37] [38] [19] [8]}

High reliability and error tolerance in AI

The Future of Life Institute (FLI)

Mathematical tools such as formal logic, probability, and decision theory have made significant insight into the foundations of reasoning and decision-making. However, there are still many open problems in the foundations of reasoning and decision. Solutions to these problems can make the behavior of the system much more reliable and predictable. Example research topics in this area include reasoning and decision-making under computational resources at Horvitz and Russell, how to take into account correlations between systems and behaviors and their environment? reason, and how to reason about uncertainty over logical consequences of beliefs or other deterministic computations. These topics can benefit from being considered together,^[23]

The priorities document quotes MIRI publications in the relevant areas: formalizing cooperation in the prisoner’s dilemma between ” superrational ” software agents; ^[39] defining alternatives to causal decision theory and evidential decision theory in Newcomb’s problem ; ^[40] and developing alternatives to Solomonoff’s theory of inductive inference for agents embedded in physical environments ^[41] and agents reasoning without logical omniscience. ^[42] [22]

Standard decision procedures are not well-specified enough (eg, with regard to counterfactuals) to be instantiated as algorithms. MIRI researcher Benja Fallenstein and then-researcher Nate Soares that decision theory is “unstable under reflection” in the sense that a rational agent following causal decision theory “correctly identifies that the agent should modify itself to stop using CDT [causal decision theory] to make decisions “. MIRI researchers identify “logical decision theories” as alternatives that perform better in general decision-making tasks. ^[40]

MIRI also studies self-monitoring and self-verifying software. The FLI research priorities document notes that “a formal system that is rather powerful in the sense of gaining assurance about the accuracy of functionally similar formal systems, on the breadth of inconsistency via Gödel’s incompleteness theorems “. ^[23] MIRI’s publications on Vingean reflection attempts to model the Gödelian limits on self-referential reasoning and identify practically useful exceptions. ^[43]

Soares and Fallenstein classifies the above research programs as a high level of transparency and transparency in agent behavior. They also recommend research into “error-tolerant” software systems, citing human error and default incentives as sources of serious risk. ^[37] [8] The FLI research priorities document adds:

If a system is selecting the actions that are best suited to a given task, then avoiding conditions is a natural subgoal (and conversely, seeking unconstrained situations is sometimes a useful heuristic). This could become problematic, however, if we wish to repackage the system, to deactivate it, or to significantly alter its decision-making process; such a system would rationally avoid these changes. Systems That do not exhibit thesis Behaviors-have-been termed correctable systems, and Both theoretical and practical work in this area Appears towable and Useful.

MIRI’s priorities in these areas are summarized in their 2015 technical agenda. ^[4]

Value specification

Soares and Fallenstein write, “the ‘intentions’ of the operators are a complex, vague, fuzzy, context-dependent notion (Yudkowsky 2011). ^[44] Concretely writing out the full intentions of the operators in a machine-readable format is implausible if not impossible, even for simple tasks. ” Soares and Fallenstein proposes that autonomous AI systems instead be designed to inductively learn the values ​​of humans from observational data. ^[4]

Soares discusses several technical obstacles to value learning in AI: changes in the agent’s beliefs may result in a mismatch between the agent’s values ​​and its ontology; agents who are well-behaved in training and human operators may make it difficult to identify or anticipate incorrect inductions. ^[22] [45] Bostrom’s Superintelligence discusses the philosophical problems raised by value learning at greater length. ^[19]

See also

References