Technical Problems in Long-Term AI SafetyAndrew Critch Technical (and Non-Technical) Problems in Long-Term AI Safety Andrew Critch Machine Intelligence Research Institute

Technical Problems in Long-Term AI SafetyAndrew Critch Motivation, Part 1: Is human-level AI plausible? There are powerful short-term economic incentives to create human-level AI if possible. Natural selection was able to produce human- level intelligence. Thus, HLAI seems plausible in the long-term. Recent surveys of experts give arrival medians between 2040 and 2050.

Technical Problems in Long-Term AI SafetyAndrew Critch From :

Technical Problems in Long-Term AI SafetyAndrew Critch Cumulative probability of AI being predicted over time, by group From :

Technical Problems in Long-Term AI SafetyAndrew Critch Motivation, Part 2: Is superintelligence plausible? In many domains, once computers have matched human performance, soon afterward they far surpassed it. Thus, not long after HLAI, it’s not implausible that AI will far exceed human performance in most domains, resulting in what Bostrom calls “superintelligence”.

Technical Problems in Long-Term AI SafetyAndrew Critch (optional pause for discussion / comparisons)

Technical Problems in Long-Term AI SafetyAndrew Critch Thought experiment: Imagine it’s 2060, and the leading tech giant announces it will roll out the world’s first superintelligent AI sometime in the next year. Is there anything you’re worried about? Are there any questions you wish there had been decades of research on dating back to 2015? Motivation, Part 3: Is superintelligence safe?

Technical Problems in Long-Term AI SafetyAndrew Critch Some Big questions Is it feasible to build a useful superintelligence that, e.g., Shares our values, and will not take them to extremes? Will not compete with us for resources? Will not resist us modifying its goals or shutting it down? Can understand itself without deriving contradictions as in Gödel’s Theorems?

Technical Problems in Long-Term AI SafetyAndrew Critch Goal: Develop these big questions past the stages of philosophical conversation and into the domain of mathematics and computer science PhilosophyMathematics/CS Big Questions Technical Understanding

Technical Problems in Long-Term AI SafetyAndrew Critch Motivation, Part 4: Examples of technical understanding Vickrey second-price auctions (1961) : – Well-understood optimality results (truthful bidding is optimal) – Real-world applications, (network routing) – Decades of peer-review

Technical Problems in Long-Term AI SafetyAndrew Critch Nash equilibria (1951) :

Technical Problems in Long-Term AI SafetyAndrew Critch Classical Game Theory (1953) : An extensive form game.

Technical Problems in Long-Term AI SafetyAndrew Critch Key Problem: Counterfactuals for Self-Reflective Agents What does it mean for a program A to improve some feature of a larger program E in which A is running, and which A can understand? def Environment (): … def Agent(senseData) : def Utility(globalVariables) : … … do Agent(senseData1) … do Agent(senseData2) … end

Technical Problems in Long-Term AI SafetyAndrew Critch (optional pause for discussion of IndignationBot)

Technical Problems in Long-Term AI SafetyAndrew Critch Example: π maximizing What would happen if I changed the first digit of π to 9? This seems absurd because π is logically determined. However, the result of running a computer program (e.g. the evolution of the Schrodinger equation) is logically determined by its source code and inputs…

Technical Problems in Long-Term AI SafetyAndrew Critch … when an agent reasons to do X “because X is better than Y”, considering what would happen if it did Y instead means considering a mathematical impossibility. (If the agent has access to its own source code, it can derive a contradiction from the hypothesis “I do Y”, from which anything follows. This is clearly not how we want our AI to reason. How do we?

Technical Problems in Long-Term AI SafetyAndrew Critch Current formalisms are “Cartesian” in that they separate an agent’s source code and cognitive machinery form its environment. This is a type error, and in combination with other subtleties, it has some serious consequences.

Technical Problems in Long-Term AI SafetyAndrew Critch Examples (page 1) Robust Cooperation in the Prisoners’ Dilemma (LaVictoire et al, 2014) demonstrates non- classical cooperative behavior in agents with open source codes; Robust Cooperation in the Prisoners’ Dilemma Memory Issues of Intelligent Agents (Orseau and Ring, AGI 2012) notes that Cartesian agents are oblivious to damage to their cognitive machinery; Memory Issues of Intelligent Agents

Technical Problems in Long-Term AI SafetyAndrew Critch Examples (page 2) Space-Time Embedded Intelligence (Orseau and Ring, AGI 2012) provides a more naturalized framework for agents inside environments; Space-Time Embedded Intelligence Problems of self-reference in self-improving space-time embedded intelligence (Fallenstein and Soares, AGI 2014) identifies problems persisting in the Orseau-Ring framework, including procrastination and issues with self- trust arising from Löb’s theorem; Problems of self-reference in self-improving space-time embedded intelligence

Technical Problems in Long-Term AI SafetyAndrew Critch Examples (page 3) Vingean Reflection: Reliable Reasoning for Self-Improving Agents (Fallenstein and Soares, 2015) provides some approaches to resolving some of these issues; Vingean Reflection: Reliable Reasoning for Self-Improving Agents … lots more; see intelligence.org/research for additional reading.intelligence.org/research

Technical Problems in Long-Term AI SafetyAndrew Critch Summary There are serious problems with superintelligence that need formalizing in the way that fields like probability theory, statistics, and game theory have been formalized. Superintelligence poses a plausible existential risk to human civilization. Some of these problems can be explored now via examples in theoretical computer science and logic. So, let’s do it!

Technical Problems in Long-Term AI SafetyAndrew Critch Thanks! To…. Owen Cotton-Baratt for the invitation to speak. Patrick LaVictoire for reviewing my slides. Laurent Orseau, Mark Ring, Mihaly Barasz, Paul Christiano, Benja Fallenstein, Marcello Herreshoff, Patrick LaVictoire, and Eliezer Yudkowsky for doing all the research I cited