But the difference between infinity and any finite value is infinity . Intelligence itself, or a substantial subset if it, is easy, given infinite resources, as AIXI shows. But that’s been of no use in developing real world AI: tractable approximations to AIXI aren’t powerful enough to be dangerous.
It would be embarrassing to MIRI if someone cobbled together AI smart enough to be dangerous, and came to the worlds experts on AI safety for some safety features, only to be told “sorry guys, we haven’t got anything that’s compatible with your system, because it’s finite”.
A Monte-Carlo approximation of AIXI can play Pac-Man and other simple games
(Veness et al. 2011), but some experts think AIXI approximation isn’t a fruitful path
toward human-level AI. Even if that’s true, AIXI is the first model of cross-domain
intelligent behavior to be so completely and formally specified that we can use it to
make formal arguments about the properties which would obtain in certain classes of
hypothetical agents if we could build them today. Moreover, the formality of AIXI-like
agents allows researchers to uncover potential safety problems with AI agents of increasingly
general capability—problems which could be addressed by additional research, as
happened in the field of computer security after Lampson’s article on the confinement
problem.
AIXI-like agents model a critical property of future AI systems: that they will need
to explore and learn models of the world. This distinguishes AIXI-like agents from
current systems that use predefined world models, or learn parameters of predefined
world models. Existing verification techniques for autonomous agents (Fisher, Dennis,
and Webster 2013) apply only to particular systems, and to avoiding unwanted optima
in specific utility functions. In contrast, the problems described below apply to broad
classes of agents, such as those that seek to maximize rewards from the environment.
For example, in 2011 Mark Ring and Laurent Orseau analyzed some classes of AIXIlike
agents to show that several kinds of advanced agents will maximize their rewards
by taking direct control of their input stimuli (Ring and Orseau 2011). To understand
what this means, recall the experiments of the 1950s in which rats could push a lever
to activate a wire connected to the reward circuitry in their brains. The rats pressed the
lever again and again, even to the exclusion of eating. Once the rats were given direct
control of the input stimuli to their reward circuitry, they stopped bothering with more
indirect ways of stimulating their reward circuitry, such as eating. Some humans also
engage in this kind of “wireheading” behavior when they discover that they can directly
modify the input stimuli to their brain’s reward circuitry by consuming addictive narcotics.
What Ring and Orseau showed was that some classes of artificial agents will
wirehead—that is, they will behave like drug addicts.
Fortunately, there may be some ways to avoid the problem. In their 2011 paper, Ring
and Orseau showed that some types of agents will resist wireheading. And in 2012,
Bill Hibbard (2012) showed that the wireheading problem can also be avoided if three
conditions are met: (1) the agent has some foreknowledge of a stochastic environment,
(2) the agent uses a utility function instead of a reward function, and (3) we define
the agent’s utility function in terms of its internal mental model of the environment.
Hibbard’s solution was inspired by thinking about how humans solve the wireheading
problem: we can stimulate the reward circuitry in our brains with drugs, yet most of us
avoid this temptation because our models of the world tell us that drug addiction will
change our motives in ways that are bad according to our current preferences.
Relatedly, Daniel Dewey (2011) showed that in general, AIXI-like agents will locate
and modify the parts of their environment that generate their rewards. For example,
an agent dependent on rewards from human users will seek to replace those humans
with a mechanism that gives rewards more reliably. As a potential solution to this problem,
Dewey proposed a new class of agents called value learners, which can be designed
to learn and satisfy any initially unknown preferences, so long as the agent’s designers
provide it with an idea of what constitutes evidence about those preferences.
Practical AI systems are embedded in physical environments, and some experimental
systems employ their environments for storing information. Now AIXI-inspired work
is creating theoretical models for dissolving the agent-environment boundary used as
a simplifying assumption in reinforcement learning and other models, including the
original AIXI formulation (Orseau and Ring 2012b). When agents’ computations must
be performed by pieces of the environment, they may be spied on or hacked by other,
competing agents. One consequence shown in another paper by Orseau and Ring is
that, if the environment can modify the agent’s memory, then in some situations even
the simplest stochastic agent can outperform the most intelligent possible deterministic
agent (Orseau and Ring 2012a).
I feel as though you’re engaging in pedantry for pedantry’s sake. The point is that if we can’t even solve the simplified version of the problem, there’s no way we’re going to solve the hard version—effectively, it’s saying that you have to crawl before you can walk. Your response was to point out that walking is more useful than crawling, which is really orthogonal to the problem here—the problem being, of course, the fact that we haven’t even learned to crawl yet. AIXI and Bayes are useful in that solving AGI problems in the context provided can act as a “stepping stone” to larger and bigger problems. What are you suggesting as an alternative? That MIRI tackle the bigger problems immediately? That’s not going to work.
You are still assuming that infinite systems count as simple versions of real world finite systems, but that is the assumption I am challenging: our best real world AIs aren’t cut down AIXI systems, they are something different entirely, so there is no linear progression from crawling to walking in your terms,
You are still assuming that infinite systems count as simple versions of real world finite systems
That’s not just an assumption; that’s the null hypothesis, the default position. Sure, you can challenge it if you want, but if you do, you’re going to have to provide some evidence why you think there’s going to be a qualitative difference. And even if there is some such difference, it’s still unlikely that we’re going to get literally zero insights about the problem from studying AIXI. That’s an extremely strong absolute claim, and absolute claims are almost always false. Ultimately, if you’re going to criticize MIRI’s approach, you need to provide some sort of plausible alternative, and right now, unfortunately, it doesn’t seem like there are any. As far as I can tell, AIXI is the best way to bet.
That’s not just an assumption; that’s the null hypothesis, the default position. Sure, you can challenge it if you want, but if you do, you’re going to have to provide some evidence why you think there’s going to be a qualitative difference.
I have already pointed out that the best AI systems currently existing are not cut down infinite systems.
And even if there is some such difference, it’s still unlikely that we’re going to get literally zero insights about the problem from studying AIXI. That’s an extremely strong absolute claim, and absolute claims are almost always false.
Something doesn’t have to be completely worthless to be sub optimal.
But the difference between infinity and any finite value is infinity . Intelligence itself, or a substantial subset if it, is easy, given infinite resources, as AIXI shows. But that’s been of no use in developing real world AI: tractable approximations to AIXI aren’t powerful enough to be dangerous.
It would be embarrassing to MIRI if someone cobbled together AI smart enough to be dangerous, and came to the worlds experts on AI safety for some safety features, only to be told “sorry guys, we haven’t got anything that’s compatible with your system, because it’s finite”.
What’s high value again?
It’s arguably been useful in building models of AI safety. To quote Exploratory Engineering in AI:
I feel as though you’re engaging in pedantry for pedantry’s sake. The point is that if we can’t even solve the simplified version of the problem, there’s no way we’re going to solve the hard version—effectively, it’s saying that you have to crawl before you can walk. Your response was to point out that walking is more useful than crawling, which is really orthogonal to the problem here—the problem being, of course, the fact that we haven’t even learned to crawl yet. AIXI and Bayes are useful in that solving AGI problems in the context provided can act as a “stepping stone” to larger and bigger problems. What are you suggesting as an alternative? That MIRI tackle the bigger problems immediately? That’s not going to work.
You are still assuming that infinite systems count as simple versions of real world finite systems, but that is the assumption I am challenging: our best real world AIs aren’t cut down AIXI systems, they are something different entirely, so there is no linear progression from crawling to walking in your terms,
That’s not just an assumption; that’s the null hypothesis, the default position. Sure, you can challenge it if you want, but if you do, you’re going to have to provide some evidence why you think there’s going to be a qualitative difference. And even if there is some such difference, it’s still unlikely that we’re going to get literally zero insights about the problem from studying AIXI. That’s an extremely strong absolute claim, and absolute claims are almost always false. Ultimately, if you’re going to criticize MIRI’s approach, you need to provide some sort of plausible alternative, and right now, unfortunately, it doesn’t seem like there are any. As far as I can tell, AIXI is the best way to bet.
I have already pointed out that the best AI systems currently existing are not cut down infinite systems.
Something doesn’t have to be completely worthless to be sub optimal.