Posts

AI: The Story So Far – Stuart Russell

stuart russell - redAwesome to have Stuart Russell discussing AI Safety – a very important topic. Too long have people been associating the idea of AI safety issues with Terminator – unfortunately the human condition seems such that people often don’t give themselves permission to take seriously non-mainstream ideas unless they see a tip of the hat from an authority figure.

During the presentation Stuart brings up a nice quote by Norbert Wiener:

If we use, to achieve our purposes, a mechanical agency with whose operation we cannot efficiently interfere once we have started it, because the action is so fast and irrevocable that we have not the data to intervene before the action is complete, then we had better be quite sure that the purpose put into the machine is the purpose which we really desire and not merely a colorful imitation of it.Norbert Wiener

P.s. Stuart Russell co-authored AI A Modern Approach with Peter Norvig – arguably the most popular textbook on AI theory.

The lecture was presented at the 2016 Colloquium Series on Robust and Beneficial AI (CSRBAI) hosted by the Machine Intelligence Research Institute (MIRI) and Oxford’s Future of Humanity Institute (FHI).

What I’m finding is that senior people in the field who have never publicly evinced any concern before are privately thinking that we do need to take this issue very seriously, and the sooner we take it seriously the better.Stuart Russell

Video of presentation:

 

The field [of AI] has operated for over 50 years on one simple assumption: the more intelligent, the better. To this must be conjoined an overriding concern for the benefit of humanity. The argument is very simple:

1. AI is likely to succeed.
2. Unconstrained success brings huge risks and huge benefits.
3. What can we do now to improve the chances of reaping the benefits and avoiding the risks?

Some organizations are already considering these questions, including the Future of Humanity Institute at Oxford, the Centre for the Study of Existential Risk at Cambridge, the Machine Intelligence Research Institute in Berkeley, and the Future of Life Institute at Harvard/MIT. I serve on the Advisory Boards of CSER and FLI.

Just as nuclear fusion researchers consider the problem of containment of fusion reactions as one of the primary problems of their field, it seems inevitable that issues of control and safety will become central to AI as the field matures. The research questions are beginning to be formulated and range from highly technical (foundational issues of rationality and utility, provable properties of agents, etc.) to broadly philosophical.

– Stuart Russell (Quote Source)

Metamorphogenesis – How a Planet can produce Minds, Mathematics and Music – Aaron Sloman

The universe is made up of matter, energy and information, interacting with each other and producing new kinds of matter, energy, information and interaction.
How? How did all this come out of a cloud of dust?
In order to find explanations we first need much better descriptions of what needs to be explained.

By Aaron Sloman
Abstract – and more info – Held at Winter Intelligence Oxford – Organized by the Future of Humanity Institute

Aaron Sloman

Aaron Sloman

This is a multi-disciplinary project attempting to describe and explain the variety of biological information-processing mechanisms involved in the production of new biological information-processing mechanisms, on many time scales, between the earliest days of the planet with no life, only physical and chemical structures, including volcanic eruptions, asteroid impacts, solar and stellar radiation, and many other physical/chemical processes (or perhaps starting even earlier, when there was only a dust cloud in this part of the solar system?).

Evolution can be thought of as a (blind) Theorem Prover (or theorem discoverer).
– Proving (discovering) theorems about what is possible (possible types of information, possible types of information-processing, possible uses of information-processing)
– Proving (discovering) many theorems in parallel (including especially theorems about new types of information and new useful types of information-processing)
– Sharing partial results among proofs of different things (Very different biological phenomena may share origins, mechanisms, information, …)
Combining separately derived old theorems in constructions of new proofs (One way of thinking about symbiogenesis.)
– Delegating some theorem-discovery to neonates and toddlers (epigenesis/ontogenesis). (Including individuals too under-developed to know what they are discovering.)
– Delegating some theorem-discovery to social/cultural developments. (Including memes and other discoveries shared unwittingly within and between communities.)
– Using older products to speed up discovery of new ones (Using old and new kinds of architectures, sensori-motor morphologies, types of information, types of processing mechanism, types of control & decision making, types of testing.)

The “proofs” of discovered possibilities are implicit in evolutionary and/or developmental trajectories.

They demonstrate the possibility of development of new forms of development, evolution of new types of evolution learning new ways to learn evolution of new types of learning (including mathematical learning: by working things out without requiring empirical evidence) evolution of new forms of development of new forms of learning (why can’t a toddler learn quantum mechanics?) – how new forms of learning support new forms of evolution amd how new forms of development support new forms of evolution (e.g. postponing sexual maturity until mate-selection mating and nurturing can be influenced by much learning)
….
…. and ways in which social cultural evolution add to the mix

These processes produce new forms of representation, new ontologies and information contents, new information-processing mechanisms, new sensory-motor
morphologies, new forms of control, new forms of social interaction, new forms of creativity, … and more. Some may even accelerate evolution.

A draft growing list of transitions in types of biological information-processing.

An attempt to identify a major type of mathematical reasoning with precursors in perception and reasoning about affordances, not yet replicated in AI systems.

Even in microbes I suspect there’s much still to be learnt about the varying challenges and opportunities faced by microbes at various stages in their evolution, including new challenges produced by environmental changes and new opportunities (e.g. for control) produced by previous evolved features and competences — and the mechanisms that evolved in response to those challenges and opportunities.

Example: which organisms were first able to learn about an enduring spatial configuration of resources, obstacles and dangers, only a tiny fragment of which can be sensed at any one time?
What changes occurred to meet that need?

Use of “external memories” (e.g. stigmergy)
Use of “internal memories” (various kinds of “cognitive maps”)

More examples to be collected here.