Will robots outsmart us? by the late Stephen Hawkins

There is a interesting article, "Will robots outsmart us?" in today's Sunday Times Magazine. While I don't accept all Stephen's predictions I was most interested to read:

When an artificial intelligence (AI) becomes better than humans at AI design, so that it can recursively improve itself without human help, we may face an intelligence explosion that ultimately results in machines whose intelligence exceeds ours by more than ours exceeds that of snails. When that happens, we will need to ensure that the computers have goals aligned with ours.

Later he says:

In short, the advent of super-intelligent AI would be either the best or the worst thing ever to happen to humanity. The real risk with AI isn’t malice, but competence. A super-intelligent AI will be extremely good at accomplishing its goals, and if those goals aren’t aligned with ours we’re in trouble. You’re probably not an evil ant-hater who steps on ants out of malice, but if you’re in charge of a hydroelectric green-energy project and there’s an anthill in the region to be flooded, too bad for the ants. Let’s not place humanity in the position of those ants. 

Of course we know what happened last time a super-intelligence came into existence. About half a million years ago the Earth was populated by a great variety of animals of a comparatively low intelligence. All the higher animals had brains that worked in roughly the same way, and how much they could learn was limited because everything they learnt was lost when they died. Then one species, which we call Homo sapiens, discovered a way to recursively increase its own intelligence. It was already good at making tools but for several million years the cost of trail and error learning had limited what it could do. But then it invented a tool to boost intelligence, which we call language.  Language not only made it possible to make better tools, but also it made it possible to recursively build a better language generation by generation. So some 5000 generations later the Earth is home to a super-intelligent species ...

And are the goals of this species aligned with the the goals of the millions of other species? Of course not. Billions of animals are kept as slaves to be killed and eaten, while the homes of countless more have been, or are being, destroyed. 

If we invent a super-intelligent AI system why should it treat us with more respect than we have  shown for our animal relatives.

A new book "Brief Answers to the Big Questions," by Stephen Hawkins, is published by John Murry on Tuesday.

Why does CODIL differ from other computer languages

This query came up on a FutureLearn Course which read "Christopher, to be honest I don't think the world needs any more computer languages or most of the ones it already has for that matter"

The vast majority of computer languages, such as COBOL, Fortran, C, Java, etc. are designed to process information on a conventional stored program computer where the memory consists of numbered boxes which contain numbers . The numbers in the boxes may represent coded data (often several different formats), numeric addresses, or coded instructions. This approach was originally designed to handle a class of well-defined mathematical tasks which humans find difficult to do quickly and accuracy, so it is not surprising that modern computers are incomprehensible black boxes when viewed by the average human being. They were deliberately designed to do efficiently things which people are bad at doing.

CODIL uses a completely different memory structure which is based on a dynamic network which attempts to mimic the working of the brain's neural network. The aim is to produce a transparent information processor (rather than a black box) which is easy for the average human to understand and use for a range of potentially complex information processing tasks. It is particularly aimed at complex tasks where a dynamically flexible human interface is advantageous - and so fills the area where conventional computers are weakest.

In CODIL the array of numbered boxes which make up a conventional computer memory is replaced by a large number of nodes, where each node consists of an on/off switch and a label (which is for the benefit of the human user). The human user defines the names of the nodes and the wires linking the nodes.

How plans for a user-friendly computer were rubbished 50 years ago

In 1968 David Caminer and John Pinkerton (who were responsible for the world's first business computer, the LEO I, and who were directors of English Electric Computer) decided to fund research into a project to build inherently user-friendly computers and it was estimated that the market for such systems would be several hundred million pounds a year.

However, as a result of the government inspired merger to make the UK computer industry more competitive, ICL was created, and the project was closed down with no serious attempt to assess the successful research into CODIL which had already been carried out.

This was, of course, about 10 years before the first personal computers, and it is interesting to speculate what might have happened if the research had not been so rudely interrupted. Perhap the UK based idea would have been successful and the first personal computers would have been inherently friendly. This would have meant that there would have been no need for the hard to use MS-DOS operating system - and no one would have heard of Microsoft.

To get details of what happened read 

The EEC ICL Story

To see how the idea originated read

The Unlikely Origins of my current Evolutionary Research

This is relevant to this blog because the rubbished 1968 proposal is orign of the model of human working memory at the heart to the current research into human short term memory.

Can you model for the impact of language, writing, etc. ???

Your Questions Answered

The Evolutionary Model of Intelligence in 200 words!

JS asked; "Can you model for the impact of language, writing, print, radio, TV and the internet?"

I am basically interested in modelling the flow of information within brains and, using a language, between human brains. It is important to note that the model is not concerned with the physical form of the decision making nodes or the messages. An eye is a decision making node which receives photons and converts them into messages to neurons in the brain. In the same way books, TVs and computer systems can be considered to be integral part of the overall network model. 

One of billions of possible complex information flow examples: The node “William Shakespeare” generated a message “Macbeth” which was sent to by the book to actor nodes living 400 years later – and the resulting performance ended up in your brain via a TV and your eye. 

My model is (at present) only an initial qualitative model – but the model shows how improved information tools can have an effect on the knowledge stored in the brains of modern humans. It is an open question whether the model could be developed to measure impact by, for instance, saying what percentage of a given individual’s knowledge came from the use of the internet.

How far has computing advanced in the last 50 years?

Computers have changed significantly over the last 50 years - but have some of the early problems been solved. The following brief excerpts came from a survey I prepared in 1968 about the difficulties that were being reported when trying to computerise complex tasks involving both computers and people:

J Nievergelt wrote:

“We are pushing against the limits of complexity that we know how to handle. ... However, we also run into problems that are commputationally light but also so complex that we do not know how to design, document, and debug them. This kind of problem is relatively new. – this complexity barrier may well be the most important limitation that the computer field will be subject to in the immediate future.”

Wesley Davis wrote:

“The ultimate success or failure of computer systems at present is overdependent on the problem definition and system design stage, The desire to simplify systems leads to a conflict between the needs of people and the computer specialist's interest in containing the project within present methodology. ... for its [computer system] purpose is to make a company responsive, flexible, and aware of its commercial situation. It is unlikely to do this without these characteristics being implicit in its design.

M A Jackson wrote:

“We should be positively looking for and developing programming methods that do a1low inconsistency, redundancy, ambiguity and incompleteness; we should recognise that these seem to be vices only because the error-prone techniques of procedural programming make them so.”

Seymour Papert wrote:

“Machines can't think,” said he, “because stupid humans don't know how to teach them to think. We try and teach them as we think machines should be taught, and it doesn't work.”

Computers have changed enormously since 1968, and we nearlly all have, in our pockets or bags, a mobile phone which is thousands of times more powerful than the computer shown in the above picture. Computers now provide a very wide range of usebuf services, Dispite all the advances, have they really solved the problems posed by complex human systems?

What do you think?

See a transcript of the original 1968 suvey

The Limitations of Existing Computer Techniques.

The Evolutionary Model of Intelligence in 200 words!

Picture from HITXP blog

My “biological computer” model of the evolution of human intelligence involves information flowing between “decision making nodes” in a infinite recursive network. Every node can receive, process, and transmit messages and may consist of clusters of simpler nodes. The simplest nodes represent individual neurons, with groups of interlinked neurons up to complete brains, individual animals, and extending to social groups of animals or humans exchanging information (including large organizations) and tools made by humans – such as the Internet. “Animal” nodes have a limited lifespan and (if the brain makes good decisions) can pass genetic information to new, initially ignorant, “animal” nodes.

The driving mechanism is based on CODIL, a computer language that mimics human working memory. It suggests how information can be used to make decisions within brains, and how it can be passed between brains so that it is not lost on death. It identifies limits on animal intelligence, predicts failings of the human mind, and identifies key tipping points in the development of tools and language, and brain size changes. Human intelligence results from recycling cultural information through the network over many thousands of generations.

Of course such a short text only scratches the surface, and further information on my research, and supporting details will be appearing on this blog, if it is not already here. 

If you can't find what you want to know why not 

ask a question.

If only this skull could still talk ...

If only this skull
could still talk

I am always happy to answer questions about how my research relates to a particular research paper and in a discussion on the FutureLearn course "A Question of Time" Paul asked about "On the antiquity of language: the reinterpretation of Neandertal linguistic capacities and its consequences" by Dan Dediu and Stephen C. Levinson. This paper concludes that Neanderthals and Denisovans may have had very similar language capacities to us - they too had complex tool making technologies. Paul asked:

When did the conditions arise to allow more rapid evolution of information (technical, such as for tools but also hunting & food gathering & processing techniques, cultural info etc.) ? After all, we possess much the same genes as our ancestors did 60,000 years ago, but our world is significantly different due to evolution of information.

Dediu and Levinson's paper looks in detail at the relevant literature about the discover of human fossils, evidence for tool making , etc. while my approach starts in a very different way by looking at how a network of neurons might evolve into an intelligent brain. The immediately relevant parts of my model are as follows: