The number of insights before superintelligence


The number of insights before superintelligence

A transcript of the above talk

Hi! I'm Tim Tyler - and this is a video about the number of insights required in order to produce superintelligent machines.

This is of interest since it affects the time required to produce superintelligent machines, our level of uncertaintly about that time, the rate of change when such machines are developed, and the best strategey to employ in producing them.

Examples

To illustrate what I mean with some examples:

If a sudden theoretical breakthrough allows superintelligent machines to be produced, then this might happen at almost any time - and it increases the liklihood of rapid change when that happens.

On the other hand, if the development of superintelligent machines takes a long and complex engineering effort, then the duration of that effort will be easier to predict - based on observations of the progress so far. Also, the impact of the onset of superintelligent machines will probably be spread out over a longer period of time.

Analysis

So: will the development of superintelligent machines involve a small number of large leaps - or many small incremental steps?

Insight into this question can be gained by considering the insights needed in other projects - both completed projects and unfinished ones.

Also, it may be possible to get some insight by considering the structure of the remaining part of the problem.

One issue with comparing against other problems is that it is difficult to know which problems to select for comparison.

Some obvious examples include powered flight, the atom bomb, space flight, the development of printing, telephony, computers, and the internet.

Consider powered flight, for example. Flight is a classical instance of a technological problem which has already been mastered. There we see a few basic principles, some insights about how to apply them, and then a complex engineering project based on them.

One might say that the basic principle of flight involves creating a low-pressure area above a wing, and the key insight is that that can be done by giving a cross section through the wing a longer upper side. However, this then leads to a relatively complex engineering project, which implements the idea - which itself has many associated sub-problems.

If machine intelligence is like flight, then probably the main insights are that intelligence is an adaptive learning process, that such a learning process can be instantiated by a self-organising system consisting of large numbers of relatively simple computing components.

This formulation places the key insights in the past - with the implication that we are currently in the "extended engineering project" phase of the project to construct machine intelligence.

However, flight might not be a useful problem to compare against. It certainly turned out to be a much easier problem - and we solved it long ago. Maybe the problems which we find harder to solve are characterised by complexity and a large numbers of interacting components.

One issue is that there are many types of problem with differing levels of serial insight required.

Another issue iss that the problems we have solved may be systematically different from the problems that remain before us - for example they might be simpler and easier.

Machine intelligence so far

Looking at the progress of specific machine intelligence applications usually leads to the idea that progress is a long and painful process. Speech recognition is an obvious example of this. People have been trying to write speech recognition programs for a very long time, and progress has been slow, in a manner not entirely explicable by hardware limitations.

Progress in machine intelligence in general leads to a similar conclusion. People have been trying for a long time - and progress is being made - but it is slow and gradual.

Magic bullets

What is the case that a sudden breakthrough will be made? This is sometimes known as the idea that there is a "magic bullet".

In my view, the best case for a magic bullet is based on the idea that much of the brain is composed of large numbers of relatively simple cells that form a neural network. Each cell is not that complicated - and the brain's properties come from linking billions of these cells together.

If we can understand roughly how the brain does what it does, then we can duplicate it - and thus produce machine intelligence.

There are three main sub-problems to this approach:

  • Figuring out how the various types of synapses work;

  • Figuring out how axon growth and atrophy operate;

  • Figuring out how the various types of neurons work.

These problems are not trivial, but it seems likely that we will eventually solve them - and then we will probably be able to create synthetic "grey matter" - functionally similar to the human neo-cortex.

This would indeed be an important step forwards - but it would not solve the entire problem. There are other inportant architectural features of the brain. In particular, there is attention, there is a heirarchical memory structure - with short and long term memory, there are emotions, there is the preprocessing circuitry associated with each of the senses, and the post-processing circuitry associated with motor control.

However, maybe understanding all this material is not necessary. Perhaps there is some kind of short-cut?

My own view is that it isn't possible to completely rule out these "magic bullet" scenarios - but the history of attempts to build machine intelligence does not seem to offer them much support.

Perhaps some day, future historians will look back, and wonder why we didn't find the secret of intelligence earlier. However, it doesn't currently look as though we are going to make machine intelligence by discovering a few secrets - even if those secrets are in some sense "out there". Rather, it looks as though it will be a long hard slog.

The structure of the remaining problem

Lastly, can we gain anything by an examination of the structure of the remaining problem?

Again, looking from the direction of the brain lets us see at least a little of the problem's structure.

Of the remaining mysteries there, one big one is attention. Attention represents a systematic shutting down of perceptual promotion into consciousness in most areas of the brain. We can see why its useful, we can crudely fake the effect, but we don't yet know how best to implement it.

Another mystery is memory architecture. Humans have multiple stages of short, medium and long-term memory, things are shuffled into long-term memory while we are asleep. We don't yet really understand how best to implement such a storage medium.

Also, there are the emotions. Presumably we will not want synthetic intelligent agents feeling anger, lust and jealousy. Indeed, most emotions look like crude biochemical hacks that we could probably improve on substantially. Some emotional states beyond pleasure and pain may well be needed - but which ones?

Another possible issue is the wirehead problem. It seems possible that - once intelligent agents learn how to modify themselves - they will either get paranoid about self-modification and avoid it, or else be over-enthusiastic, and mess themselves up. It seems likely that they will eventually find the line between these extremes, but it may not be easy to get them to this point initially.

Of these, only the last looks like it might be resistant to research and development - and we probably don't need to crack that problem in order to develop pretty sophisticated machine intelligences.

Conclusions

If the analysis here is correct, some of the implications would seem to be:

  • Machine intelligence will not suddenly enjoy a massive competitive advantage over humans when it arrives. Instead, each generation of machine intelligence will find itself competing with the products of the previous generation. Machines will compete against last years machines - and humans will incrementally augment their own intelligence - by preprocessing their senses with machines, and post-processing their motor outputs with machines.

  • Machine intelligence is not terribly close at hand - there still seems to be quite a way to go;

  • The time of arrival of machine intelligence should be relatively predictable - based on what we already know about progress so far;

  • Machine intelligence is likely to be produced by a large organisation with considerable resources to devote to research and development. Reserchers working on the problem in their basements and garages should probably look into getting hired if they actually want to contribute.

Enjoy,


Tim Tyler | Contact | http://alife.co.uk/