How long before superintelligence?

How long before superintelligence?

A transcript of the above talk

Hi, I'm Tim Tyler, and today I'll be addressing the question of: "How long before superintelligence?"

What is superintelligence?

Firstly, to explain the terminology: a superintelligence is an agent which is vastly smarter than a human in most domains.

Are companies superintelligences?

Perhaps the nearest thing we have to superintelligence today are companies. A company can posess intelligence which exceeds that of an individual human in many domains. However companies don't really qualify as as superintelligent agents, because they're not smart enough across a wide enough range of different problems. Companies may attain super-human performance in some areas - and yet exhibit relatively poor performance in other ones.

Often, a company is only as smart as its smartest employee. For example, consider the case of ability at the game of go. A company may be able to play go better than its best employee - but it will probably not be a superintelligent go player - simply because there is no known effective algorithm for parallelizing the problem of playing go and distributing it across multiple employees.

This issue can be illustrated by a diagram:

A company usually consists of a network of individual human brains. There are communication bottlenecks between the brains and algorithms must split a problem into modular chunks in order to run efficiently over such a network. A genuine superintelligence would probably not have that architecture. It would dispense with the communications bottlenecks between the nodes - and thus be able to handle larger problems without requiring that they first be divided up.

When will genuine superintelligence arrive?

If we do not have superintelligence today, when will it arrive? I'll argue that superintelligence will come soon after synthetic intelligence reaches human performance levels.

Once human-level synthetic intelligence is attained, machines will be able to contribute extensively to the research and development needed to produce the next generation of intelligent machines - thus accelerating their development. Also, by the time we have human-level machine intelligence, progress is likely to be taking place rapidly - simply on the grounds that technological development is constantly accelerating.

When will we get broadly human-level performance from intelligent machines?

So, the next question is: when will we get broadly human-level performance from intelligent machines?

Hardware considerations

One way in which this question has traditionally been addressed is by looking at the hardware requirements for something with broadly-equivalent functionality to the human brain.

Hans Moravec produced one of the first estimates of when this might happen - way back in the 1980s. He looked at the computational properties of cells in the retina performing edge detection and motion detection, and compared these to some electronic signal processing equipment which he considered to perform a similar function. He came up with some MIPS-per-neuron figures, and then extrapolated from these to produce an estimate representing the computational power of the entire human brain.

Other teams have subsequently analysed other mental subsystems - including parts of the auditory cortex and the cerebellum - and come up with broadly comparable figures.

Plotting these on a graph with the computational power of existing computers suggests that human-level computing hardware will become available in supercomputers around 2010 - and will become available in desktops around 2020.

Software considerations

However - according to most researchers in the field - hardware is not the limiting factor. We don't know how to utilise the hardware we currently have, let alone what will be available in 2020.

It is common for computer software to lag behind the capabilities of its associated hardware. The effect will be familiar to anyone who has owned a computer games console near the time of its launch. Software contains the complex parts of the system, and it is those which are difficult to create and maintain.

So: hardware estimates provide a lower bound for when we will have superintelligence - but are not the whole story. Are there estimates of how difficult it would be to create the required software?

The brain and the genome

Ray Kurzweil has attempted such an estimate - arguing that the brain's design is of a managable size - since it is stored in the human genome. Here he is at Stanford in 2006, making his case:

[Ray Kurzweil footage]

However, this argument was criticised by Douglas Hofstadter - at the same event.

[Douglas Hofstadter footage]

So, who is right? Does the brain's design fit into the genome? - or not?

The detailed form of proteins arises from a combination of the nucleotide sequence that specifies them, the cytoplasmic environment in which gene expression takes place, and the laws of physics.

We can safely ignore the contribution of cytoplasmic inheritance - however, the contribution of the laws of physics is harder to discount. At first sight, it may seem simply absurd to argue that the laws of physics contain design information relating to the construction of the human brain. However there is a well-established mechanism by which physical law may do just that - an idea known as the anthropic principle. This argues that the universe we observe must necessarily permit the emergence of intelligent agents. If that involves a coding the design of the brains of intelligent agents into the laws of physics then: so be it. There are plenty of apparently-arbitrary constants in physics where such information could conceivably be encoded: the fine structure constant, the cosmological constant, Planck's constant - and so on.

At the moment, it is not even possible to bound the quantity of brain-design information so encoded. When we get machine intelligence, we will have an independent estimate of the complexity of the design required to produce an intelligent agent. Alternatively, when we know what the laws of physics are, we may be able to bound the quantity of information encoded by them. However, today neither option is available to us.

Anyway, even if Kurzweil was right - and the design of the human brain fits onto a CD-ROM - that would still represent an utterly enormous search space - which would easily take far longer than the history of the universe to exhaustively search. So this whole approach doesn't really permit us to say anything useful about how difficult the overall problem is.

Argument from evolution

What about the idea that the human brain evolved over the last 600 million years - in the time since the Cambrian explosion? Doesn't that give us a clue about how hard the problem is? If we can apply a fudge factor to account for the advantage of engineering design over blind mutations - and another fudge factor to account for the fact that we can crib from nature's solution to the problem, doesn't that help us estimate the level of difficulty? Unfortunately, no: the anthropic principle again blocks this kind of approach. We have no idea whether our brain evolved via a series of lucky flukes - we don't know if our own evolution was typical - or not. For example, if we imagine a constraint that says that a big meteorite destroys all multi-cellular life every 700 million years, the evolution of intelligent life had better take place within that timescale - no matter how many apparently-improbable events that involves.

Reverse-engineering the human brain

How about reverse-engineering the human brain? If we can estimate how hard that problem is won't that at least give us an upper bound on how difficult the overall problem is? Yes, but reverse-engineering the human brain looks pretty tricky - and the task might take us well into the second half of this century. So while such an approach does - in principle - allow us to bound the difficulty of the problem, the resulting bound is not very tight. We will have probably machine intelligence long before such projects gets very far off the ground.

Intelligence testing

So, what does all that leave us with? It mostly leaves us with intelligence testing. We can test the intelligence of humans - and of machines. We can plot the increase of machine intelligence over time, and see when it reaches human level.

I think this is the most promising approach to estimating when we will first have access to superintelligent machines, on other grounds besides sheer hardware capability.

So, what intelligence tests exist - where we already have access to historical data about machine intelligence? One such area is computer chess. Before Deep Blue beat Kasparov it was possible to look at the history of chess computer ratings, and plot a graph that showed when computers would be able to beat the best humans. That illustrates the potential of the technique - but it also highlights one of its limitations - any individual test may "crap out" before broad human-level intelligence is reached, and fall to a machine intelligence that specialises in solving the test.

The most obvious solution is to use multiple tests, from a variety of domains: strategy games, robot control, compression, speech recognition, language translation, IQ tests - and so on.

One of the best tests we currently have is the game of go - a classical oriental board game. Playing go exercises pattern recognition circuitry in the human visual cortex - which represents a substantial proportion of the human brain by volume. It is a taxing game for both humans and machines. Also there is a rating scale whose uniformty and linearity is backed up by a handicap system, and there is a long history of computer players and tournaments.

Currently the best computer go programs rate at around 1 or 2 kyu - on a scale that goes from beginners at 30-kyu down to 1-kyu, and then from 1-dan up to 9-dan - which represents world champion level.

As an intelligence test, Go may "crap out" before broad human-level intelligence is reached - but to me, it looks as though it may well take us to within spitting distance of the target.

Go is probably the best single way of measuring intelligence for which we have a good history of machine intelligence in which to ground extrapolations.

A 1997 survey of go programmers produced a wide range of answers to the question of when a program would be world champion.

19992005Mei-Kou Tei 9-dan [P]20102150Kobayashii
20002010Darren Cook [P]20152035David Keeble
20002010Chihiro Mizuuchi20172090Hiroyuki 8-dan
20002010Naritatsu Yamamoto20202050Mick Reiss [P]
20022040Martin Mueller [P]20202100Jay Burmeister
20052050Hirooka20202100Chen Zhixing [P]
20052100Amano20202100Prof. Hsu
20072097Ken Chen [P]20202100Yoshikawa
20072097Kojima 9-dan20202200Oyama
20072197Redmond 8-dan20302050Masahiro Okazaki
20102023Shinichi Sei [P]20502100Kim
20102030Tristan Cazenave [P]2100-Izuka
20102045David Fotland [P]22002500Fujisawa
20102050Yung Jye Hunag [P]25003000Oyaizu
20102100Jun Saito

Even with go - where we have some of the best available evidence, the situation is not yet clear-cut.

My estimate

To finish this talk, I'd like to present a slide which illustrates my estimate of when we will obtain superintelligent machines:

Tim Tyler's estimate

The curve is a probabilty density function, illustrating the probability of superintelligence first arising on the specified date.

A roughly-bell-shaped curve, peaking around 2025 - with a fairly substantial spread - indicating my estimate of my level of uncertainty about the issue.

Other estimates

I've listed estimates from those interested in the issue enough to produce probability density functions.

David Lucifer's estimate [source]

Shane Legg's estimate [source]

Roko's estimate [source - now private]

Ben Goertzel poll [source]

Bruce K poll [source]

Winter intelligence Conference 2010 [source]

Winter intelligence Conference 2010 [source]

Michael Vassar [source]
This graph has an unusual shape - and the y-axis seems to be miscalibrated.

Less Wrong 2011 Survey [source]
I discarded data before 2010 and above 2150 on the grounds that these were outliers.
The question was:
"By what year do you think the Singularity will occur?
Answer such that you think there is an even chance of the Singularity falling before or after that year.
If you don't think a Singularity will ever happen, leave blank."
Note: each point (rather misleadingly) represents data for the next 10 years.


These are in reverse date order.

When Will AI Be Created? - Luke Muehlhauser (2013)

AI overlords and their deciles (2011)

When will we get our robot overlords? (2011)

How long until human-level AI? - 2010 Ben Goertzel paper on the subject (2010)

Artificial General Intelligence (AGI) coming soon? (2010)

The Maes-Garreau Point - Kevin Kelly (2007)

[AI@50] First Poll (2006)


Tim Tyler | Contact |