What Happens When Many AIs Are Connected Together?

The result might be a network that looks like the internet, except where routers are replaced with AIs. That could be good, or bad.


IBM Deep Blue is old news and so is Watson. While few of us understand how these artificially intelligent machines (AIs) actually work, we have a good feel for what they do. Most of us will be also able to appreciate what a self driving car is supposed to do. There are also lots of other AI’s: Google Now, Dragon Naturally Speaking (which is the AI I’m using to dictate this article) and self-piloting drones, and many more.

If we include all the other AI’s that underpin the daily workings of the financial markets, our communications networks and logistics infrastructures, then there are already thousands of different ‘AIs’ each optimized for a specific function and all operating independently.

But we are now starting to see something new: here and there, bit by bit some of these AI’s are being connected together which raises major questions about where this could lead. Here are some high-profile examples of where AI-AI interconnections are already happening.

Google: Combining AIs to richly describe pictures

In November last year Google research scientists announced an innovation which combined two previously separate AI’s in order to richly describe images in words:

The AIs that were connected together are quite different: one is a convolution neural network (CNN) - which identifies objects within an image - and the other is a recurrent neural network (RNN) - which can translate languages.

Firstly, here are examples of the sort of objects that Google’s CNN can identify in arbitrary photographs:

Google’s idea was to take the individual words which the CNN had extracted from a picture and then feed those words into the RNN, which would then try to ‘translate’ those words into a single sentence.

The idea was that if enough words were provided to the RNN for a given picture then it would fill in the missing words to produce a full sentence which described the picture. 

After a lot of work this is what it did:

This is quite a remarkable achievement and while there are still errors, the fact that both component parts are based on deep learning algorithms means that the accuracy will improve over time - more data, more human training and faster processing will inevitably mean that eventually, this type of technology will become very accurate indeed.

This technology is already being extended to help Google image search: by richly describing images that do not have tags it will be possible for users to more quickly find a given image.

It does not take too much imagination to see how this concept could be extended to video content, which comprises a series of pictures which are shown one after the other. A future version of Google’s software could produce a rich narrative describing the content of a video which could then be programmatically analysed to extract useful information, such as which actors appeared or responses to a search query like “swimming with dolphins”.

Microsoft: Real-time spoken language translation 

In December last year Microsoft announced a new feature for Skype, Skype Translator, which can translate spoken Spanish into spoken English and vice versa in near real-time. This currently works by first translating spoken English into English text (dictation AI) and then English text into Spanish text (translation AI). The Spanish text is then spoken out loud at the other end in Spanish (read-aloud AI). The process works the same in reverse and the overall speed is sufficient for two speakers to hold a normal conversation with minimal delay.

With 40 additional languages available for pairing based on displaying text (rather than spoken voice), the system is already quite capable.

Skype Translator is based on AI technologies that have been in development for many years at Microsoft’s research facility and combine a number of different deep learning technologies.

After demonstrating the new product at Code Conference 2014, Microsoft CEO Satya Nadella noted that it was displaying some strange, although beneficial characteristics:

“You teach in English, it learns English. Then you teach it Mandarin — it learns Mandarin, but it becomes better at English. And then you teach it Spanish and it gets good at Spanish, but it gets great at both Mandarin and English. And quite frankly none of us know exactly why.”

The significance of this is not just that multiple AI’s are being combined to produce remarkable results but also that certain aspects of the performance of the combined structure are not understood, were not predictable and are surprising to the creators.

The trend is towards increasingly good outcomes, but with a growing risk of bad outcomes

From the examples outlined above it seems clear that AI-AI interconnectivity could produce many positive outcomes, and quite probably some astonishingly positive outcomes. 

But one does not have to think too hard to see that, equally, the wrong sort of AI-AI interconnectivity could produce undesirable outcomes.

Here are a few other types of AI that could be connected together:

  • Family and friends AI (analyses billions of photographs to establish relationships between individuals in different contextual settings);
     
  • Self driving AI (provides driverless taxi and road transportation services);
     
  • Face recognition AI (allows real-time mapping of a face to a given person and their complete demographic and behavioural profile);
     
  • Criminal profiling AI (predicts criminal activity before it is actually happened);
     
  • Hacker AI (able to out-hack the best human hackers)

There are obviously a number of other AI’s which we could add to this list, which are being worked on by the military or those who have malicious intent.

To see how things could go sideways, consider the scenario which has been well articulated by many commentators in the field of self driving cars.

The scenario is that a self driving car is suddenly faced with a life or death decision. There are two options but both options will mean that someone will be seriously injured. 

The debate so far about this moral and ethical dilemma has focused on the rules that the AI should follow: should the AI take no particular action and let fate decide the outcome? Or should it make a random selection about which life to save?

But this assumes that the self-driving car AI is acting as an autonomous object that is somehow isolated from the connected world within which it exists. But this is incorrect: it will clearly have a network connection which will be essential to obtain navigational information and also to provide feedback about its own environment, for example local road conditions.

So what happens if the self driving car AI asks the network for assistance? 

The self driving car has high-resolution cameras so maybe it could take photographs of the two individuals who are going to be involved in this tragedy and provide their approximate locations to a network level profiling AI which would then be able to identify these individuals. This AI could then send their identities to another AI who would have access to richer data about both candidate victims.

What say one person was found to be a 28-year-old woman who was pregnant with twins and was a leading scientist in the field of clinical embryology. Let’s say the other person was a 38-year-old single male who was being investigated for serious criminal wrongdoing and who had a terminal illness.

Bear in mind that this level of insight could be obtained in a few milliseconds – well before the accident has actually occurred.

The point here is that at some point with a sufficient level of AI-AI interconnectivity it may be that AIs begin making complex judgements like this – without us even knowing.

Hyper-innovation through AI-AI interconnectivity

The rate at which the commercial benefits that come from AI-AI interconnectivity, will increase very rapidly because three types of benefit are being multiplied together:

  • Individual AIs get better: The first type of commercial benefit comes from improvements which allow existing AI’s to become more capable. One example would be incremental improvements in AI algorithms. Another example would be where AIs improve as a result of gaining access to more training data. Another example would be a breakthrough in AI theory which would then have a positive impact on all AI’s at the same time as this innovation was adopted on a  sector-wide basis;
     
  • New types of AI emerge: The second type of benefit comes from the development of new types of AI which will become possible as technology advances. For example, cloud services as we understand them today will very soon reach the level where it becomes feasible to deliver AI functions over the web to third-party developers. So rather than advanced AI’s being confined to powerful network computers, these capabilities will be provided as a commodity service to developers worldwide.

    As another example, personal devices will have storage capacities of 100s of TB in a few decades. At this point it will be feasible for every user to have a personal AI with capabilities vastly exceeding the capabilities of IBM Watson today. 
     
  • AI-AI Interconnectivity: The third type of benefit comes from the connections between different AI’s: Metcalfe’s law was originally proposed as a way to explain why the value of a telecommunications network increases with the number of connections, but it also applies in this case – albeit at a much higher level. Every new AI that is added to the network could become accessible by all other AI’s. Likewise, each new AI that has been added can access the capabilities of all AI’s in the network.

Because these three effects are all mutually-reinforcing, the overall impact on the economy will be multiplicative not additive, which leads to an exponentially increasing rate of AI innovation, or equally, an exponentially increasing commercial impact.

We are already moving towards of network of AIs

One way of looking at what is happening here is to think about the development of the Internet. The Internet is essentially a single network comprising thousands of individual data networks. Each constituent network is broadly as portrayed below:

The purpose of the network is to allow computer A to send information to tablet B. The information to be sent is divided up into packets which are then individually routed across the network and reassembled at the other end. Routing algorithms allow individual data packets to take arbitrary paths through the network. The key functions performed by the transport layer of the Internet involve disassembly and reassembly of information and the automatic routing of packets.

As more and more data networks were deployed it became clear to some that benefits would result from connecting the networks together, so that data packets could ‘hop’ from one network to the next in order to cover great distances, without having to build a special network that covered that distance.

The act of connecting lots of similar networks together results in benefits that transcend those which can be harvested if the networks work in isolation.

In this way, the Internet essentially consists of tens of thousands of data networks which are connected together at thousands of points (technically called points of presence, or POPs or at Internet Exchange Points, of which there are a few hundred).

Importantly, while the Internet is complicated in terms of its scale and scope, it is not complicated in terms of the rules which govern its underlying behaviour. Hence, there is no possibility of it exhibiting unpredictable behaviour.

But an AI 'network' is very different to a data network because its purpose is to behave in ways which are not entirely predictable. 

In a way the more predictable an AI’s behaviour is, the less ‘intelligent’ that behaviour must be. Equally, assuming that the AI operates within a behavioural boundary that has been defined by the architect, then the more surprising the outcomes are - the more valuable they might be.

The above diagram shows that this intrinsic lack of complete predictability could be caused by a feedback mechanism whereby the output is fed back to adjust how the result was obtained in the first place. Unpredictability might also come from rules that are extremely complex.

Unlike the way traffic flows around a data network, which is based on each router in the network making what are essentially very simple decisions, some AI’s use rule sets or self-learning algorithms that are extremely sophisticated and which lead to complex and unpredictable behaviour.

Just as was the case with the internet – when individual data networks started to be connected together, we are now beginning to see the same thing happen with AIs.

The examples regarding Google and Microsoft outlined above clearly show that we are already beginning to see several, formally autonomous AIs being connected together to produce more valuable outcomes than were possible when the individual AI’s were used separately.

It seems likely that this process will continue until, perhaps, the world comprises hundreds of thousands, or millions and eventually billions of different AI’s which are all connected together, one way or another. Each of the constituent AI will be an expert within its given field and within that field it will be producing outcomes which are not predictable.

It is either exciting or worrying to wonder what might happen if a number of such unpredictable outcomes were combined together and used as part of the input conditions for other AI’s.

But we do not have to think very far into the future to get a feel for what AI-AI interactions might look like: very soon we will see the arrival of the first cloud-based AI services. These will be offered by companies like Google, Microsoft, Amazon, HP, IBM and possibly Facebook – to an army of developers, worldwide. These developers have been used to using cloud services to host their websites and to provide on demand computing, but soon they will be able to access deep learning algorithms and other AI related technologies which will be delivered on tap over the web, as a commodity service. Developers will be able to provide services that are, in effect, mash ups that combine multiple AIs.

When the market reaches this point and there are, let’s say, 10 providers of such services, then the market will begin accelerating because competitive pressures will force innovation among the 'cloud AI' service providers.

Further ahead, when mobile devices become powerful enough to host their own AI’s, they will be able to interwork with network level AI’s in the same way that smartphone and tablet apps interwork with network level services today.

An open letter was published in January 2015 by the Future of Life Institute which rightly called for research into a number of critical areas which are applicable to ongoing developments in the field of artificial intelligence. We note that the specific area of AI-AI connectivity is not explicitly mentioned in the 8-page document which suggests a range of candidate areas research.

If we return to the analogy of the Internet – which is a network of data networks – then we see that computer scientists and engineers developed a broad range of technical specifications which defined how the different data networks interwork with each other (e.g. TCP/IP, MPLS, HTTP, SMTP etc.).  But in the case of the emerging network of AI’s, which we could call the ‘InterAI’, there are no such specifications, or even an understanding of how such specifications should be developed.

It is not a question of whether we will see AI-AI interconnectivity occurring on a mass scale but simply when - and I feel we should understand more about the implications before that point arrives.