Tag Archives: AI

Some trees just don’t get barked up

Now and then, someone asks me for an old document and as I search for it, I stumble across others I’d forgotten about. I’ve been rather frustrated that AI progress hasn’t kept up with its development rate in the 90s, so this was fun to rediscover, highlighting some future computing directions that offered serious but uncertain potential exactly 20 years ago, well 20 years ago 3 weeks ago. Here is the text, and the Schrodinger’s Computer was only ever intended to be silly (since renamed the Yonck Processor):

Herrings, a large subset of which are probably red

Computers in the future will use a wide range of techniques, not just conventional microprocessors. Problems should be decomposed and the various components streamed to the appropriate processing engines. One of the important requirements is therefore some means of identifying automatically which parts of a problem could best be tackled by which techniques, though sometimes it might be best to use several in parallel with some interaction between them.

 Analogs

We have a wider variety of components available to be used in analog computing today than we had when it effectively died out in the 80s. With much higher quality analog and mixed components, and additionally micro-sensors, MEMs, simple neural network components, and some imminent molecular capability, how can we rekindle the successes of the analog domain. Nature handles the infinite body problem with ease! Things just happen according to the laws of physics. How can we harness them too? Can we build environments with synthetic physics to achieve more effects? The whole field of non-algorithmic computation seems ripe for exploitation.

 Neural networks

  • Could we make neural microprocessor suspensions, using spherical chips suspended in gel in a reflective capsule and optical broadcasting. Couple this with growing wires across the electric field. This would give us both electrical and optical interconnection that could be ideal for neural networks with high connectivity. Could link this to gene chip technology to have chemical detection and synthesis on the chips too, so that we could have close high speed replicas of organic neural networks.
  • If we can have quantum entanglement between particles, might this affect the way in which neurons in the brain work? Do we have neural entanglement and has this anything to do with how our brain works. Could we create neural entanglement or even virtual entanglement and would it have any use?
  • Could we make molecular neurons (or similar) using ordinary chemistry? And then form them into networks. Might need nanomachines and bottom-up assembly.
  • Could we use neurons as the first stage filters to narrow down the field to make problems tractable for other techniques
  • Optical neurons
  • Magnetic neurons

Electromechanical, MEMS etc

  • Micromirror arrays as part of optical computers, perhaps either as data entry, or as part of the algorithm
  • Carbon fullerene balls and tubes as MEM components
  • External fullerene ‘décor’ as a form of information, cf antibodies in immune system
  • Sensor suspensions and gels as analog computers for direct simulation

Interconnects

  • Carbon fullerene tubes as on chip wires
  • Could they act as electron pipes for ultra-high speed interconnect
  • Optical or radio beacons on chip

Software

  • Transforms – create a transform of every logic component, spreading the functionality across a wide domain, and construct programs using them instead. Small perturbation is no longer fatal but just reduces accuracy
  • Filters – nature works often using simple physical effects where humans design complex software. We need to look at hard problems to see how we might make simple filters to narrow the field before computing final details and stages conventionally.
  • Interference – is there some form of representation that allows us to compute operations by means of allowing the input data to interact directly, i.e. interference, instead of using tedious linear computation. Obviously only suited to a subset of problems.

And finally, the frivolous

  • Schrodinger’s computer – design of computer and software, if any, not determined until box is opened. The one constant is that it destroys itself if it doesn’t finding the solution. All possible computers and all possible programs exist and if there is a solution, the computer will pop out alive and well with the answer. Set it the problem of answering all possible questions too, working out which ones have the most valuable answers and using up all the available storage to write the best answers.

Who controls AI, controls the world

This week, the fastest supercomputer broke a world record for AI, using machine learning in climate research:

https://www.wired.com/story/worlds-fastest-supercomputer-breaks-ai-record/

I guess most readers thought this is a great thing, after all we need to solve climate change. That wasn’t my thought. The first thing my boss told me when I used a computer for the first time was: “shit in, shit out”. I don’t remember his name but I remember that concise lesson every time I read about climate models. If either the model or the data is garbage, or both, the output will also be garbage.

So my first thought reading about this new record was: will they let the AI work everything out for itself using all the raw, unadjusted data available about the environment, including all the astrophysics data about every kind of solar activity, human agricultural, industrial activities, air travel, all the unadjusted measurements of or proxies for surface, sea and air temperatures, ever collected, any empirical evidence for any corrections that might be needed on such data in any direction, and then let it make its own deductions, form its own models of how it might all connected and then watch eagerly as it makes predictions?

Or will they just input their own models, CO2 blinkering, prejudices and group-think, adjusted datasets, data omissions and general distortions of historical records into biased models already indoctrinated with climate change dogma, so that it will reconfirm the doom and gloom forecasts we’re so used to hearing, maximizing their chances of continued grants? If they do that, the AI might as well be a cardboard box with a pre-written article stuck on it. Shit in, shit out.

It’s obvious that the speed and capability of the supercomputer is of secondary important to who controls the AI, and its access to data, and its freedom to draw its own conclusions.

(Read my blog on Fake AI: https://timeguide.wordpress.com/2017/11/16/fake-ai/)

You may recall a week or two ago that IBM released a new face database to try to address bias in AI face recognition systems. Many other kinds of data could have biases for all sorts of reasons. At face value reducing bias is a good thing, but what exactly do we mean by that? Who decides what is biased and what is real? There are very many potential AI uses that are potentially sensitive, such as identifying criminals or distinguishing traits that correlate with gender, sexuality, race, religion, or indeed any discernible difference. Are all deductions by the AI permissible, or are huge swathes of possible deductions not permitted because they might be politically unacceptable? Who controls the AI? Why? With what aims?

Many people have some degree of influence on  AI. Those who provide funding, equipment, theoreticians, those who design hardware, those who design the learning and training mechanisms, those who supply the data, those who censor or adjust data before letting the AI see it, those who design the interfaces, those who interpret and translate the results, those who decide which results are permissible and how to spin them, and publish them.

People are often impressed when a big powerful computer outputs results of massive amounts of processing. Outputs may often be used to control public opinion and government policy, to change laws, to alter balance of power in society, to create and destroy empires. AI will eventually make or influence most decisions of any consequence.

As AI techniques become more powerful, running on faster and better computers, we must always remember that golden rule: shit in, shit out. And we must always be suspicious of those who might have reason to influence an outcome.

Because who controls AI, controls the world.

 

 

Future AI: Turing multiplexing, air gels, hyper-neural nets

Just in time to make 2018 a bit less unproductive, I managed to wake in the middle of the night with another few inventions. I’m finishing the year on only a third as many as 2016 and 2017, but better than some years. And I quite like these new ones.

Gel computing is a very old idea of mine, and I’m surprised no company has started doing it yet. Air gel is different. My original used a suspension of processing particles in gel, and the idea was that the gel would hold the particles in fixed locations with good free line of sight to neighbor devices for inter-device optical comms, while acting also as a coolant.

Air gel uses the same idea of suspending particles, but does so by using ultrasound, standing waves holding the particles aloft. They would form a semi-gel I suppose, much softer. The intention is that they will be more easily movable than in a gel, and maybe rotate. I imagine using rotating magnetic fields to rotate them, and use that mechanism to implement different configurations of inter-device nets. That would be the first pillar of running multiple neural nets in the same space at the same time, using spin-based TDM (time division multiplexing), or synchronized space multiplexing if you prefer. If a device uses on board processing that is fast compared to the signal transmission time to other devices (the speed of light may be fast but can still be severely limiting for processing and comms), then having the ability to deal with processing associated with several other networks while awaiting a response allows a processing network to be multiplied up several times. A neural net could become a hyper-neural net.

Given that this is intended for mid-century AI, I’m also making the assumption that true TDM can also be used on each net, my second pillar. Signals would carry a stream of slots holding bits for each processing instance. Since this allows a Turing machine to implement many different processes in parallel, I decided to call it Turing multiplexing. Again, it helps alleviate the potential gulf between processing and communication times. Combining Turing and spin multiplexing would allow a single neural net to be multiplied up potentially thousands or millions of times – hyper-neurons seems as good a term as any.

The third pillar of this system is that the processing particles (each could contain a large number of neurons or other IT objects) could be energized and clocked using very high speed alternating EM fields – radio, microwaves, light, even x-rays. I don’t have any suggestions for processing mechanisms that might operate at such frequencies, though Pauli switches might work at lower speeds, using Pauli exclusion principle to link electron spin states to make switches. I believe early versions of spin cubits use a similar principle. I’m agnostic whether conventional Turing machine or quantum processing would be used, or any combination. In any case, it isn’t my problem, I suspect that future AIs will figure out the physics and invent the appropriate IT.

Processing devices operating at high speed could use a lot of energy and generate a lot of heat, and encouraging the system to lase by design would be a good way to cool it as well as powering it.

A processor using such mechanisms need not be bulky. I always assumed a yogurt pot size for my gel computer before and an air gel processor could be the same, about 100ml. That is enough to suspend a trillion particles with good line of sight for optical interconnections, and each connection could utilise up to millions of alternative wavelengths. Each wavelength could support many TDM channels and spinning the particles multiplies that up again. A UV laser clock/power source driving processors at 10^16Hz would certainly need to use high density multiplexing to make use of such a volume, with transmission distances up to 10cm (but most sub-mm) otherwise being a strongly limiting performance factor, but 10 million-fold WDM/TDM is attainable.

A trillion of these hyper-neurons using that multiplexing would act very effectively as 10 million trillion neurons, each operating at 10^16Hz processing speed. That’s quite a lot of zeros, 35 of them, and yet each hyperneuron could have connections to thousands of others in each of many physical configurations. It would be an obvious platform for supporting a large population of electronically immortal people and AIs who each want a billion replicas, and if it only occupies 100ml of space, the environmental footprint isn’t an issue.

It’s hard to know how to talk to a computer that operates like a brain, but is 10^22 times faster, but I’d suggest ‘Yes Boss’.

 

Biomimetic insights for machine consciousness

About 20 years ago I gave my first talk on how to achieve consciousness in machines, at a World Future Society conference, and went on to discuss how we would co-evolve with machines. I’ve lectured on machine consciousness hundreds of times but never produced any clear slides that explain my ideas properly. I thought it was about time I did. My belief is that today’s deep neural networks using feed-forward processing with back propagation training can not become conscious. No digital algorithmic neural network can, even though they can certainly produce extremely good levels of artificial intelligence. By contrast, nature also uses neurons but does produce conscious machines such as humans easily. I think the key difference is not just that nature uses analog adaptive neural nets rather than digital processing (as I believe Hans Moravec first insighted, a view that I readily accepted) but also that nature uses large groups of these analog neurons that incorporate feedback loops that act both as a sort of short term memory and provide time to sense the sensing process as it happens, a mechanism that can explain consciousness. That feedback is critically important in the emergence of consciousness IMHO. I believe that if the neural network AI people stop barking up the barren back-prop tree and start climbing the feedback tree, we could have conscious machines in no time, but Moravec is still probably right that these need to be analog to enable true real-time processing as opposed to simulation of that.

I may be talking nonsense of course, but here are my thoughts, finally explained as simply and clearly as I can. These slides illustrate only the simplest forms of consciousness. Obviously our brains are highly complex and evolved many higher level architectures, control systems, complex senses and communication, but I think the basic foundations of biomimetic machine consciousness can be achieved as follows:

That’s it. I might produce some more slides on higher level processing such as how concepts might emerge, and why in the long term, AIs will have to become hive minds. But they can wait for later blogs.

AI that talks to us could quickly become problematic

Google’s making the news again adding evidence to the unfortunate stereotype of the autistic IT nerd that barely understands normal people, and they have therefore been astonished at the backlash that normal people would all easily have predicted. (I’m autistic and work in IT mostly too, and am well used to the stereotype it so it doesn’t bother me, in fact it is a sort of ‘get out of social interactions free’ card). Last time it was Google Glass, where it apparently didn’t occur to them that people may not want other people videoing them without consent in pubs and changing rooms. This time it is Google Duplex, that makes phone calls on your behalf to arrange appointment using voice that is almost indistinguishable from normal humans. You could save time making an appointment with a hairdresser apparently, so the Googlanders decided it must be a brilliant breakthrough, and expected everyone to agree. They didn’t.

Some of the objections have been about ethics: e.g. An AI should not present itself as human – Humans have rights and dignity and deserve respectful interactions with other people, but an AI doesn’t and should not masquerade as human to acquire such privilege without knowledge of the other party and their consent.

I would be more offended by the presumed attitude of the user. If someone thinks they are so much better then me that they can demand my time and attention without the expense of any of their own, delegating instead to a few microseconds of processing time in a server farm somewhere, I’ll treat them with the contempt they deserve. My response will not be favourable. I am already highly irritated by the NHS using simple voice interaction messaging to check I will attend a hospital appointment. The fact that my health is on the line and notices at surgeries say I will be banned if I complain on social media is sufficient blackmail to ensure my compliance, but it still comes at the expense of my respect and goodwill. AI-backed voice interaction with better voice wouldn’t be any better, and if it asking for more interaction such as actually booking an appointment, it would be extremely annoying.

In any case, most people don’t speak in fully formed grammatically and logically correct sentences. If you listen carefully to everyday chat, a lot of sentences are poorly pronounced, incomplete, jumbled, full of ums and er’s, likes and they require a great deal of cooperation by the listener to make any sense at all. They also wander off topic frequently. People don’t stick to a rigid vocabulary list or lists of nicely selected sentences.  Lots of preamble and verbal meandering is likely in a response that is highly likely to add ambiguity. The example used in a demo, “I’d like to make a hairdressing appointment for a client” sounds fine until you factor in normal everyday humanity. A busy hairdresser or a lazy receptionist is not necessarily going to cooperate fully. “what do you mean, client?”, “404 not found”, “piss off google”, “oh FFS, not another bloody computer”, “we don’t do hairdressing, we do haircuts”, “why can’t your ‘client’ call themselves then?” and a million other responses are more likely than “what time would you like?”

Suppose though that it eventually gets accepted by society. First, call centers beyond the jurisdiction of your nuisance call blocker authority will incessantly call you at all hours asking or telling you all sorts of things, wasting huge amounts of your time and reducing quality of life. Voice spam from humans in call centers is bad enough. If the owners can multiply productivity by 1000 by using AI instead of people, the result is predictable.

We’ve seen the conspicuous political use of social media AI already. Facebook might have allowed companies to use very limited and inaccurate knowledge of you to target ads or articles that you probably didn’t look at. Voice interaction would be different. It uses a richer emotional connection that text or graphics on a screen. Google knows a lot about you too, but it will know a lot more soon. These big IT companies are also playing with tech to log you on easily to sites without passwords. Some gadgets that might be involved might be worn, such as watches or bracelets or rings. They can pick up signals to identify you, but they can also check emotional states such as stress level. Voice gives away emotion too. AI can already tell better then almost all people whether you are telling the truth or lying or hiding something. Tech such as iris scans can also tell emotional states, as well as give health clues. Simple photos can reveal your age quite accurately to AI, (check out how-old.net).  The AI voice sounds human, but it is better then even your best friends at guessing your age, your stress and other emotions, your health, whether you are telling the truth or not, and it knows far more about what you like and dislike and what you really do online than anyone you know, including you. It knows a lot of your intimate secrets. It sounds human, but its nearest human equivalent was probably Machiavelli. That’s who will soon be on the other side of the call, not some dumb chatbot. Now re-calculate political interference, and factor in the political leaning and social engineering desires of the companies providing the tools. Google and Facebook and the others are very far from politically neutral. One presidential candidate might get full cooperation, assistance and convenient looking the other way, while their opponent might meet rejection and citation of the official rules on non-interference. Campaigns on social issues will also be amplified by AI coupled to voice interaction. I looked at some related issue in a previous blog on fake AI (i.e. fake news type issues): https://timeguide.wordpress.com/2017/11/16/fake-ai/

I could but won’t write a blog on how this tech could couple well to sexbots to help out incels. It may actually have some genuine uses in providing synthetic companionship for lonely people, or helping or encouraging them in real social interactions with real people. It will certainly have some uses in gaming and chatbot game interaction.

We are not very far from computers that are smarter then people across a very wide spectrum, and probably not very far from conscious machines that have superhuman intelligence. If we can’t even rely on IT companies to understand likely consequences of such obvious stuff as Duplex before thy push it, how can we trust them in other upcoming areas of AI development, or even closer term techs with less obvious consequences? We simply can’t!

There are certainly a few such areas where such technology might help us but most are minor and the rest don’t need any deception, but they all come at great cost or real social and political risk, as well as more abstract risks such as threats to human dignity and other ethical issues. I haven’t give this much thought yet and I am sure there must be very many other consequences I have not touched on yet. Google should do more thinking before they release stuff. Technology is becoming very powerful, but we all know that great power comes with great responsibility, and since most people aren’t engineers so can’t think through all the potential technology interactions and consequences, engineers such as Google’s must act more responsibly. I had hoped they’d started, and they said they had, but this is not evidence of that.

 

Futurist memories: The leisure society and the black box economy

Things don’t always change as fast as we think. This is a piece I wrote in 1994 looking forward to a fully automated ‘black box economy, a fly-by-wire society. Not much I’d change if I were writing it new today. Here:

The black box economy is a strictly theoretical possibility, but may result where machines gradually take over more and more roles until the whole economy is run by machines, with everything automated. People could be gradually displaced by intelligent systems, robots and automated machinery. If this were to proceed to the ultimate conclusion, we could have a system with the same or even greater output as the original society, but with no people involved. The manufacturing process could thus become a ‘black box’. Such a system would be so machine controlled that humans would not easily be able to pick up the pieces if it crashed – they would simply not understand how it works, or could not control it. It would be a fly-by-wire society.

The human effort could be reduced to simple requests. When you want a new television, a robot might come and collect the old one, recycling the materials and bringing you a new one. Since no people need be involved and the whole automated system could be entirely self-maintaining and self-sufficient there need be no costs. This concept may be equally applicable in other sectors, such as services and information – ultimately producing more leisure time.

Although such a system is theoretically possible – energy is free in principle, and resources are ultimately a function of energy availability – it is unlikely to go quite this far. We may go some way along this road, but there will always be some jobs that we don’t want to automate, so some people may still work. Certainly, far fewer people would need to work in such a system, and other people could spend their time in more enjoyable pursuits, or in voluntary work. This could be the leisure economy we were promised long ago. Just because futurists predicted it long ago and it hasn’t happened yet does not mean it never will. Some people would consider it Utopian, while others possibly a nightmare, it’s just a matter of taste.

How can we make a computer conscious?

This is very text heavy and is really just my thinking out loud, so to speak. Unless you are into mental archaeology or masochistic, I’d strongly recommend that you instead go to my new blog on this which outlines all of the useful bits graphically and simply.

Otherwise….

I found this article in my drafts folder, written 3 years ago as part of my short series on making conscious computers. I thought I’d published it but didn’t. So updating and publishing it now. It’s a bit long-winded, thinking out loud, trying to derive some insights from nature on how to make conscious machines. The good news is that actual AI developments are following paths that lead in much the same direction, though some significant re-routing and new architectural features are needed if they are to optimize AI and achieve machine consciousness.

Let’s start with the problem. Today’s AI that plays chess, does web searches or answers questions is digital. It uses algorithms, sets of instructions that the computer follows one by one. All of those are reduced to simple binary actions, toggling bits between 1 and 0. The processor doing that is no more conscious or aware of it, and has no more understanding of what it is doing than an abacus knows it is doing sums. The intelligence is in the mind producing the clever algorithms that interpret the current 1s and 0s and change them in the right way. The algorithms are written down, albeit in more 1s and 0s in a memory chip, but are essentially still just text, only as smart and aware as a piece of paper with writing on it. The answer is computed, transmitted, stored, retrieved, displayed, but at no point does the computer sense that it is doing any of those things. It really is just an advanced abacus. An abacus is digital too (an analog equivalent to an abacus is a slide rule).

A big question springs to mind: can a digital computer ever be any more than an advanced abacus. Until recently, I was certain the answer was no. Surely a digital computer that just runs programs can never be conscious? It can simulate consciousness to some degree, it can in principle describe the movements of every particle in a conscious brain, every electric current, every chemical reaction. But all it is doing is describing them. It is still just an abacus. Once computed, that simulation of consciousness could be printed and the printout would be just as conscious as the computer was. A digital ‘stored program’ computer can certainly implement extremely useful AI. With the right algorithms, it can mine data, link things together, create new data from that, generate new ideas by linking together things that haven’t been linked before, make works of art, poetry, compose music, chat to people, recognize faces and emotions and gestures. It might even be able to converse about life, the universe and everything, tell you its history, discuss its hopes for the future, but all of that is just a thin gloss on an abacus. I wrote a chat-bot on my Sinclair ZX Spectrum in 1983, running on a processor with about 8,000 transistors. The chat-bot took all of about 5 small pages of code but could hold a short conversation quite well if you knew what subjects to stick to. It’s very easy to simulate conversation. But it is still just a complicated abacus and still doesn’t even know it is doing anything.

However clever the AI it implements, a conventional digital computer that just executes algorithms can’t become conscious but an analog computer can, a quantum computer can, and so can a hybrid digital/analog/quantum computer. The question remain s whether a digital computer can be conscious if it isn’t just running stored programs. Could it have a different structure, but still be digital and yet be conscious? Who knows? Not me. I used to know it couldn’t, but now that I am a lot older and slightly wiser, I now know I don’t know.

Consciousness debate often starts with what we know to be conscious, the human brain. It isn’t a digital computer, although it has digital processes running in it. It also runs a lot of analog processes. It may also run some quantum processes that are significant in consciousness. It is a conscious hybrid of digital, analog and possibly quantum computing. Consciousness evolved in nature, therefore it can be evolved in a lab. It may be difficult and time consuming, and may even be beyond current human understanding, but it is possible. Nature didn’t use magic, and what nature did can be replicated and probably even improved on. Evolutionary AI development may have hit hard times, but that only shows that the techniques used by the engineers doing it didn’t work on that occasion, not that other techniques can’t work. Around 2.6 new human-level fully conscious brains are made by nature every second without using any magic and furthermore, they are all slightly different. There are 7.6 billion slightly different implementations of human-level consciousness that work and all of those resulted from evolution. That’s enough of an existence proof and a technique-plausibility-proof for me.

Sensors evolved in nature pretty early on. They aren’t necessary for life, for organisms to move around and grow and reproduce, but they are very helpful. Over time, simple light, heat, chemical or touch detectors evolved further to simple vision and produce advanced sensations such as pain and pleasure, causing an organism to alter its behavior, in other words, feeling something. Detection of an input is not the same as sensation, i.e. feeling an input. Once detection upgrades to sensation, you have the tools to make consciousness. No more upgrades are needed. Sensing that you are sensing something is quite enough to be classified as consciousness. Internally reusing the same basic structure as external sensing of light or heat or pressure or chemical gradient or whatever allows design of thought, planning, memory, learning and construction and processing of concepts. All those things are just laying out components in different architectures. Getting from detection to sensation is the hard bit.

So design of conscious machines, and in fact what AI researchers call the hard problem, really can be reduced to the question of what makes the difference between a light switch and something that can feel being pushed or feel the current flowing when it is, the difference between a photocell and feeling whether it is light or dark, the difference between detecting light frequency, looking it up in a database, then pronouncing that it is red, and experiencing redness. That is the hard problem of AI. Once that is solved, we will very soon afterwards have a fully conscious self aware AI. There are lots of options available, so let’s look at each in turn to extract any insights.

The first stage is easy enough. Detecting presence is easy, measuring it is harder. A detector detects something, a sensor (in its everyday engineering meaning) quantifies it to some degree. A component in an organism might fire if it detects something, it might fire with a stronger signal or more frequently if it detects more of it, so it would appear to be easy to evolve from detection to sensing in nature, and it is certainly easy to replicate sensing with technology.

Essentially, detection is digital, but sensing is usually analog, even though the quantity sensed might later be digitized. Sensing normally uses real numbers, while detection uses natural numbers (real v  integer as programmer call them). The handling of analog signals in their raw form allows for biomimetic feedback loops, which I’ll argue are essential. Digitizing them introduces a level of abstraction that is essentially the difference between emulation and simulation, the difference between doing something and reading about someone doing it. Simulation can’t make a conscious machine, emulation can. I used to think that meant digital couldn’t become conscious, but actually it is just algorithmic processing of stored programs that can’t do it. There may be ways of achieving consciousness digitally, or quantumly, but I haven’t yet thought of any.

That engineering description falls far short of what we mean by sensation in human terms. How does that machine-style sensing become what we call a sensation? Logical reasoning says there would probably need to be only a small change in order to have evolved from detection to sensing in nature. Maybe something like recombining groups of components in different structures or adding them together or adding one or two new ones, that sort of thing?

So what about detecting detection? Or sensing detection? Those could evolve in sequence quite easily. Detecting detection is like your alarm system control unit detecting the change of state that indicates that a PIR has detected an intruder, a different voltage or resistance on a line, or a 1 or a 0 in a memory store. An extremely simple AI responds by ringing an alarm. But the alarm system doesn’t feel the intruder, does it?  It is just a digital response to a digital input. No good.

How about sensing detection? How do you sense a 1 or a 0? Analog interpretation and quantification of digital states is very wasteful of resources, an evolutionary dead end. It isn’t any more useful than detection of detection. So we can eliminate that.

OK, sensing of sensing? Detection of sensing? They look promising. Let’s run with that a bit. In fact, I am convinced the solution lies in here so I’ll look till I find it.

Let’s do a thought experiment on designing a conscious microphone, and for this purpose, the lowest possible order of consciousness will do, we can add architecture and complexity and structures once we have some bricks. We don’t particularly want to copy nature, but are free to steal ideas and add our own where it suits.

A normal microphone sensor produces an analog signal quantifying the frequencies and intensities of the sounds it is exposed to, and that signal may later be quantified and digitized by an analog to digital converter, possibly after passing through some circuits such as filters or amplifiers in between. Such a device isn’t conscious yet. By sensing the signal produced by the microphone, we’d just be repeating the sensing process on a transmuted signal, not sensing the sensing itself.

Even up close, detecting that the microphone is sensing something could be done by just watching a little LED going on when current flows. Sensing it is harder but if we define it in conventional engineering terms, it could still be just monitoring a needle moving as the volume changes. That is obviously not enough, it’s not conscious, it isn’t feeling it, there’s no awareness there, no ‘sensation’. Even at this primitive level, if we want a conscious mic, we surely need to get in closer, into the physics of the sensing. Measuring the changing resistance between carbon particles or speed of a membrane moving backwards and forwards would just be replicating the sensing, adding an extra sensing stage in series, not sensing the sensing, so it needs to be different from that sort of thing. There must surely need to be a secondary change or activity in the sensing mechanism itself that senses the sensing of the original signal.

That’s a pretty open task, and it could even be embedded in the detecting process or in the production process for the output signal. But even recognizing that we need this extra property narrows the search. It must be a parallel or embedded mechanism, not one in series. The same logical structure would do fine for this secondary sensing, since it is just sensing in the same logical way as the original. This essential logical symmetry would make its evolution easy too. It is easy to imagine how that could happen in nature, and easier still to see how it could be implemented in a synthetic evolution design system. Such an approach could be mimicked in natural or synthetic evolutionary development systems. In this approach, we have to feel the sensing, so we need it to comprise some sort of feedback loop with a high degree of symmetry compared with the main sensing stage. That would be natural evolution compatible as well as logically sound as an engineering approach.

This starts to look like progress. In fact, it’s already starting to look a lot like a deep neural network, with one huge difference: instead of using feed-forward signal paths for analysis and backward propagation for training, it relies instead on a symmetric feedback mechanism where part of the input for each stage of sensing comes from its own internal and output signals. A neuron is not a full sensor in its own right, and it’s reasonable to assume that multiple neurons would be clustered so that there is a feedback loop. Many in the neural network AI community are already recognizing the limits of relying on feed-forward and back-prop architectures, but web searches suggest few if any are moving yet to symmetric feedback approaches. I think they should. There’s gold in them there hills!

So, the architecture of the notional sensor array required for our little conscious microphone would have a parallel circuit and feedback loop (possibly but not necessarily integrated), and in all likelihood these parallel and sensing circuits would be heavily symmetrical, i.e. they would use pretty much the same sort of components and architectures as the sensing process itself. If the sensation bit is symmetrical, of similar design to the primary sensing circuit, that again would make it easy to evolve in nature too so is a nice 1st principles biomimetic insight. So this structure has the elegance of being very feasible for evolutionary development, natural or synthetic. It reuses similarly structured components and principles already designed, it’s just recombining a couple of them in a slightly different architecture.

Another useful insight screams for attention too. The feedback loop ensures that the incoming sensation lingers to some degree. Compared to the nanoseconds we are used to in normal IT, the signals in nature travel fairly slowly (~200m/s), and the processing and sensing occur quite slowly (~200Hz). That means this system would have some inbuilt memory that repeats the essence of the sensation in real time – while it is sensing it. It is inherently capable of memory and recall and leaves the door wide open to introduce real-time interaction between memory and incoming signal. It’s not perfect yet, but it has all the boxes ticked to be a prime contender to build thought, concepts, store and recall memories, and in all likelihood, is a potential building brick for higher level consciousness. Throw in recent technology developments such as memristors and it starts to look like we have a very promising toolkit to start building primitive consciousness, and we’re already seeing some AI researchers going that path so maybe we’re not far from the goal. So, we make a deep neural net with nice feedback from output (of the sensing system, which to clarify would be a cluster of neurons, not a single neuron) to input at every stage (and between stages) so that inputs can be detected and sensed, while the input and output signals are stored and repeated into the inputs in real time as the signals are being processed. Throw in some synthetic neurotransmitters to dampen the feedback and prevent overflow and we’re looking at a system that can feel it is feeling something and perceive what it is feeling in real time.

One further insight that immediately jumps out is since the sensing relies on the real time processing of the sensations and feedbacks, the speed of signal propagation, storage, processing and repetition timeframes must all be compatible. If it is all speeded up a million fold, it might still work fine, but if signals travel too slowly or processing is too fast relative to other factors, it won’t work. It will still get a computational result absolutely fine, but it won’t know that it has, it won’t be able to feel it. Therefore… since we have a factor of a million for signal speed (speed of light compared to nerve signal propagation speed), 50 million for switching speed, and a factor of 50 for effective neuron size (though the sensing system units would be multiple neuron clusters), we could make a conscious machine that could think at 50 million times as fast as a natural system (before allowing for any parallel processing of course). But with architectural variations too, we’d need to tune those performance metrics to make it work at all and making physically larger nets would require either tuning speeds down or sacrificing connectivity-related intelligence. An evolutionary design system could easily do that for us.

What else can we deduce about the nature of this circuit from basic principles? The symmetry of the system demands that the output must be an inverse transform of the input. Why? Well, because the parallel, feedback circuit must generate a form that is self-consistent. We can’t deduce the form of the transform from that, just that the whole system must produce an output mathematically similar to that of the input.

I now need to write another blog on how to use such circuits in neural vortexes to generate knowledge, concepts, emotions and thinking. But I’m quite pleased that it does seem that some first-principles analysis of natural evolution already gives us some pretty good clues on how to make a conscious computer. I am optimistic that current research is going the right way and only needs relatively small course corrections to achieve consciousness.

 

Why superhumans are inevitable, and what else comes in the box

Do we have any real choice in the matter of making  super-humans? 20 years ago, I estimated 2005 as the point of no return, and nothing since then has changed my mind on that date. By my reckoning, we are already inevitably committed to designer babies, ebaybies, super-soldiers and super-smart autonomous weapons, direct brain-machine links, electronic immortality, new human races, population explosion, inter-species conflicts and wars with massively powerful weaponry, superhuman conscious AI, smart bacteria, and the only real control we have is relatively minor adjustments on timings. As I was discussing yesterday, the technology potential for this is vast and very exciting, nothing less than a genuine techno-utopia if we use the technologies wisely, but optimum potential doesn’t automatically become reality, and achieving a good outcome is unlikely if many barriers are put in its way.

In my estimation, we have already started the countdown to this group of interconnected technologies – we will very likely get all of them, and we must get ready for the decisions and impacts ahead. At the moment, our society is a small child about to open its super-high-tech xmas presents while fighting with its siblings. Those presents will give phenomenal power far beyond the comprehension of the child or its emotional maturity to equip it to deal with the decisions safely. Our leaders have already squandered decades of valuable preparation time by ignoring the big issues to focus on trivial ones. It is not too late to achieve a good ending, but it won’t happen by accident and we do need to make preparations to avoid pretty big problems.

Both hard and soft warfare – the sword and the pen, already use rapidly advancing AI, and the problems are already running ahead of what the owners intended.

Facebook, Twitter, Instagram and other media giants all have lots of smart people and presumably they mean well, but if so, they have certainly been naive. They maybe hoped to eliminate loneliness, inequality, and poverty and create a loving interconnected global society with global peace, but instead created fake news, social division and conflict and election interference. More likely they didn’t intend either outcome, they just wanted to make money and that took priority over due care and attention..

Miniaturising swarming smart-drones are already the subjects of a new arms race that will deliver almost un-killable machine adversaries by 2050. AI separately is in other arms races to make super-smart AI and super-smart soldiers. This is key to the 2005 point of no return. It was around 2005 that we reached the levels of technology where future AI development all the way to superhuman machine consciousness could be done by individuals, mad scientists or rogue states, even if major powers had banned it. Before 2005, there probably wasn’t quite enough knowledge already on the net to do that. In 2018, lots of agencies have already achieved superiority to humans in niche areas, and other niches will succumb one by one until the whole field of human capability is covered. The first machines to behave in ways not fully understood by humans arrived in the early 1990s; in 2018, neural nets already make lots of decisions at least partly obscured to humans.

This AI development trend will take us to superhuman AI, and it will be able to accelerate development of its own descendants to vastly superhuman AI, fully conscious, with emotions, and its own agendas. That will need humans to protect against being wiped out by superhuman AI. The only three ways we could do that are to either redesign the brain biologically to be far smarter, essentially impossible in the time-frame, to design ways to link our brains to machines, so that we have direct access to the same intelligence as the AIs, so a gulf doesn’t appear and we can remain relatively safe, or pray for super-smart aliens to come to our help, not the best prospect.

Therefore we will have no choice but to make direct brain links to super-smart AI. Otherwise we risk extinction. It is that simple. We have some idea how to do that – nanotech devices inside the brain linking to each and every synapse that can relay electrical signals either way, a difficult but not impossible engineering problem. Best guesses for time-frame fall in the 2045-2050 range for a fully working link that not only relays signals between your organic brain and an IT replica, but by doing so essentially makes external IT just another part of your brain. That conveys some of the other technology gifts of electronic immortality, new varieties of humans, smart bacteria (which will be created during the development path to this link) along with human-variant population explosion, especially in cyberspace, with androids as their physical front end, and the inevitable inter-species conflicts over resources and space – trillions of AI and human-like minds in cyberspace that want to do things in the real world cannot be assumed to be willingly confined just to protect the interests of what they will think of as far lesser species.

Super-smart AI or humans with almost total capability to design whatever synthetic biology is needed to achieve any biological feature will create genetic listings for infinite potential offspring, simulate them, give some of them cyberspace lives, assemble actual embryos for some of them and bring designer babies. Already in 2018, you can pay to get a DNA listing, and blend it in any way you want with the listing of anyone else. It’s already possible to make DNA listings for potential humans and sell them on ebay, hence the term ebaybies. That is perfectly legal, still, but I’ve been writing and lecturing about them since 2004. Today they would just be listings, but we’ll one day have the tech to simulate them, choose ones we like and make them real, even some that were sold as celebrity collector items on ebay. It’s not only too late to start regulating this kind of tech, our leaders aren’t even thinking about it yet.

These technologies are all linked intricately, and their foundations are already in place, with much of the building on those foundations under way. We can’t stop any of these things from happening, they will all come in the same basket. Our leaders are becoming aware of the potential and the potential dangers of the AI positive feedback loop, but at least 15 years too late to do much about it. They have been warned repeatedly and loudly but have focused instead on the minor politics of the day that voters are aware of. The fundamental nature of politics is unlikely to change substantially, so even efforts to slow down the pace of development or to limit areas of impact are likely to be always too little too late. At best, we will be able to slow runaway AI development enough to allow direct brain links to protect against extinction scenarios. But we will not be able to stop it now.

Given inevitability, it’s worth questioning whether there is even any point in trying. Why not just enjoy the ride? Well, the brakes might be broken, but if we can steer the bus expertly enough, it could be exciting and we could come out of it smelling of roses. The weak link is certainly the risk of super-smart AI, whether AI v humans or countries using super-smart AI to fight fiercely for world domination. That risk is alleviated by direct brain linkage, and I’d strongly argue necessitates it, but that brings the other technologies. Even if we decide not to develop it, others will, so one way or another, all these techs will arrive, and our future late century will have this full suite of techs, plus many others of course.

We need as a matter of extreme urgency to fix these silly social media squabbles and over-reactions that are pulling society apart. If we have groups hating each other with access to extremely advanced technology, that can only mean trouble. Tolerance is broken, sanctimony rules, the Inquisition is in progress. We have been offered techno-utopia, but current signs are that most people think techno-hell looks more appetizing and it is their free choice.

AIs of a feather flocking together to create global instability

Hawking and Musk have created a lot of media impact with their warnings about AI, so although terminator scenarios resulting from machine consciousness have been discussed, as have more mundane use of non-conscious autonomous weapon systems, it’s worth noting that I haven’t yet heard them mention one major category of risks from AI – emergence. AI risks have been discussed frequently since the 1970s, and in the 1990s a lot of work was done in the AI community on emergence. Complex emergent patterns of behavior often result from interactions between entities driven by simple algorithms. Genetic algorithms were demonstrated to produce evolution, simple neighbor-interaction rules were derived to illustrate flocking behaviors that make lovely screen saver effects. Cellular automata were played with. In BT we invented ways of self-organizing networks and FPGAs, played with mechanism that could be used for evolution and consciousness, demonstrated managing networks via ANTs – autonomous network telephers, using smart packets that would run up and down wires sorting things out all by themselves. In 1987 discovered a whole class of ways of bringing down networks via network resonance, information waves and their much larger class of correlated traffic – still unexploited by hackers apart from simple DOS attacks. These ideas have slowly evolved since, and some have made it into industry or hacker toolkits, but we don’t seem to be joining the dots as far as risks go.

I read an amusing article this morning by an ex-motoring-editor who was declined insurance because the AI systems used by insurance companies had labelled him as high risk because he maybe associated with people like Clarkson. Actually, he had no idea why, but that was his broker’s theory of how it might have happened. It’s a good article, well written and covers quite a few of the dangers of allowing computers to take control.

http://www.dailymail.co.uk/sciencetech/article-5310031/Evidence-robots-acquiring-racial-class-prejudices.html

The article suggested how AIs in different companies might all come to similar conclusions about people or places or trends or patterns in a nice tidy positive feedback loop. That’s exactly the sort of thing that can drive information waves, which I demonstrated in 1987 can bring down an entire network in less than 3 milliseconds, in such a way that it would continue to crash many times when restarted. That isn’t intended by the algorithms, which individually ought to make good decisions, but when interacting with one another, create the emergent phenomenon.  Automated dealing systems are already pretty well understood in this regard and mechanisms prevent frequent stock market collapses, but that is only one specific type of behavior in one industry that is protected. There do not seem to be any industry-wide mechanisms to prevent the rest of this infinite class of problems from affecting any or all of the rest, simultaneously.

As we create ever more deep learning neural networks, that essentially teach themselves from huge data pools, human understanding of their ‘mindsets’ decreases. They make decisions using algorithms that are understood at a code level, but the massive matrix of derived knowledge they create from all the data they receive becomes highly opaque. Often, even usually, nobody quite knows how a decision is made. That’s bad enough in a standalone system, but when many such systems are connected, produced and owned and run by diverse companies with diverse thinking, the scope for destructive forms of emergence increases geometrically.

One result could be gridlock. Systems fed with a single new piece of data could crash. My 3 millisecond result in 1987 would still stand since network latency is the prime limiter. The first AI receives it, alters its mindset accordingly, processes it, makes a decision and interacts with a second AI. This second one might have different ‘prejudice’ so makes its own decision based on different criteria, and refuses to respond the way intended. A 3rd one looks at the 2nd’s decision and takes that as evidence that there might be an issue, and with its risk-averse mindset, also refuse to act, and that inaction spreads through the entire network in milliseconds. Since the 1st AI thinks the data is all fine and it should have gone ahead, it now interprets the inaction of the others as evidence that that type of data is somehow ‘wrong’ so itself refuses to process any further of that type, whether from its own operators or other parts of the system. So it essentially adds its own outputs to the bad feeling and the entire system falls into sulk mode. As one part of infrastructure starts to shut down, that infects other connected parts and our entire IT could fall into sulk mode – entire global infrastructure. Since nobody knows how it all works, or what has caused the shutdown, it might be extremely hard to recover.

Another possible result is a direct information wave, almost certainly a piece of fake news. Imagine our IT world in 5 years time, with all these super-smart AIs super-connected. A piece of fake news says a nuke has just been launched somewhere. Stocks will obviously decline, whatever the circumstances, so as the news spreads, everyone’s AIs will take it on themselves to start selling shares before the inevitable collapse, triggering a collapse, except it won’t because the markets won’t let that happen. BUT… The wave does spread, and all those individual AIs want to dispose of those shares, or at least find out what’s happening, so they all start sending messages to one another, exchanging data, trying to find what’s going on. That’s the information wave. They can’t sell shares of find out, because the network is going into overload, so they try even harder and force it into severe overload. So it falls over. When it comes back online, they all try again, crashing it again, and so on.

Another potential result is smartass AI. There is always some prat somewhere who sees an opportunity to take advantage and ruins if for everyone else by doing something like exploiting a small loophole in the law, or in this case, most likely, a prejudice our smartass AI has discovered in some other AI that means it can be taken advantage of by doing x, y, or z. Since nobody quite knows how any of their AIs are making their decisions because their mindsets ate too big and too complex, it will be very hard to identify what is going on. Some really unusual behavior is corrupting the system because some AI is going rogue somewhere somehow, but which one, where, how?

That one brings us back to fake news. That will very soon infect AI systems with their own varieties of fake news. Complex networks of AIs will have many of the same problems we are seeing in human social networks. An AI could become a troll just the same as a human, deliberately winding others up to generate attention of drive a change of some parameter – any parameter – in its own favour. Activist AIs will happen due to people making them to push human activist causes, but they will also do it all by themselves. Their analysis of the system will sometimes show them that a good way to get a good result is to cause problems elsewhere.

Then there’s climate change, weather, storms, tsunamis. I don’t mean real ones, I mean the system wide result of tiny interactions of tiny waves and currents of data and knowledge in neural nets. Tiny effects in one small part of a system can interact in unforeseen ways with other parts of other systems nearby, creating maybe a breeze, which interacts with breezes in nearby regions to create hurricanes. I think that’s a reasonable analogy. Chaos applies to neural net societies just as it does to climate, and 50 year waves equivalents will cause equivalent havoc in IT.

I won’t go on with more examples, long blogs are awful to read. None of these requires any self-awareness, sentience, consciousness, call it what you will. All of these can easily happen through simple interactions of fairly trivial AI deep learning nets. The level of interconnection already sounds like it may already be becoming vulnerable to such emergence effects. Soon it definitely will be. Musk and Hawking have at least joined the party and they’ll think more and more deeply in coming months. Zuckerberg apparently doesn’t believe in AI threats but now accepts the problems social media is causing. Sorry Zuck, but the kind of AI you’re company is messing with will also be subject to its own kinds of social media issues, not just in its trivial decisions on what to post or block, but actual inter-AI socializing issues. It might not try to eliminate humanity, but if it brings all of our IT to a halt and prevents rapid recovery, we’re still screwed.

 

Emotion maths – A perfect research project for AI

I did a maths and physics degree, and even though I have forgotten much of it after 36 years, my brain is still oriented in that direction and I sometimes have maths dreams. Last night I had another, where I realized I’ve never heard of a branch of mathematics to describe emotions or emotional interactions. As the dream progressed, it became increasingly obvious that the most suited part of maths for doing so would be field theory, and given the multi-dimensional nature of emotions, tensor field theory would be ideal. I’m guessing that tensor field theory isn’t on most university’s psychology syllabus. I could barely cope with it on a maths syllabus. However, I note that one branch of Google’s AI R&D resulted in a computer architecture called tensor flow, presumably designed specifically for such multidimensional problems, and presumably being used to analyse marketing data. Again, I haven’t yet heard any mention of it being used for emotion studies, so this is clearly a large hole in maths research that might be perfectly filled by AI. It would be fantastic if AI can deliver a whole new branch of maths. AI got into trouble inventing new languages but mathematics is really just a way of describing logical reasoning about numbers or patterns in formal language that is self-consistent and reproducible. It is ideal for describing scientific theories, engineering and logical reasoning.

Checking Google today, there are a few articles out there describing simple emotional interactions using superficial equations, but nothing with the level of sophistication needed.

https://www.inc.com/jeff-haden/your-feelings-surprisingly-theyre-based-on-math.html

an example from this:

Disappointment = Expectations – Reality

is certainly an equation, but it is too superficial and incomplete. It takes no account of how you feel otherwise – whether you are jealous or angry or in love or a thousand other things. So there is some discussion on using maths to describe emotions, but I’d say it is extremely superficial and embryonic and perfect for deeper study.

Emotions often behave like fields. We use field-like descriptions in everyday expressions – envy is a green fog, anger is a red mist or we see a beloved through rose-tinted spectacles. These are classic fields, and maths could easily describe them in this way and use them in equations that describe behaviors affected by those emotions. I’ve often used the concept of magentic fields in some of my machine consciousness work. (If I am using an optical processing gel, then shining a colored beam of light into a particular ‘brain’ region could bias the neurons in that region in a particular direction in the same way an emotion does in the human brain. ‘Magentic’ is just a playful pun given the processing mechanism is light (e.g. magenta, rather than electrics that would be better affected by magnetic fields.

Some emotions interact and some don’t, so that gives us nice orthogonal dimensions to play in. You can be calm or excited pretty much independently of being jealous. Others very much interact. It is hard to be happy while angry. Maths allows interacting fields to be described using shared dimensions, while having others that don’t interact on other dimensions. This is where it starts to get more interesting and more suited to AI than people. Given large databases of emotionally affected interactions, an AI could derive hypotheses that appear to describe these interactions between emotions, picking out where they seem to interact and where they seem to be independent.

Not being emotionally involved itself, it is better suited to draw such conclusions. A human researcher however might find it hard to draw neat boundaries around emotions and describe them so clearly. It may be obvious that being both calm and angry doesn’t easily fit with human experience, but what about being terrified and happy? Terrified sounds very negative at first glance, so first impressions aren’t favorable for twinning them, but when you think about it, that pretty much describes the entire roller-coaster or extreme sports markets. Many other emotions interact somewhat, and deriving the equations would be extremely hard for humans, but I’m guessing, relatively easy for AI.

These kinds of equations fall very easily into tensor field theory, with types and degrees of interactions of fields along alternative dimensions readily describable.

Some interactions act like transforms. Fear might transform the ways that jealousy is expressed. Love alters the expression of happiness or sadness.

Some things seem to add or subtract, others multiply, others act more like exponential or partial derivatives or integrations, other interact periodically or instantly or over time. Maths seems to hold innumerable tools to describe emotions, but first-person involvement and experience make it extremely difficult for humans to derive such equations. The example equation above is easy to understand, but there are so many emotions available, and so many different circumstances, that this entire problem looks like it was designed to challenge a big data mining plant. Maybe a big company involved in AI, big data, advertising and that knows about tensor field theory would be a perfect research candidate. Google, Amazon, Facebook, Samsung….. Has all the potential for a race.

AI, meet emotions. You speak different languages, so you’ll need to work hard to get to know one another. Here are some books on field theory. Now get on with it, I expect a thesis on emotional field theory by end of term.