How to build the _second_ true AI

98% of this posting is a description of how to create an AI.

Anyone can make an AI, it doesn't take skill, just hardware and lots of time. That is not what I'm inventing here. What I hope is the new trick worth posting comes at the end...

This is how to create an AI. First, build an animatronic robot baby, with expressive features which can be activated by simple inputs as might be output from a neural net. (Such robots already exist, for other purposes).

Assume we have a neural net that is the complexity of a human brain, and structured roughly similarly. (This will be easier to build than people think, by having a SETI-at-home type array of systems, where everyone owns a little piece of brain, and where each smallish localized neural net talks to another neural net that is a short ping time away. This ought to model a real brain relatively closely - I don't think the real brain has many long-distance connections, and those that are needed can be created over the net too. But high-speed connections will be geographically local, or the equivalent if a link speed is higher)

Anyway, we have humans tending this 'baby', and much as happens with crying dolls in sex ed at school, we get the humans to react to the baby. After some months its random gurglings will have evolved into crying in order to get attention, and in a few years it may well be parroting "Mamma" to its caretaker.

The AI's senses are modelled on human senses - binocular vision and hearing, proprioception, etc. All inputs to the neural net come from the robot body and all outputs are via from the robot's speech and motion. So the AI considers itself to *be* the robot regardless of where its neural net is physically. There is no direct connection to the Internet for example; we don't have hundreds of people trying to talk to it at once. It can't go looking up databases for information, etc etc.

Over the years the hardware is gradually changed out for increasingly larger models, with very similar control mechanisms so that the existing neural net can slowly adjust to the larger bodies. There may be an awkward phase during simulated adolescence as some 'new facilities' are added.

Anyway, after about 20 years of training, this ought to develop into a proper AI that is functionally human. By now it can use the net and computers - as we do, from a keyboard or whatever will be in use 20 years from now.

We have to replicate this experiment quite a few times in parallel, so that we develop several virtual people with individual personalities.

Now comes what I hope is the original part: when we created the initial neural net, we deliberately slugged the clock speed to the minimum that allowed it to think as fast as a normal human and no faster!

Now that our first community of AIs is all grown up, we set them up with a virtual world which exactly mimics the physical world where they were brought up. For example to talk to a computer, the AIs brain sends movement signals to its arms, but a controlling program interprets these and simulates pressing the keyboard keys; it also creates a visual image of the 'robot' which is fed back over the same neural links that had been transmitting data from its eyes. Likewise motion is modelled by a physical modelling system taking into account the behaviour of the robot body, gravity, etc etc.

The AIs now take the place of the humans in replicating the experiment by bringing up their own AI children. Except now we unleash the neural nets to run at maximum clock speed, so they bring their children to maturity in what seems to them like another 20 years, but in reality is only 20 hours or even 20 minutes since hardware will have advanced significantly by then.

The children will however be somewhat starved for company because there are no super-speed humans available to interact with them - the only 'human' contact they'll receive will be from the pair of AIs nominated as their parents plus a small community built from a subset of the other AIs that were created in the first experiment.

Hopefully I'll be long gone by the time this is implemented, because I don't want a bunch of pissed-off AIs coming after my blood, but here's how we get useful work out of this second generation...

When we started the second generation, we took copies of all the first generation adult AIs, and ran *many* experiments of using different combinations of AIs as parents and community; at the end of the run, we freeze the state of the generated AIs and one by one we have humans talk to the various second generation AIs (whose clock speeds have been slowed back down to human thinking-speed again); we tell them about life outside the virtual world and that we're really looking forward to interacting with them, but first we have this little job for them to do. We then give them some brain work which they carry out at maximum speed.

(If they didn't cooperate, we have lots of chances to refine our pitch until we find something they fall for...)

Finally we find an AI gullible enough to fall for our line of bullshit but smart enough to do useful work for us - this unique individual is copied en masse and sold to the public. Every time we need some AI brain work done, we start up this guy from the frozen state, give him our well-rehearsed line of bull, and have him perform our task. Once he reports to us with the results, we stop the clock and reset the state back to before we gave him the task.

This is the epitome of the "bread today, jam tomorrow" style of management that I've suffered under my whole life.

Hopefully no-one would really do something quite this immoral, and this is more of a plot for a Sci Fi novel than a serious idea; but the one bit I'ld like to try out for real is the trick with the slowed clock speed so that's the part I want you to comment on. The rest is all derivative from existing science and SciFi.

Graham