AI

I know a bunch of you are just as nerdy as I am about AI and tech, so I think you will appreciate this.

https://arxiv.org/pdf/1803.03453v1.pdf

It's a research paper showing the mistakes that have arisen from poor AI goal programming. The fun stuff starts around page 6.

Comments

[cvvvrt]

If you're interested in AGI I can recommend Robert Miles' YouTube channel. He makes amazing videos.

[Damaged]

4825038 There's been a great series from MIT running just this year: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4

Lots of stuff from all angles. Some lecturers are giving a ground-up view, others are giving a top-down, while one was a psychologist giving insight into what a human brain needs to establish allostasis. Really great stuff.

[Booster Spice]

What suprised me was one YouTube link that claimed to demonstrate a dep learning method, and all it was was a Genetic Evolution algorithm to find a solution neural net using a phase space search, with nothing that could be considered diirected learning at all. the second problem was that there was three parts to the learning proess, but only the neural net was evolving, neither the net builder or the net tester was being evolved as well. Fixed rote teaching in an enviroment they dont really know whats occuring? So very Season 8 Premier episode.

There is an alternate creation method Ive seen mentioned, Baysean builds, which I was hoping to use as an experimental method because its relative simple to implement. Then again, aneural net of unlimited complexity is, as far as I understand it, simple to implement, because once you have a synapse model and a neuron model, you just iterate them over the net values database.

Which leads to the reasons for why AI has taken so long to get going. Not just processing for the turnover point, but storage access times. If you only have spinning hard drives, then the pure random data access needed for a generalised neural net limits you to less than 100 look ups a second, which are then processed to give a single result. Its barely fast enough to do predictive texting. A modern SSD can reach 50 Thousand accesses a second, and even a $5 Pi has more direct access RAM than even the Hard drive store of a mid 90s machine when neuralnet.library first came out.

The best description that I can come up with for the announced standard method of training a Deep AI with hundreds of thousands of images, is lets say I give you a book. This book has a hundred thousand hi res images. There is one thing in common with these images.

So far, so good.

Each image is made up or a maxiimum colur, shade etc pixel variation SIRD, Magic Eye.

You can only view it with one eye.

You dont know what the object is you are looking for, you have never seen it before.

And computers are so powerful, they can do this.

just imagine how much faster that shere power can learn, if it first goes to computer kindergarten. A is for Apple. This simplified stylised symbol on a plain background is a representation.

If we can do this on a Pi, Combat Pie just imagine the code that can run on Nvidias new 30 Teraflop SoC, or someones attempt to make a Petaflop PCI express card by putting 32 of those chips with water cooling on one. Meaning a Exascale supercomputer can be as little as 64 racks, 8 cabinets or a Megawatt electrical power, within a couple years.

Combine that with small world, hypercube connectivity so that theres a trade off by more than halving power consumption through having far more parralel, lower speed links, and you can be considering a supercomputer which doesnt even have to stop, never mind shut down, for repair and upgrade.

the thing that really strikes me, as research is done into AI and data processing, is just how little is actually needed when its discovered how it works. Sort of like the difference between a complex trestle railway bridge, and a self deploying Baily Beam Bridge?

Ive seen a GPS and microcontroller setup that can run for 2 years off a single AA alkaline battery for a wall clock, but I havent yet seen anyone announce a vibration harvesting unit and clock setup thats capable of running perpetually. Maybe it would make BOLO too dangerous if they never need to go to sleep because they can run at low level off water ZPR etc?

The biggest problem I think Ive seen mentioned with Beysean designs, is that they appear to be able, in a restricted energetics enviroment with dynamic load ajustment, to be capable of suffering what seems to be the analogs of various mental disorders.