|
|
|
 |
AI is neither Artificial nor Intelligent.
The Wikipedia link (on the left) is balanced in that it insists AI is a kind of "machine intelligence"
as opposed to a human or "animal intelligence". The Stephen Hawking link has a more popular view that
AI robots will take over the world and kill us all...and that must be true it's in all the movies.
Sorry. If you've ever tried to build an AI you'll find how frustratingly limited they are. They give the
illusion of intelligence but that is because humans are so good at filling in the gaps.
I'll start with the story of Clever Hans the horse
that could do math. Hans was a sensation and travelled from town to town showing off his math skills. In 1907
and investigation showed his trainer was doing the math and Hans was just tapping out the answers with
his hoof.
Okay just one more exposed magic trick. But "The Clever Hans Effect" has become the topic of serious research in
both psychology and computing science.
When you start poking around in the guts of something like Mycroft you being to understand two things:
AI is just Clever Hans and more important you are the trainer. If you do it right the AI appears to
have human intelligence. There is a profit in having everyone believe the AI you built is truly intelligent.
To understand the limitations of an AI you need to understand AI should stand for
Arithmetic Interpolation
The Arithmetic Part
The basis for the current best AI math for AI is called a
Convolution Neural Network (CNN). There are
two parts to this method: the "convolution" and the "neural network".
A Convolution is a very general term that just means multiplying two functions together to create a third. The two
functions can be anything and if you take any two functions the result will probably of no use (garbage).
If your first function is something you recognize like a picture or a audio recording you can chose the
second function to be a filter then the third function (output) is a "better" version. For example you
could make a picture of a sunset have more reds and oranges. You could remove the crackles and pops from
a song on an old vinyl record.
If you are interested in the math of the process there are lots of
simple tutorials .
If you know what matrix multiplication is skip this paragraph: In the digital world a convolution is just the term used
for multiplying two matrices together. You multiply each
the numbers in the cells on the top row of the first matrix against each of numbers in the first column of the second
matrix and add them all together. That "sum of products" fills the top left corner cell of the third matrix. You then do
each row of the first matrix against each column of the second matrix until you fill all the cells in the third matrix.
Matrix Multiplication
When training an AI the first function is of course the input training set. The second function is a set of filters. Getting the
right size and values for the each of the cells in these filters can make a huge difference in how well the training works.
Better filters mean better identification of the elements and better separation so one element will not be confused with another.
This would yield a better translation of inputs to classifications during the execution of the AI.
A Neural Network is way taking many inputs and routing them to many
outputs. In simple terms a NN takes inputs and sorts them into categories. The work is done in a tree structure where weights
at each node direct the flow of data from the input nodes thought the hidden layers to the output nodes. The NN just
multiplies the incoming data value by the weight in that node and passes it on to one of several outputs based on the result.
Again this is elementary school math were you just multiply two numbers and make a "is it bigger or is it smaller" decision.
So all the math is CNN is just elementary school arithmetic. How does it get smart? The CNN is a massive arithmetic problem.
During training the input is a large sample of all possible inputs the AI will every see. For an STT that could be hours
of recorded voice covering thousands of possible sentences. This is passed through a convolution process that creates a
pool of identifiable chunks. These are used to set the weighting values of each of the nodes in a many
layered neural network.
But smart? It's smart like a dictionary is smart or an instruction manual is smart. It has almost all the answers to all
the questions you can ask it. For example let's say I wrote a dictionary where the entry for "stupid" was "See idiot".
When you looked up "idiot" the entry was "What you again?" So the dictionary told a joke. You could write an AI that,
like that dictionary, could tell jokes. Just train it on all the possible jokes and it would make a joke out of everything
you said to it. The AI is not different to the joke telling dictionary. But it is massively more complex.
To repeat: an AI is a very complex classification engine but the math is simple arithmetic. During training it has to do trillions of simple
calculations for many hours to calculate the parameter values. It takes billions of calculations to run that classifier
and turn a spoken "what's the weather tomorrow?" into spoken weather forecast.
But it is not magic .... it's just arithmetic.
The Interpolation Part
When an AI is trained it is given a very large number of samples. To train a voice you need five hours of audio of a speaker
saying several thousand sentences that make up a vocabulary of 20 to 30 thousand words. That training set is the bounds
of its database. If it must speak a word that is not in the training set there are rules in the software to approximate.
In practice it often says the new word the way the training voice would say it.
Filling in the holes missing from the training set is called interpolation. This is finding a new value approximately
between two known values. If an AI were required to use the word "wonderful" as an adverb in a sentence it might invent the
word "wonderfulness" because it has basic English rules built into its exception handling.
If there was something the was outside the bounds of the training, for example it had to pronounce Worcester
(the city in Massachusetts) it would pronounce it as three syllables. Without a specific entry in its training set it
could never guess the correct two syllable pronunciation.
In an application where the plan is to put general purpose AI in control of machine in a factory any holes
in its training could be disastrous. If a problem arose that was outside its training set it would
try an interpolated solution. For example if a pipe
broke and there was no flow at the next station it might increase the input pressure to compensate.
The obvious solution is to increase the size of the training set to cover all possible situations.
But that assumes the humans
building the training set know all possible situations (see 737 Max).
The AI is trapped within the bounds of its training set. It it very good at interpolating between the elements within those
bounds but it is non-functional outside those bounds.
Where Does an Arithmetic Interpolator (AI) Succeed?
An AI is a marvelous librarian. It can take a question, sort through the vast store of human knowledge on the internet,
and provide a concise answer (usually). More important the human asking the question needs no computer skills to
make it work and the communications tool is not some elaborate electronic protocol. It it is just the human voice.
Once a child can speak (about 18 months) they have access to all human knowledge. Now that is as close to real magic
as you can get.
As you can tell I am a huge fan of AI. I love it for its strengths. I just want everyone to understand it is not god,
it is not even smart. It is just a tool among many tools on the pegboard each with their own applications and limitations.
Humans are tool using critters and AI is just a tool and no more.
An AI is also an marvelous assistant. Some chess history. The best human chess players are called grandmasters. In 1997
Deep Blue beat Garry Kasparov who was considered one of the best grandmasters in existence.
Chess playing programs were following Moore's Law
and it was clear within a few years you could have a chess application on your laptop that could beat any human.
Then a couple of average chess playing humans with a laptops running chess programs showed they could beat the best
chess playing AI. There are now contests called Advanced Chess
or Centaur Chess (introduced by the same defeated Garry Kasparov) where humans augmented by chess programs compete
at a level well beyond grandmaster and outperform any non-augmented AI.
Exponential Growth in AI Technologies
Like many technologies AI is following a kind of Moore's Law exponential with a very short doubling time.
When a field is in an exponential everything that you've just done looks trivial and everything in front of you looks impossible. In
Jobs Jobs Jobs I discuss the costs of using very new applications like chatGPT for dlexa is in the impossible realm.
AI technology is moving so fast that within the 5 year project plan it is entirely possible to make it a component.
Christopher Potts from Stanford Online (January 31, 2023) in GPT-3 and Beyond discusses the rapid
advances being made. He repeatedly says that things he thought were impossible are being solved surprisingly quickly. He shows the exponential growth in
the size of the models and their remarkable ability to provide very satisfying (human like) responses to complex requests.
Better Liars
Unfortunately it is becoming clear that the more advanced chat AIs turn out to be better liars. It's not intentional because AIs cannot have motives.
They just scour the internet for phrases that look like assertions and offer those in well written text.
A recent 60 Minutes episode challenged ChatGPT with
a topic it was not allowed to discuss. The AI was designed to change the topic with a "surprising" fact. ChatGPT asserted that 3% of the Antarctic ice sheet
was penguin pee. Certainly a surprisingly cleaver way to change the topic. The truth however is that penguins don't pee: they excrete their salty waste as a paste.
I'm not sure a $billion computer application that gets caught lying on a global TV program can be called "mature" or "ready for serious employment."
Christof Koch of the Allen Institute for Neuroscience in Seattle said:
You can simulate weather in a computer, but it will never be ‘wet.'
I have tried to find the original quote for this without success but there are several versions posted in 2014. It may also be that Koch said it best but
the idea has been around for as long as there have been weather forecasting systems (1950s) ... a mixture of both surprising skill and frustration with their predictions.
The Last Mile
Christopher Potts suggests there is a lot of work to be done on the "last mile problem" That is getting these AI systems to do something practical in the real world.
His interest is in search (like Google) and in context specific tasks where the AI is an intelligent assistant to a skilled worker. He explains the best AIs
record human feedback while they are being used and include that input in training sets.
He also suggests the best way to get progress in AIs is to put them in as many real world challenges (the last mile) as possible.
So building a lexa is exactly one of those "last mile" projects that (may) find things that lead to more advanced AIs. If nothing else
such a project allows you to keep up and appreciate the advances that are being made. In a exponentially growing technology just
keeping up is hard work but will allow you to see both business and new learning opportunities.
The Dlexa Project
I think AIs could become good story tellers. And these stories could augment an inanimate object so it has a presence and personality.
I think the world would be a more magical place if we could have conversations with boulders and trees and hedges.
The Dlexa project is an attempt to employ an AI to make that magic happen in my front yard.
Next:
Check out the Builds tab to see the Dlexa project progress.
|
Wikipedia Definition