There's a Dark Secret at the Heart of Artificial Intelligence: No One Really Understands How it Works

..."AI works by combining large amounts of data with fast, iterative processing and intelligent algorithms, allowing the software to learn automatically from patterns or features in the data. ... Cognitive computing is a subfield of AI that strives for a natural, human-like interaction with machines."...

..."Artificial intelligence (AI) is the simulation of human intelligence processes by machines, especially computer systems. These processes include learning (the acquisition of information and rules for using the information), reasoning (using rules to reach approximate or definite conclusions) and self-correction."...

https://futurism.com/1-evergreen-making-sense-of-terms-deep-learning-machine-learning-and-ai

https://www.google.com/search?q=how+does+artificial+intelligence+work&oq=how+does+artificial+intelligence+work&aqs=chrome..69i57j0l5.14378j1j7&sourceid=chrome&ie=UTF-8

This is certainly thought provoking and a little hint at the possible future of AI.

"I bring the news from"D" street,and the news is not good."

https://www.independent.co.uk/life-style/gadgets-and-tech/news/facebook-artificial-intelligence-ai-chatbot-new-language-research-openai-google-a7869706.html

GA.

This article, like most of what is published on the internet is misleading.

I will add this clarification for those who do not train a Artificial Neural Networks (ANN) in their daily activities.

We know exactly how ANNs work but once trained, it is extremely difficult to read their knowledge. This means that when most types of ANN learn something, we can copy the knowledge into another (usually identical) ANN but we cannot extract this knowledge and publish it in a book like we do when human learn something. The knowledge cannot (yet) be synthesized in an understandable form.

For the moment, we do not loose much as this knowledge is sub-human level (actually, sub-insect). It is basically simple tasks like recognizing a part or perform a diagnostic of the state of a machine. Nothing threatening. That statement could change in 20-30 years if we keep on progressing.

The power of ANNs is that it can treat large amount of data with a logic and rules that are more flexible than classic computer logic. What we train in an ANN is usually very difficult to code in a computer program. It is similar to Fuzzy logic taken to the extreme.

For the moment, it is another tool to solve some problems on a computer. That is all.

There is some human knowledge that can be written down and put into books, and there is some knowledge that cannot. I doubt that one could learn to ride a bicycle by reading a book, and most bicyclists could not explain how it's done.

"...and, thus, SKYNET was born." (wink,wink)

In the immortal words of Alfred E. Neuman . . . .

(much ado about nothing, let the panic begin)

In 1997 the computer, Deep Blue, beat world chess champion Garry Kasparov. Interestingly, in the first game of the match, Deep Blue made an inconsequential move which was the result of a programming bug. Kasparov could not understand the purpose of the move and wrongly concluded that it emerged from a deeper strategic or positional insight than he could fathom. Kasparov nevertheless won the first game, but it is speculated that this shaking of his confidence (caused by a computer glitch, not by computer processing power), contributed to his loss of the match. There is a lesson here: As AI develops and we all become enthralled by its power, we could base our actions on computer glitches that we misinterpret as deep analysis.

Skynet is pleased

Sounds like the understanding issue is the sheer mass of data and size of the total combined algorithm. I suspect an individual could run the data an the problem with a slide rule and ledger pad given enough lifetimes, forests worth of ledger pads and enough graphite mines to make pencils.

Similar to the original Monte Carlo analysis developed for nuclear reactor core design. Individual interactions can be well understood, but the cumulative total of activity is daunting.

I agree.Sort of like predicting the weather:billions of variables to contend with,each with it's eventual effect on the outcome,and the output of one calculation is simply the input for another result,ad infinitum.

However,it would behoove us to know exactly what OUR machines are doing so that the "Black Box" does not turn into a "Pandora's Box".

Perhaps quantum computing will help with this, that is until they have quantum computing performing AI.

Face book recently shut down a couple of their computers because they had developed a language of their own,something they were not programmed to do.

Something perhaps tangent to this discussion...I notice that in the last couple of decades of new engineers that many are not able to recognize a nonsensical answer from a simulation software program.

I started in the era of slide rules and ledger pads and a back of the envelope calculation to set orders of magnitude on an engineering problem was expected. I went through the transitions to calculators, cad and then simulation, but always did that back of the envelope to do a "sanity check" on simulation results to avoid taking a trip down the rabbit hole.

With AI, the ability to do that sanity check is at least somewhat impaired and if the output of one AI effort is accepted as input for another AI process, weirdness is highly probable. I liked the generated image of the dog sandwich in the AI descriptive article...troubling, but in a cute way.

There is a difference between "understanding how it works" and "understanding how as specific result was achieved".
It's not that hard to write software that combines several inputs and rules which then gives a surprising behaviour.
Del

The embryonic form of Skynet is here now:

https://www.allaboutcircuits.com/news/intel-introduces-pohoiki-beach-its-64-chip-neuromorphic-system/

15-legitimate-fears-about-artificial-intelligence

Legitimate is subjective. While there is at least some basis for these 15, I believe most of these are over-stated. My biggest concern with AI is not with AI at all, it is with the humans that interact with it. As mentioned previously by Jpfalt, newer generations of engineers are more likely to 'believe' a bogus simulation result. The popular thought that AI will be the panacea for <fill in the blank here> is the most dangerous one.

Just look at self-driving cars. Four and more years ago, more than a few engineers were calling out the Emperor's clothes, but their concerns were drowned out by the cheerleaders' accusations of Luddites, and so on. Now that autonomous vehicles have had their reality check, so too will be AI's turn soon.

Recall that the root purpose of AI is to make machines as unreliable as people.

That's okay; hardly anyone knows how natural intelligence works, either.

Yeah, especially if you're talking about women.

IMHO:

We don't really totally understand anything.The deeper we dig,the more unknowns we uncover;the more mysterious our very existence becomes.

IMHO,it is folly to trust a machine not to inherit the faults of it's creator,of which we do not know or will not acknowledge.

If AI is eventually able to program itself,it will still incorporate the basic faults of it's genesis.

Computers can calculate billions of times faster than humans,and can also make billions of mistakes faster,and our own errors may eventually come back to doom us.

We already rely on computers to control our electrical grid,and they suffer from internal errors as well as outside nefarious players.No one person knows how a nuclear generating plant works,not all of the minute details.

Our vehicles can be hacked and eventually everything in our lives will have a IP address,subjecting it to monitoring and control.

They will probably eventually embed code into the DNA of crop seed to control every grain harvested and track it though the entire system,all the way to the waste treatment plant,into the rivers,and into the ocean.Each contributor to the waste will be automatically be billed for the proportion of waste generated.

DNA is the most durable of all memory storage systems known,surviving for hundreds of million of years.

The origin of a single grain of wheat could be traced back millions of years.

At some point,AI may start tinkering with our DNA and hybrid computer/humans will be created,sort of like the Borg.

But,that is just me,letting my imagination run wild.

Here's another type of hybrid:

youtube.com/watch

I seem to recall a science fiction book probably by Arthur C. Clark where there was a war (with AI involved?). One side was learning how to do multiplication with a pencil in order to get an advantage over the other side. Anyone remember the name and details of the book?

It's the illusion of intelligence...

After some 800,000 cycles, or more, through some algorithm, some resultant version appears to be (statistically?) ''superior'' for the moment, and maybe it is, but maybe, in the longer run, it is not...

...and that is only assuming that the algorithm was well-written in the first place...

So we will understand the day your toaster wakes up believing it's Optimus Prime and zaps it's nemesis, the latte machine.

... while simultaneouly hoping that the latte machine is not any sort of Deceptacon.(?)...

Somewhere Alan Turing is smiling.

Rise of the machines and the matrix was born...

With all the innovations and discoveries in the information technology world, we are seeing something very exciting and that is machine learning. You might have heard the world artificial intelligence by now. To clearly describe this to you let me explain it in simple words.

Artificial intelligence is a branch of computer science that helps machines to learn with the initial data given to them. With the inbuilt learning capabilities of the machine, we try to create machine intelligence which can minimize human effort.