"Make shit up for a weekly column on an IT news website"
I bet the editor choked on his tea when he read that. I simply laughed out loud...wonderful.
Never again. As Gods are my witnesses, you will never catch me [insert gerund here] in future. I have learnt my lesson. You won’t catch me because I’ll be more careful next time. Learning from one’s mistakes is a sign of intelligence, they say. This is why machine-learning is at the heart of artificial intelligence. Give a …
Reminds me of an old tale (possibly an urban myth) I was told some years ago.
The military had the idea of creating and training an AI to spot potential weapons installations, possible hidden bases etc. from snapshots of satellite footage. Programmers created the AI, and they proceeded to train it by showing it numerous satellite pictures of weapon installations, hidden bases and the likes, then showed it a bunch of nice scenery shots where there was nothing suspect.
All seemed to go well!
Then they tried it on some real data, and the AI promptly started flagging *everything* as a weapons base. After some head-scratching, someone figured it out. The photos of weapon installations they'd used for training were all gloomy, murky shots. The normal scenery shots they'd used were taken on a brighter sunnier day.
In essence, all they'd done was train the AI to recognise a sunny day...
It may be true that AI learns from every single case it sees, the problem is that you don't know what it learned, can't ask it and can't correct it. No doubt it will some times get the right answer, but that just makes it Accidental Intelligence.
And that's put off my looming work for all of 30 seconds: must be time to make a cup of tea.
That's marketing speak. No AI is capable of learning at all. It's pattern recognition aided by human curated data and human initial labelling.
But indeed, you don't really know if it's matching (data comparisons) for what you intended or some other feature in the images. We see patterns, familiar objects in clouds, toast, flames in the fire, scattered stones. Because once we understand chair, we can decide to use a crate as a chair. A child that has eaten bread and sausages will assume a sausage-in-a-bun or a hot dog is edible. A two year old can do things easily that are impossible for AI.
It's called the AI paradox and it was known nearly 60 years ago.
Expert systems were the big thing in AI in the 1980s because they used text. The problem was capturing the expert. Faster cpus, bigger databases and more RAM simply made actually simpler so called image recognition possible. There is no recognition. Just matching.
It's all marketing. None use "machine learning" or "neural networks" as those don't mean what they mean outside of AI marketing.
Even machine translation has gone backwards. It now uses a brute force approach like a giant Rosetta stone and matching phrases and words.
Text to speech isn't much better than nearly 40 years ago and so called smart agents are just voice to text front ends using pattern matching to search engines and chat bots hardly better than Eliza or ALICE. Speech recognition has moved from being a program on your car radio, phone or PC to something creepy running on a 3rd party system, the so called cloud. That's a backward step in privacy and needs the Internet.
@Mage - "A child that has eaten bread and sausages will assume a sausage-in-a-bun or a hot dog is edible. A two year old can do things easily that are impossible for AI."
A good thing too. It'll be bad enough when the machines take over, without them eating all the damn hotdogs!
Proof, or actual evidence, or substantive argument, or displaying any knowledge of the subject area are discouraged in Reg forum debates about machine learning or related areas.
Hell, just using the term "AI" makes anything someone says on the subject suspect. (Those O'Neill quotes in the article were painful.)
"Actually medical "AI" systems can be quite good - better than doctors."
For very, very, specific things that they have been told to look for. For example, a dude falls off his horse/bike/roof. Complains of sore ribs. Doctor orders X-rays. AI comes back "no busted ribs in that X-Ray, tell him to take two aspirin and call us in the morning" ... all while missing a little, tiny dark spot on the left lung that would stand out like a searchlight to any halfway competent radiologist.
But wait, it gets worse ... with today's digital X-rays, it's just numbers being fed into the AI. So said radiologist will probably never see them rendered as an image, and thus won't even be able to spot it accidentally. At least it's all lovely and modern, though, right?
Many years ago, somebody I knew was working on a military project to get a tank launched missile to recognise and destroy other tanks.
They were testing on Boscombe down or somewhere and launched their missile from their tank.
Off it went, turned around, came back and hit them. It had decided that their tank was the best threat.
Good job it was a wooden payload.
It reminds me of the also possibly apocryphal story of the AI trained on chest X-rays to spot cases where a chest drain would be beneficial. It was given a nice set of X-rays to train on, which had been vetted by the thoracic consultant, half of which were cases which didn't need a chest drain, and half were cases which did. It worked really well, and picked up all the ones needing a chest drain. When tried on real data, it didn't perform so well at all. It turns out that those training images from patients who needed a chest drain had already had one fitted (for ethical reasons!), and the "AI" was matching the chest drain in the X-Ray and not the symptoms.
I'd say it's a little from column A and a little from column B. The problem with trained "AI" is you can't know what it has actually been trained to do, you can only look at the results, without knowing how they are arrived at. For example, if you trained an AI to spot red BMWs amongst a sea of blue Fords, is it spotting blue cars, BMWs, a combination of both, or personalised number plates? Unless you have suitably controlled test cases to measure the output, you can't know.
At least with a real person, you can ask them how they arrived at an answer. If they give a reply along the lines of "it just felt right the right answer", they've just failed their Voight Kampff test.
This is not actually impossible to do, for example you can feed in reams of perturbed example data and see which perturbations affect the outcome. More efficient methods might be possible for specific learning frameworks.
@Loyal Commentator - 'If they give a reply along the lines of "it just felt right the right answer", they've just failed their Voight Kampff test.'
Not really, often people find it difficult to explain why something is "not quite right", the Uncanny Valley can provide examples. In the opposite sense, conmen try to manipulate people into trusting them with the right triggers, "but he seemed so nice".
There is the expression of the "ring of truth". I find that when I'm trying to remember something out of many possibilities then the right match really does have "ding" feeling. It also happens when I'm trying to solve an IT problem - before the then exacting application of logic to prove it.
it just felt right the right answer
That's an interesting point.
Sometimes - with humans - a 'gut feeling' might still be based on input data. It's just that the person in question can't identify and isolate the particular data that ultimately gives rise to the subsequent decision.
When I worked in the rat race many years ago, I could always seem to tell if something was up (involving other people, not machines), but could never identify why. At the time, I assumed I was just being paranoid or something, but as time went by, my 'gut feelings' tended to be strangely accurate.
In the job that I do now, picking up 'the vibes' of the people I deal with is quite important. I can tell if someone's mood is different to the last time I met them within a few seconds, even if they say they're fine. Then, later, it'll come out they've had a negative experience somewhere and it's that I'm picking up. It's impossible to say precisely what it is I'm detecting, but I can just tell it's somehow different. It's my 'gut feeling'.
I'm sure everyone else can detect it, too. They just don't realise or put it to use.
It's most likely what can be attributed to subconscious training. That's part of that makes measuring intelligence so difficult; we don't even understand how we arrive at everything, as a good chunk of our brain functions are autonomic and outside our conscious perception.
A good example of this is where the words come from when we speak. We don't actually select which words we are going to use. We sort of find out as we say them.
Even more freakily; Brain research shows that the signal to move a muscle occurs before we decide to move the muscle. (Can't find/be bothered to look for the reference).
I'm not all that clued in to others and I've had co-workers where the words "completely and utterly oblivious" wouldn't even begin to describe it. But then again I didn't choose to become a mechanical engineer because I like working with people. I prefer things I can hit with a hammer if they don't work. While that method works in humans too, its usually frowned upon.
"Other people" are annoying and confusing. Best avoid them imho.
When tried on real data, it didn't perform so well at all. It turns out that those training images from patients who needed a chest drain had already had one fitted (for ethical reasons!), and the "AI" was matching the chest drain in the X-Ray and not the symptoms.
And this is the other big lie that AI people/companies tell - the more you train it, the better it becomes, or like in TFA:
"A system with two years’ learning under its belt will do the job far better than a new one that has to learn everything from scratch. It's the same reason that someone who's been with your institution for two years works more efficiently than a newly hired employee. Experience matters."
In scenarios like this, where you eventually find that you've trained the wrong thing, you go back ot the start and redesign something, and start from scratch again. AI is like having a job role that you keep firing people from, "the model" is constantly being adjusted and retrained.
We have an AI team; they currently training a model to crop images to the right size for the website. 5 developers, 4-5 years of time so far, and the thing still isn't a) in production, or b) any better than the hand-coded auto-cropper which gets adjusted by a mechanical turk.
The aim of course is to do away with the people who currently do this as 10% of their working time, or "free them up to do other things" in manglementspeak - there are 10 of them, probably earning 30-40% of what each developer does (and probably just 10% of what AWS earns from the project...) Long time to get the payoff from this.
You know this sort of error is older than computers -
Another serious training mistake was revealed later; the Soviets used their own diesel engine tanks to train the dogs rather than German tanks which had gasoline engines. As the dogs relied on their acute sense of smell, the dogs sought out familiar Soviet tanks instead of strange-smelling German tanks.
The hotdog problem has been solved (you can still download it for iOS).
I am reminded of the wonderful game Deus Ex.
The Illuminati created an AI called Daedalus to advise them, and eventually to analyse the data captured by the Echelon system and identify terrorist groups and threats to the Illuminati. Unfortunately, due to a "pattern matching" error, the Illuminati themselves were classified as a terrorist organisation as well.
Unfortunately, due to a "pattern matching" error, the Illuminati themselves were classified as a terrorist organisation as well. .... Franco
Who/What decided on that particular "pattern matching" error being a false positive and not a fake negative? I'm calling out MRDA on that if the answer is the Illuminati themselves.
... intentionally using the Butterfly Effect in AI training data as a means of surreptitiously manipulating the end result. He's doing it for fun, as a distraction from his Doctoral studies.
Or rather it WAS fun ... until I asked him what would happen if "the bad guys" did it.
"Volunteer my services, only to find everyone else is getting paid for theirs." Never again, not on your life bub! The first thing out of any self respecting contractor's mouth would be "What's in it for me?" Just once I tried to be nice, turned into the longest week of my life!
If life gives you lemons, peel them, taking care only to remove the outer skin with no pith. Seal the peel in a container with about 100ml of pure alcohol (95% will do if 100% not available) per lemon for a few weeks. Strain the liquid and mix it about 1:2 with simple syrup. Ideally lay this up for another week or two. The attractive yellow liquid you have prepared is limoncello.
After peeling the lemons can be juiced to use as an ingredient in a cocktail such as a whisky sour, gin fizz or French 75. For the French 75 you will need gin, champagne and some additional sugar syrup, you may also wish to put aside a little of the zest.
(Yes, ideally you would use Amalfi or Sicilian lemons for this, but if life is giving away lemons I prefer not to look them in the mouth. Also assuming that any lemons life has in stock will be organic.)
Nice recipe for the 75 @ibm.
The juice should be put in the jug that comes with your electric whisk.
Then add a teaspoon of salt flakes, a teaspoon of Dijonais and a teaspoon of organic white wine vinegar that you have pre-prepared by soaking with some home grown tarragon in its bottle.
Take an organic egg and put it one side, to get to room temperature, while you swirl the mix in the jug, by hand.
Add about 10 rotations of a pepper mill, preferably using white peppercorns as they are more aesthetic than black in this case.
Crack the egg into the jug and mix it with the electric whisk.
Slowly pour in about 250ml of organic vegetable oil and move the whisk around whilst you do so.
After a couple of minutes you should have a near perfect mayonnaise.
Until the beancounter finds out what the experience really contributed and how much it costs at freelance rates.
Most never do as that is another budget and even if they do, it is always way too late, the competition pays better for that experience (and knowledge of the inner workings of their competition).
a) Your comment needs help, it's not coherent.
b) It's actually much harder to make a decent lighter beer like a lager than a darker, heavier stout or ale because there is no place off flavo(u)rs to hide. Simple Ales are a lot easier to make than American industrial lager clones. If you don't believe me, try it.
Now then, you've hit on a nugget of real world truth there. The curriculum for teaching young kids to do creative writing is very much focussed on having lots of adjectives. (And other "parts of speech" but mostly adjectives/adverbs).
The Powers-That-Be have decided that there is a formula for good writing that has a high bias toward using these "wow words"- especially in opening sentences.
Imagine an AI trained by such people, or indeed the current crop of youngsters when they've grown up...
I had to check back on that one myself -despite a) having done a fair bit of classroom teaching since my retirement from being a literacy specialist and b) having an 'A' level and a large chunk of my degree in Eng Lit.
Because it isn't something anyone should need to bother about with, if not for the fact that it's part of that same "Wow words" bollocks.
"Suddenly, Francine shot the politician".
"Suddenly" being an adverb. And it's fronted because it comes before the verb it describes.
But the point to it is that in the Behaviourist inspired world of curriculum design UK style it's an element of their model of "impactful" (I think the word is actually in a curriculum document somewhere- certainly in training materials) writing.
Being very much a Behaviourist model of literacy teaching, everything has to be taught from mechanical, testable components rather than the messy, intuitive, subjective, real life activity that is literacy. It's also easy and cheap to publish then sell training materials to hard pressed teachers at inflated prices. Teachers are under pressure for the kids to get ticks in boxes, so must use this stuff in the prescribed manner.
It suits politicians because it is testable and measurable - whether it adds up to decent writing is another matter. IMHO it creates an army of clone Zombies - every kid churning out the same rubbish to get the marks. And by the way, the same goes for the focus on "Phonics". Totally Behaviourist in method. Easy to teach, easy to test, easy to design programmes, easy to sell, both financially and politically- because it seems logical - even if it doesn't match how we actually read.
And, to briefly draw this back to "Artificial Intelligence", it seems to me that the approaches I've read about also seem quite Behaviourist in the underlying thinking- I may be wrong.
It is very sensible. As one ages ones memory often becomes a little, thingy, you know. Particular problems come from trying to remember a specific instance of a regular event, like parking in the supermarket. The brain just drops the info as soon as it comes in. So wandering out with 3 large carrier-bags you're standing there like a plonker for ages trying to remember where you parked THIS time. Much better to stick to the same place, and stick a flag on your car radio aerial.
I have a similar problem parking in town. Probably a dozen spots where I can sometimes find a slot, so drive from one to the next until I find a space. An hour later: where the heck did I park! I have to mentally replay my route until it clicks.
There is no Artificial Stupidity ... Stupidity is the most common thing in the infinite Universe, therefore all examples of stupidity already exist naturally.
"Only two things are infinite, the universe and human stupidity, and I'm not sure about the former." —Albert Einstein (supposedly)
"Apart from hydrogen, the most common thing in the universe is stupidity." —Harlan Ellison
"There is more stupidity than hydrogen in the universe, and it has a longer shelf life." —Frank Zappa
Reminds me of my research days. I was trying to figure out HOW to show that a particular neural net (not that that was the name used) did learn. A friend suggested that I look even more closely at some standard optimisation techniques I was using, and all seemed well.
Until I realised that I was essentially showing how fragile the whole shit was.
Repeat after me, neural nets are not generic.
And a few years ago started reading about this deep learning and had smoke out of my ears. Called a friend who is still in academia and asked him about "old wine in new bottles". He laughed and said "It is not that bad".
As a sometime neuroscientist it’s in part because our brains work more like massively parallel analogue computers than digital ones so using digital computers to make intelligence like ours is doomed to fail.
For those who like a challenge I recommend Peter Ulric Tse’s The Neural Basis of Free Will: Criterial Causation. A background in neurophysiology is recommended but there is also an argument in formal logic at the end.
Basically when a thought goes once round the brain as the signal passes through neurons there are mechanisms which alter the set point of the neurons on the fly. So when the thought comes around again you can think about it differently, make connections from it etc. The range of mechanisms which do this is quite large and it is likely we have not detected all of them yet.
It MAY be possible to make silicon emulate this but the computing power for a simple network modelling all the permutations will be enormous. The human brain is the most complex thing we know in the universe. Modern AI is just big data, it is NOT the route to machine intelligence.
Oh and if we ever make a brain, it will be a baby and will need to be taught, and corrected just like a human being. Since we can make real human beings fairly easily and in a fun way, other than for interest’s sake why would you do this? So you can treat the conscious robot like a slave?
Since we can make real human beings fairly easily and in a fun way, other than for interest’s sake why would you do this?"
The expectation is:
1) such creations will live much longer than humans.
2) their state at any time can be replicated in clones.
As human knowledge gets distributed into manageable portions - then it creates silos whose knowledge is incomplete or built on misconceptions. People learn levels of abstraction as a "truth" - without an understanding of the constraints beneath them.
The Enlightenment was three centuries ago. Looking at the news recently it appears that several generations have been raised without the benefits of such thinking.
> Since we can make real human beings fairly easily and in a fun way, other than for interest’s sake why would you do this?
Because, first of all an AI doesn't get pay, and second, it doesn't mind working 24/7.
That's about all the reasons you need. Who cares if they do substandard work, humans aren't always super-efficient either. The big difference is AI is CAPEX, while a human is OPEX.
Watch trash TV after midnight
Make shit up for a weekly column on an IT news website
Alistair "Antoine" Dabbs? SFTWS as a Eurotrash for the IT crowd? (Ok, may be that is going a bit too far).
Eurotrash was broadcast by Channel 4 on a Friday (2230?). Just like Eurotrash, SFTWS is "produced" in France.
We even had Pipi and Popo make an appearance here recently.
As somebody much smarter than me once remarked "Artificial Intelligence will never beat Natural Stupidity!" ....... Willy Ekerslike
Advanced IntelAIgents do not engage or compete against Natural Stupidity. Such is how and why ITs Progress is so rapid and rabid in those dumb sectors and terrified vectors drowning in its Base Subprime See Scapes ....... Perverse Subverted Corrupt COSMIC Guidance Systems
Does humanity Do you like to think the converse/obverse/reverse, because it gives you something a warm blanket of cold comfort, that Natural Stupidity engages and competes against Advanced IntelAIgents?
How well do you imagine any prize element or sundry vital component encountered or launched from either side, faring in that entanglement?
Be honest now, for your life certainly depends upon the honesty of such a reply.
00ps .... I almost forgot to give you ... COSMIC is an abbreviation for "Control Of Secret Material in an International Command" ......to try further help you understand the dire straits present condition of the current future situation you be in.
"Advanced IntelAIgents do not engage or compete against Natural Stupidity."
Don't be silly, amfM. That's EXACTLY what we are building them for ... to do the drudge work formerly done by the under-educated and ineducable.
If you want something truly worth railing against, how about the sorry state of education in the Western World? In theory we've had the time, money and capability to give every child a real education for over a century ... and yet we've intentionally crippled the education budget pretty much everywhere. One of the first places any government makes spending cuts has been in education these last fifty years or so ... and people wonder how Trump got elected? I give you exhibit A ...
"Advanced IntelAIgents do not engage or compete against Natural Stupidity."
Don't be silly, amfM. That's EXACTLY what we are building them for ... to do the drudge work formerly done by the under-educated and ineducable. ..... jake
?????? Pray tell, jake, how the "we" in all or any of their wisdom, expect the Naturally Stupid to do the drudge work formerly done by the under-educated and ineducable, and for the practise not to create a perfect base breeding ground for revolt and revolution, madness and mayhem in their midst and surrounding them and feeding them their daily bread, milk and honey.
That is a permanent ACTive vulnerability always ready for exploitation by more than just the Naturally Stupid anywhere and everywhere. To say such is one of the silliest of building projects is not the most accurate of descriptive superlatives to use whenever the program is so devastatingly self-destructive of elite executive officers and non-state support actors. ..... the nucleus core of the operation gone critical and rogue and into runaway China Syndrome meltdown phase.
Your last paragraph may perfectly highlight the underlying root cause of all of the present problems with madness and mayhem, conflict and chaos and why the novel solutions for viable resolutions will not be found in the conventional and traditional/homegrown harvested..
The reason education keeps getting cut is because no one can agree on what needs to be taught: not even the basics, no thanks to all those mental liberation movements and talks about conspiracies and everything being a lie over the past few decades. And if that means civilization stops having a stable footing, anarchists in the bunch will emphatically reply, "THANK YOU!"
And what could possibly go wrong with the following Fuzzy Wuzzy type action? .....
The Biden administration is gearing up to carry out cyberattacks aimed at Russian networks, the New York Times has reported, describing the provocation as a retaliatory measure designed to send Moscow a message. ..... https://www.rt.com/usa/517481-cyber-attack-biden-russia-solarwinds/
What are they toking/dropping/injecting in the USA?
Good rule of thumb-; If it's claimed that a government is about to carry out a sneaky action somewhere it won't be. Or at least not that particular one. Because you don't warn them of your move before you make it.
I read a SF story set in what was then the near future (the date in question is now in the past) in which the author had Roomba-like Artificial Stupid devices which sufficiently annoyed users that researchers did a bit of genetic engineering on assorted rodents, felids, and canids because it was easier to give them thumbs and training than to get the AS devices to work properly. The researchers failed to consider exactly what rats and cats with thumbs might get up to. Mayhem ensued. (The dogs behaved. It is possible that the author was not a cat person.)
“ “some of humanity’s greatest achievements arise from hunches, guesswork and pure luck rather than from the painstaking evaluation of evidence.”
I’m not sure if this is a positive trait, one of these days we’re going to be messing with stuff that can destroy whole universes, I’m still waiting for those pricks at CERN to throw some samples into the test chamber and open the Gates of Hell, albeit accidentally and unintentionally.
That prat Oppenheimer and his team, somehow managed to create a terrifying weapon and this fact eluded them, all the way up to detonation, only then did the fucking penny drop.
Yes mate, while you been busy solving problems, you failed to notice the multitudes of problems you’re dumb fucking brains are creating
Or so say Gartner.
This is the end of Big Data, now it's Small & Wide Data!