Schneier on Security

Clive Robinson • July 3, 2024 5:26 AM

@ Bruce,

Re : Title and sub

I’d suggest,

“AI and the Death of Democracy : How logic will eliminate reason.”

Yes it sounds a little on the “Machines are coming to get us” side of things.

But realistically seeing how politics thus legislation is becoming way more authoritarian and those behind them are becoming more and more like modern day “Robber Barons” using guard labour and look the other way politically selected judiciary, that know which way to find for their minority view (see recent SCOTUS vote).

You know that current AI that has no reason or ability to do so will be used and abused to act as the new guard labour judiciary.

The current AI builds statistical models on deep patterns in data by simple logic used to insane levels of depth. Why insane, because we know that changing the order the data is supplied in changes the results. Thus the system can be gamed to give not “reasoned results” but “desired results”.

Yes it will cost billions in run time to get the “desired models” but think of what that will give those who have the resources…

We know from history that this is more or less how it’s going to play out, because for some their ultimate desire is “control of others” and they see everything else as just a tool to reach their desired goal.

Society is about checks and balances. The checks limit excesses by the few over the many, the balances are formed by harms to the majority.

History shows the reason why,

“Feel my pain”

Is so important to society, it gives reason over logic.

Ask yourself what happens when a machine that does not and can not have perception of pain “by design”, is given data in ways that evokes fake patterns that get brought to the fore?

But also ask a question, why has the statistical predictions of election outcomes being increasingly divergent from what actually happens?

That is why the invention of the notion of “The Shy Voter” as an excuse for the errors. That is why so many voters are so ashamed of the way they vote we have mass lying to those polling them?

It’s more likely that it’s an error somewhere else down stream of the polling, that is the statistical models being built are wrong.

Pause to think what that meens for the building of current AI “Digital neural Networks”(DNN)…

At the end of the day human thus societal reason is based on perception of the environment and self. Current AI has no such perception or even agency to acquire it. Thus it will fall to it’s internal logic that is at best unstable and prone to chaotic divergence from the reality of human existence.

We already know this, but those who have so much vested in current AI are following to basic short term destructive syndrome paths

1, Not Invented Here (NIH).
2, Don’t Kill The Golden Goose.

The “short term personal gain” they see, is blinding them to the “long term societal harm” that their get rich quick schemes will cause. As I’ve pointed out before there is a balance

“Individual Rights v. Societal Responsibility”

Currently AI is well over to the “build an investment bubble and get rich” side. Where “It’s all ‘me Me ME’ thinking”. We can already see the harms to society via the harms to the environment.

A fun thought to consider,

Some have faith in “The Law of Code” thus we have “smart contracts” and people putting everything of stored value they have into them. We are seeing others doing “rug pulls” and other attacks and stealing the stored value. Now imagine you get an AI to the point it can find all those “code defects” and is driven by rules to find them and exploit them. How long do you think it would be before the AI had all the stored value?

Not because it has any use for the stored value (though it might). But because it’s been designed to “find and acquire”.

So stored value loss by “Logic over Reason” giving rise to loss of choice in society which is a form of “Death of Democracy”.

Clive Robinson • July 3, 2024 8:54 AM

@ Bruce, Sten,

Re : Hoaxes that passed sniff tests

“Aside: Wikipedia’s list of hoaxes is surprisingly interesting”

Yup it has three of note

1, Dihydrogen monoxide
2, Grunge speak
3, ID Sniper rifle

The first is “technically” not a hoax, as the molecule certainly exists, and it does follow the naming conventions.

It’s kind of been adopted in a Tongue-in-cheek way for various reasons. Which brings us onto the second yes it was a joke/hoax but it actually spawned for a while a slang. What that said about those that formed the “in-group” clique is something I’ll leave to their psychoanalysts and anthropologists.

But the third is one I’m sure many old timers from this blog will remember as it was featured,

https://www.schneier.com/blog/archives/2005/02/implanting_chip.html

And not unexpectedly got a “zombie angle” to it.

Also the fact that something is a hoax today does not mean it will not be true tomorrow or at some point in the future (eg Paul McCartney dying).

Clive Robinson • July 3, 2024 8:36 PM

@ CyberShow,

When I read,

“What changed was the size of the camera and hence
people’s willingness to let utility overcome embarrassment.”

I cringed, and immediately went into “To Much Information” mode.

Because my mind instantly realised that “camera” is just on of a thousand or more modern technologies that would fit right into that statement.

And we are both old enough to know that there will always be a few who will do something that defies reason, sense, and good taste, with the problem of where one leads, others will follow.

Bearing in mind what today now is in the UK, I think you will realise why the urban dictionary called it “The Farage Effect” from the number of entries,

https://www.urbandictionary.com/define.php?term=Faraging

Clive Robinson • July 4, 2024 7:02 AM

@ Bruce, ALL,

Something to consider is “dynamic algorithms” created by AI that will effect those at the bottom of the power hierarchy where all they have is “collective bargaining”.

There is an old joke called “Rules of the House” where the first rule is,

“If any of the rules become known by anyone other than the rule maker the rules will be changed!”

Usually when it comes to workers such a rule fails due to business process inertia and complexity. That is those administering and using the system take time to implement any changes.

Well AI will almost certainly change that.

Because AI systems will find patterns that are non obvious and too complex for most to be able to follow let alone sufficiently understand. Thus employers will exploit AI to gain an advantage.

Imagine that your average pay gets cut by 10% or more but you can not see why… Especially when a published rule change indicates that your pay should have risen.

That is the sort of thing AI systems will relatively easily find.

Is this likely to happen?

As has been observed in the past “Heck yes it will” in the US and other places the duty of the seniors is not to the workers but to the share holders. The only thing stopping major exploitation of workers is legislation and regulation. However as we know lobbying waters down legislation or restricts it’s scope.

In many places “gig workers” do not come under any legislation or regulation to protect workers. Which is why it is rapidly becoming a large organisations way to employ workers for what would be their primary employment.

If people have doubts about what Gig Employers will do even despite the well published info on Uber and the like, maybe they should read,

https://spectrum.ieee.org/shipt

Which shows how “collective action” by workers showed a gig employer “Shipt” was deliberately cutting their income to as little as $7/hour when legislation generally puts the minimum wage for employees at three or more times that and with health care etc.

Clive Robinson • July 4, 2024 7:25 AM

@ Bruce, ALL,

Speaking of AI books, anyone remember “Superintelligence”?

Well I’d more or less forgotten it as I thought it was hype bordering on fear mongering and disagreed with much of it at the time.

But I was reading “Humanity Redefined” and the author there “Conrad Gray” pointed out the book has just hit it’s tenth birthday in the UK,

https://www.humanityredefined.com/p/superintelligence10-years-later

Along with an interesting discussion as to the back ground and how the book has aged.

Clive Robinson • July 4, 2024 9:37 AM

@ Bruce, ALL,

Re : AI existential threat by environment.

I’ve mrntioned before that the current AI are chewing up energy and creating green house gasses faster than the cryoto-coin “work-factor” lunacy started.

Unlike the other “AI Existential threats” this is a very “clear an present danger” to not just humanity but much else in the world, not just living things.

Well it appears the MSM are starting to pick up on this as rather more than a “talking point”. For instance today “The Guardian” a fairly well regarded newspaper published,

“Can the climate survive the insatiable energy demands of the AI arms race?”

https://www.theguardian.com/business/article/2024/jul/04/can-the-climate-survive-the-insatiable-energy-demands-of-the-ai-arms-race

It’s not just energy but potable water as well, the spending is expected to actually go up and into the trillions of dollars even if the AI systems become more efficient…

Why? Because of a quirk noted with the first steam engines they were grossly inefficient untill James Watt made a series of engineering changes and they became more efficient so burned less coal. However it was quickly found that as the engines became more efficient, they became more economical to run, thus the number of users went up and so did the consumption of coal. As the article notes,

In economics, that phenomenon is known as “Jevons’ paradox”, after the economist who noted that the improvement of the steam engine by James Watt, which allowed for much less coal to be used, instead led to a huge increase in the amount of the fossil fuel burned in England. As the price of steam power plummeted following Watt’s invention, new uses were discovered that wouldn’t have been worthwhile when power was expensive.

Something that is almost certainly going to happen with Current AI in the short term[1], even if it proves to be useless for various reasons. In fact we are seeing the start of it with the reduction in DNN sizes to try and get them into PCs and laptops.

Along with that, are some other interesting perspectives, that might make people some what further disquieted.

[1] Something people should look at is a graph of the Nvida share price rise and graphs of energy and water demand around certain data centers.

Clive Robinson • July 7, 2024 11:49 AM

@ Bruce, ALL,

Re ” AI means “Post Open Source”.

It’s something that people need to consider.

The Court ruling on the monkey taking selfies and who “owns the copyright” of the resulting images, has implications. In that the Court ruling found for a “Human Only” on IP rights. Worse they in effect were passed to the owner of “the equipment used” to make the images.

So an argument could be made that anything you make with Co-Pilot or other AI assistance belongs to the owner of the AI.

So Microsoft would own or have rights in any code produced…

The other problem is that all the current Open Source licences are at best ambiguous on this as the assumption when they were first developed and subsequently be inheritance is that “the creator” is human and has the right to make “the work” Open Source.

Clearly either that is nolonger true, or it is now under question.

It appears that Bruce Perens is enquiring into this but Courts especially US courts tend to have their own views in ways that appear to favour the person with the largest wallet / deepest pockets.

Some have estimated it could cost upwards of ten million in legal fees to get a definitive answer through the US Court system.

The only thing Post-Open appears to have to say so far with regards AI is with regards input for training,

https://postopen.org/2024/04/30/april-30-make-post-open-ai-possible/

Clive Robinson • July 8, 2024 4:30 AM

@ Bruce, JonKnowsNothing,

Re : Conscious or not?

This Scientific American article looks into the subject of consciousness and why like intelligence it’s an illusive butterfly to pin down,

https://www.scientificamerican.com/article/the-mystery-of-consciousness-is-deeper-than-we-thought/

Back last century when doing the MSc course there had been a presentation about biological systems and part of it was about humans and pain as a feedback mechanism in evolution. I got chatting with one of the University Readers and pointed out that I thought as an engineer it was superficial, and he asked why. To which I answered because it does not explain the foundation issue. Which got a raised eye brow, to which I continued, it does not explain the evolution of the pain receptor only what happened after that. He looked surprised for a moment then laughed and said he’d not thought about it that way.

Just recently we’ve heard a lot about AGI, and it’s got into the MSM and even Government circles (The Bletchley Declaration and six monthly repeating summits). I find AGI even less than superficial for similar reasons.

At the bottom of the SciAm article you will find,

“I believe we can make rigorous scientific sense of teleological laws that work from future to present, ensuring that what happens in the present is dependent on the need to get closer to some future goal, such as the goal of harmonious alignment of consciousness and behavior.”

I don’t like the expression “teleological laws” not just because very few have ever heard the word let alone know it’s meaning, but because it sounds scary.

Most however have heard of “goal seeking” by the time they leave school and quite a few have already thought about “their future” and the steps of how to get there. So the concept is “not unknown”.

In short it assumes “purpose”. Interestingly it’s a subject that has been arising in legal and philosophical circles this past decade with the question of,

“Does the law have purpose?”

https://academic.oup.com/book/8280/chapter-abstract/153885282

In a way it’s the question about behaviour being a down hill gradient of “reactive” or an uphill one of “proactive” as most of us know “going with the flow” of the down hill gradient is usually a lot easier and requires almost no effort.

Thus the question as to “Why go uphill?” arises not just with law but evolution in general.

It’s a “foundation question” and one that underlies not just consciousness but actual intelligence.

Like the author of the SciAm article Prof Philip Goff indicates with,

“Some philosophers have argued that psychophysical harmony points to God. I think that’s a bit of an overreaction, but I have argued that dealing with psychophysical harmony takes us in radical directions, uprooting our most fundamental assumptions about reality.”

The usual “because deity” is really a silly argument because it is a “lesser flea” or “turtles all the way down” avoidance by pretense the question is unknowable / unanswerable.

But the Prof is right about one thing

“takes us in radical directions, uprooting our most fundamental assumptions about reality”

And all the law, politics, and AGI arguments of AI are just going to be “grist to the mill” of that.

Clive Robinson • July 11, 2024 4:32 AM

@ Vint Cerf,

Re : Running with scissors

“Incidentally, we have the same risks with software in general when we give it real-world autonomy to take actions.”

Yup giving the majority of software any kind of agency makes me nervous.

In part because back last century I used to do “safety critical” and “intrinsic safety” design. Some of which was used for billion dollar worth systems in the late 1980’s that are still in operation today, still using what I’d designed…

As my beard turns more white than grey, I find myself dwelling on what makes systems “adaptive” to the environment they are in. Which is arguably the second or even later step to actual AI[1].

Yes I smile when I hear about soldiers hiding in cardboard boxes and giggling past an AI system designed to recognise “Human threats”… But the real lesson is not lost on me.

https://www.businessinsider.com/marines-fooled-darpa-robot-hiding-in-box-doing-somersaults-book-2023-1?op=1

[1] The first step and one we do not have a clue on how it came about is “goal seeking” even in it’s most primitive forms it’s an “arm wavery” subject. But as it is what underpins all our notions about evolution, it underpins all that we know as life and so actual intelligence as we currently think of it.

Clive Robinson • July 11, 2024 11:30 AM

@ ResearcherZero,

Re : AI hoax, threat or both?

“People who believe that AI is all a hoax, greatly underestimate it’s power.”

I know that what we currently call AI that is LLMs and ML is a hoax, as far as “Inteligence” goes and that the route we are currently taking is a cul-de-sac in that regard.

Likewise I know that the “AGI” stuff that has been spouted is to be polite a load of at best “faux marketing blather”.

But as I’ve frequently pointed out the current AI systems are an almost perfect “arms length” way to implement prejudice or political mantra or worse as the RoboDebt systems have clearly shown significant harm.

Also the,

“Computer says No”

Curtailment of people trying to get “faults” rectified is just the first of many harms that a reasonable person can not fight.

It is therefore an excuse that gets used frequently to get rid of people on phones. Also because nobody who pushes such discrimination ever has to take responsibility and punishment for it.

So I’m very definitely in the AI as is, is “Both a hoax and harmful threat”.

Clive Robinson • July 11, 2024 5:24 PM

@ Bruce, ALL,

Re : AI Hallucinations bad in knowledge domains.

We are aware of what the AI industry calls “Hallucinations” but “the term of art”[1] in an overseeing knowledge domain is either “Soft Bullshit” or occasionally “Hard Bullshit”.

Where “soft” is in effect producing any old nonsense output as long as it sounds good. And “hard” is deliberately producing known to be factually incorrect output.

But how often does this happen?

Well according to an ACM article[2],

“Stanford University’s RegLab, showed that hallucinations are “pervasive and disturbing” in response to verifiable legal queries, with rates ranging from 69% to 88%.”

If I read that correctly what you get back between effectively 7 and 9 times out of 10 queries is “Soft Bullshit”.

Think on that for a moment a near 90% of the time the answers are neither accurate or useful… And from other sources I understand that the 10% of answers that are verifiable are actually not of much quality or use (in effect direct quotes of legislation).

Yes the more specialised the knowledge domain the less likely current AI systems are to give you even close to accurate, therefore useful or unharmful information.

But another article quote,

‘As Brent Mittelstadt, director of research at the “Oxford Internet Institute(OII) of the University of Oxford, U.K., explained, “There is no absolute requirement for them to be factual, accurate, or line up with some sort of ground truth.”’

From where I come from that’s called “an ouch moment” and thus most definitely “Soft Bullshit”.

But it gets worse a team of computing experts at the “National University of Singapore”(NUS) “AI Institute”,

“used learning theory to argue that hallucinations are inevitable and cannot be eliminated. LLMs will always hallucinate as they “cannot learn all of the computable functions“

Note carefully what “computable functions” actually means with regards work from the 1930’s.

But back to Brent Mittelstadt and colleagues at Oxford University OII they have a somewhat quaint view point,

“Rather than using a language model as a knowledge repository, or to retrieve some sort of knowledge, you are using it as a system to translate data from one type to another or from one format to another.”

Anyone who has used Google Translate should know this is something you really should treat with considerable caution, especially when you go beyond “conversational translation” to things that have actual technical or scientific meaning.

So, what to conclude from these quotes by experts?

My view is,

“The current LLM and ML systems are even less fit for market than bug ridden consumer software.”

Like that Microsoft pushed out with say Vista…

Further that they are designed as “Surveillance Systems” with a business model of “The five Be’s” of

Bedazzle, Beguile, Bewitch, Befriend, and Betray”.

And to do this any old junk-in-a-box tech behind the curtain will do even if it pushes 100% “Soft Bullshit”.

There may be more hope for “small models” that run on augmented PC hardware that are built from highly verified and complete corpus’ input.

[1] The problem as I’ve noted before is the word now used as a “term of art” from a book in 2005 not just in it’s original knowledge domain but increasingly others, has frequently been found on word “naughty lists” for profanity and the like,

https://link.springer.com/article/10.1007/s10676-024-09775-5

[2] The ACM article by Karen Emslie,

https://cacm.acm.org/news/llm-hallucinations-a-bug-or-a-feature/

Clive Robinson • July 11, 2024 7:56 PM

@ Bruce, ALL,

Re : The irrationality of LLMs is not human…

This one about research from one of my old stamping grounds “University College London”(UCL) tells us something we should have expected,

“Cognitive psychology tests show AIs are irrational—just not in the same way that humans are”

https://techxplore.com/news/2024-06-cognitive-psychology-ais-irrational-humans.html

Clive Robinson • July 14, 2024 7:21 AM

@ JonKnowsNothing, ALL,

Re : Power to the AI is heating up many places.

More on the AI cost and other issues. That contains this jolly little statement,

“Major corporations have even now started looking at modular nuclear power plants just to ensure that their massive AI data centers can get the power they require.”

Yup read that again and let it sink in…

This is not a prediction from what some would call “a crank group of environmentalists, fear mongering”.

But one of those institutions very very close to the beating heart of the Capitalist way “Goldmam Sachs”,

https://www.tomshardware.com/tech-industry/artificial-intelligence/goldman-sachs-says-ai-is-too-expensive-and-unreliable

And Goldman Sachs appear to be in two minds as to if AI as currently practiced will even break even,

“We still see room for the AI theme to run, either because AI starts to deliver on its promise, or because bubble take a long time to burst.”

With one “mind” saying that,

“The capital expenditure cycle on AI technologies seems promising and is similar to what prior technologies went through.”

Which is the brain fever madness of “just throw money in and double-down over and over”. A strategy which has apparently already got to the point where Sequoia Capital, a fairly well known Tech VC, has recently taken a sober look into “AI investments”. And has calculated that the visible part of the AI tech sector,

“Needs to make $600 billion annually just to break even on its initial expenditure.”

And on the other hand, the other “mind” of Goldman Sachs, more rationally voicing the contrary view that says,

“AI will only deliver limited returns to the American economy and that it won’t solve complex problems more economically than current technologies.”

As you know I’m on the far side of even that “optimism” of the latter mind.

I do not think the current AI LLM systems will ever break even when it comes to the way LLM AI is currently being developed. Also I do not think LLM AI systems will meet the many claims made of it in the AGI hype that has more recently accompanied it. And all the while it is being pushed beyond hype by “Venture Capitalist”(VC) types and shrill shills into an “Investment Bubble” for those with more money than sense.

But even I have one uncertainty, that is if the bubble will burst explosively or just deflate like a weak old party balloon hidden out of sight under the bed as an intoxicated prank.

And as I’ve previously pointed out the CPU Cycle investment bubble issue so far has a succession of failures from crypto coins, through NFT’s, Smart Contracts, Web3, and VR based Metaverse,

https://www.bbc.co.uk/news/articles/c51yl7q8z42o

All of which needs the same basic power hungry hardware tech.

In this series of failures view I’m not alone, Molly White who was getting more attention than she expected for her hobby so went full time is as, if not more, skeptical,

https://www.web3isgoinggreat.com/what

And the occasional poster to this blog Nicholas Weaver has observed in “The Web3 Fraud”, Web3 is just another way to push the CPU cycle burning crypto ponzi scheme,

“So why this hype? Because the cryptocurrency space, at heart, is simply a giant ponzi scheme where the only way early participants make money is if there are further suckers entering the space. The only “utility” for a cryptocurrency (outside criminal transactions and financial frauds) is what someone else will pay for it and anything to pretend a possible real-word utility exists to help find new suckers.”

Exactly the same logic applies to the current AI LLM and ML systems, that are just throwing more fuel on the bonfire of the world.

But I admit it is a “surprise surprise” moment to see that Goldman Sachs, at least in one mind, is in agreement.

Clive Robinson • July 15, 2024 2:38 AM

@ JonKnowsNothing, Bruce, ALL,

Re : Does the foot fall or fall off?

Whilst not actually “AI” the subject of “robots” is never far away as people look to giving AI systems “Agency”. It appears to be rarely out of the MSM or Tech/Trade journals, be it delivery drones for civilian or military use or electric vehicles with a safe self driving capability.

But all those robots are “Hard Robots” in that the in effect have a rigid frame on which sensors and actuators are fixed. The biological equivalent is a skeleton be it internal as with vertebrates or external as it is with most insects but there are other life forms that do not have hard structures such as Squid, Octopus, Starfish, sea cucumber and similar.

As an overly general rule Vertebrates are animals that have a backbone inside their body that is articulated. Not just using bone but cartilage and ligaments. In structures that surprise even engineers.

But even more surprising for engineers are those creatures that lack skeletons. These “soft creatures” have given rise to the design of robot analogues that are called “Soft Robots”.

On aspect of which is to mimic what some creatures like lizards and insects do. That is to detach parts of their body or in the case of some ants fuse bodies together.

NASA amongst others are looking into this, especially the idea that small general parts can build larger specialised function robots as required. Thus giving space craft the ability to self-repair rather than use redundancy, and ultimately self replicate such that exploring the vast volumes of space becomes much more efficient.

But less obvious is what soft robotics can potentially do for prosthetics. As I mentioned in my post above an active area of current prosthetic research is “sensory feedback” where by a prosthetic device gains the ability to adjust automatically to it’s environment.

Feet and footing are the obvious candidates as unlike our hands nearly all the movement of the parts of the foot are non conscious and in response to what the foot sensors ascertain about the environment.

After a few moments thought it will can be realised that AI based “sensory feedback” and “soft robotics” will form the underlying technologies in such prosthetic feet.

But a very recent and interesting area of research in soft robotics which will end up involving AI is the ability to detach and attach soft robot parts to not just mimic but improve on what lizards and ants can do.

https://spectrum.ieee.org/soft-modular-robot

The potential for use in low gravity environments is immense.

But think about a mixture of hard and soft robotics, and vertebrate spines. Do you build limbs like boneless squid tentacles, snake or centipede bodies, or what we think of as more normal long bone limbs?

The possibilities are shall we say not just interesting but potentially bring SciFi ideas into reality.

Also consider that we are looking at prosthesis to argument humans as force multipliers. In medical environments much is limited by what a human can physically do mechanically without risk of harming themselves. An individual unassisted nurse can not lift patients in and out of bed into chairs, baths etc thus either multiple nurses or hoists and other lifting equipment is required and this is very limiting and resource awkward, hence the interest in exoskeleton systems.

Further anyone who has injured a leg or even an arm knows how debilitating it can be to ordinary existence whilst it heals. And how even minimal assistance makes a heck of a difference to much quality of life and independence. There is not an engineer alive who has not at some time wished for a third hand, and most people will realise that similar thoughts have happened to them.

Soft robotics and AI will make these things not just possible but eventually everyday. What concerns me is the route we take to get there.

Few people under the age of thirty realise that the primary use thus development arena of technology was by the areas of Government involving international relations and social control via various “civil servants” and “guard labour”. It was not till the 1960’s and through into the 1990’s that this substantially changed. Now we think of civilian and industrial technology trickling down into the military and such like.

However some are realising that with robot technology the old “military first” pathway is reasserting it’s self through the likes of drones and loitering munitions and what is now called “Sixth Generation” weapons systems.

As noted Hayek’s excusing sociopathic behaviour in economics is bringing society to the brink of disaster.

Thus how we move forward with AI and the agency it gets given by robots is of concern to rather more than Google employees. But letting the fox dictate the rules of guarding the hen house has never been a good idea. Remember whilst a poacher might become a gamekeeper, they almost always “act in service” rather than be “masters”.

Clive Robinson • July 15, 2024 5:09 AM

@ Bruce, ALL,

Re : The Red Queen dictates but does not rule.

The current so called “AI market” is at best an illusion created by sleight of hand. At worst a Ponzi scheme or investor bubble depending on which side of a quaint legal divide you take others money.

The thing is history shows that to get going new technology has to be dishonest in some way to get the funding needed to get past the inertia and thus the ball rolling.

Because few will invest in what will take more than the rest of their life to pay real dividends.

Some claim AI will be the next “Industrial Revolution” not realising that it all ready has been but under a variety of different names. The thing is industrial revolutions take the better part of a couple of centuries to happen (although it is speeding up as communications improves). Because the first signs are generally unrecognised untill much later by historians, timescales appear shorter.

For those that do not know AI systems were in practical use back in the 1980’s with the recognisable initial work going back into the 1950’s. For my sins I was working in developing them in industrial control systems in the 1980’s and some are still running today, much to my surprise four decades later.

The thing about the technology side of all industrial revolutions is it starts in curiosity then research long before it becomes usable. And in usage it goes from the highly specific nich to the general and ubiquitous.

One classic example of this would be the electric light bulb. It replaced gas lights that had replaced oil lamps that had replaced candles. Each clearly demonstrated advantages that not just outweighed it’s disadvantages but the disadvantages of the proceeding technology. But few realise that quite a while before the wire filament lamp became practical the arc light was being used quite successfully. Sir Humphrey Davey demonstrated both the platinum filament and carbon arc as light sources in the first decade of the 19th century. This was 75years, or a lifetime before Swan and 80 before Edison. During that period both Limelight and the carbon arc light were developed and used to great effect. During the first forty years or so the filament bulb was a curiosity in what we would call researchers laboratories or inventers workshops. As such it had little visibility or interest to the general public even as a quaint novelty.

At the moment with AI we are kind of at the tipping point, where it becomes recognisable to most that there is some advantage over what is already there.

I was reminded of this, when the current news letter of Jamin Ball came to my attention over the weekend,

https://cloudedjudgement.substack.com/p/clouded-judgement-71224-the-red-queen?utm_source=post-email-title&publication_id=56878&post_id=146443935&triedRedirect=true

Note the premise of the “Red Queen Race” thought up by Victorian logician and mathematician Charles Dodgson for the entertainment of a child.

It’s actually an indicator of “falling for the hype”. Yes whilst existing AI systems under different names are showing quite profitable returns, the current LLM and ML systems probably will not ever. They are at best niche solutions looking for general problems to solve, at an exorbitant use/cost of resources.

The resource cost is these AI systems Achilles heal and within a fairly short time they will either be dropped or something more efficient will replace them.

So “my money” would be on the latter of being replaced by something more efficient.

There is an expression that almost always applies to new technologies,

“The leading edge is the bleeding edge.”

And few things survive exsanguination or the traditional method of slaughter. Which is why the newer expression of,

“It’s not only the cheese the second mouse gets”.

As far as the current AI LLM and ML systems go, you don’t want to be a “first mouse”.

Clive Robinson • July 15, 2024 1:49 PM

@ Winter,

Re : Power and water resources.

I guess you are not looking at things in the whole with,

“I utterly fail to see the connection with crypto currencies. I have actually seen many practical applications of ML and deep learning and used them myself.”

Well let me think,

1, The use of Nvidia GPUs in vast arrays.
2, The electrical power to run those arrays[1].
3, The potable water to cool those arrays[2].
4, Similar costs for comms support[3].

It’s been estimated that a single run in building a LLM from training data can be up in the tens if not hundreds of millions in electrical power alone. Then there is similar high costs for potable water. As for the cost of the GPU array and the infrastructure hundreds of millions again. With I’ve no idea how to assess it but the environmental costs from component manufacture onwards.

But it would appear that LLM systems are really quite grossly inefficient and the DNNs not what many think for various reasons,

https://robleclerc.substack.com/p/general-theory-of-neural-networks

As I mentioned subject me time ago the “Multiply and Add”(MAD) network is usually configured in a way that is not exactly effective or efficient (see difference between “inside the sum” and “outside the sum” for the nonlinear components (it’s at the end of the appendix).

[1] From what has been said Alphabet and Microsoft now consume as much electricity as two of the larger EU States.

[2] It’s difficult to get reliable information about water consumption but I’ve seen over a litre per query being given, and as for training well lets just say enough hot water to make a large cup of coffee for every one in the US at least.

[3] You might not be aware of it but the total world power usage in comms for watching videos off of the Internet is estimated at being 1% of the world electricity consumption…

Clive Robinson • July 27, 2024 11:26 AM

@ Bruce Schneier, ALL,

Re : No democracy with model collapse.

There is a problem with current AI systems based on LLM and ML systems.

The are way to susceptible to “model collapse” which has serious side effects. Some of which are detailed in

“AI models collapse when trained on recursively generated data”

https://www.nature.com/articles/s41586-024-07566-y

“We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’ and show that it can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs). We build theoretical intuition behind the phenomenon and portray its ubiquity among all learned generative models.”

(Open Access Published : 24 July 2024)

In short “data based on previous data” amplifies the GIGO effect.

Perhaps not new news to some but the effects as in analogue and DSP filters used in regenerative feedback loops can be “oscillatory”. At the very least it causes ‘Q-multiplication’ effect the results of which are the “tails of the original signal distribution disappear” rather rapidly.

Unfortunately the “Democratic Political Process” is very much all about “feedback loops” within “feedback loops”. Even without AI systems the effect of MSM and Social Media often cause the very polarised effects we see and the consequent problems that follow. Current AI systems can not make this better, in fact they can only make it worse.

Clive Robinson • July 30, 2024 10:14 PM

@ Bruce, ALL,

Re : Can an AI be an observer?

In,

“Metaphysical Experiments’ Probe Our Hidden Assumptions About Reality”

https://www.quantamagazine.org/metaphysical-experiments-test-hidden-assumptions-about-reality-20240730

A discussion is had that initially appears detached.

However after some convoluted ground work, towards the end a rather important question arises, about what it means to be an observer.

The article points out a rather thorny issue currently,

“Some philosophers of mind believe in the possibility of strong AI, but certainly not all. Thinkers in what’s known as embodied cognition, for instance, argue against the notion of a disembodied mind, while the enactive approach to cognition grants minds only to living creatures.”

And yes there are other views as well with Roger Penrose and David Deutsch getting mentioned,

“… he read books by the physicists Roger Penrose and David Deutsch, each offering up a radically different metaphysical story to account for the facts of quantum mechanics. Should we give up the philosophical assumption that there’s only one universe, as Deutsch suggested? Or, as Penrose preferred, perhaps quantum theory ceases to apply at large scales, when gravity gets in on the action.
“

Roger Penrose proposed the notion of “quantum effects behind our intelligence” a subject that still caus s argument, and David Deutsch gets credited as being the mind behind quantum computing.

But nobody has yet answered the question of,

“What it takes to be an observer”

Some would think any sensor from a simple mechanical switch through to the likes of the LHC are “observers of events” but that is just a fraction of what “an observer does”.

As I’ve noted several times in the past,

1, Technology is agnostic to use.
2, The use is chosen by a directing mind.
3, It is an observer who decides if that use is good or bad.

After a few moments thought many will realise that machines whilst they can follow “the rules embedded in them”, can not decide good or bad for a novel use.

So can AI be an observer?

Well certainly not with the current LLM and ML systems, no matter how arm wavery the AGI proponents get.

Clive Robinson • August 3, 2024 2:16 PM

@ Bruce,

Will the AGI bubble will be gone before year end?

Some certainly think so, the author and AGI skeptic Gary Marcus certainly thinks so,

“I’ve always thought GenAI was overrated. In a moment, though, I will tell you why the collapse of the generative AI bubble – in a financial sense – appears imminent, likely before the end of the calendar year.

To be sure, Generative AI itself won’t disappear. But investors may well stop forking out money at the rates they have, enthusiasm may diminish, and a lot of people may lose their shirts. Companies that are currently valued at billions of dollars may fold, or stripped for parts. Few of last year’s darlings will ever meet recent expectations, where estimated values have often been a couple hundred times current earnings. Things may look radically different by the end of 2024 from how they looked just a few months ago.”

https://garymarcus.substack.com/p/why-the-collapse-of-the-generative

(Warning it’s a “pay to read” page on his site that if you don’t have a reason to be already paying, I’d give it a miss and read his recently preceding posts).

Clive Robinson • August 14, 2024 3:01 PM

@ Bruce, ALL,

In case people think I’m being ‘unfair’ in some way on current AI LLM etc systems, remember back a long time ago at a University far far away I used to play not just with AI systems but Robots as well.

So I’ve seen some of the issues most have not with the combination of them both, that is,

1, Giving robots limited action capability via primitive AI
2, Giving AI ML systems agency to explore the environment they are in

But I’m not alone in this there are others,

Over at MIT CS/AI labs in the Robotics domain Roboticist Prof. Rodney Brooks,

https://people.csail.mit.edu/brooks/

Has come up with his,

“Three Laws of AI”

https://rodneybrooks.com/rodney-brooks-three-laws-of-artificial-intelligence/

” 1, When an AI system performs a task, human observers immediately estimate its general competence in areas that seem related. Usually that estimate is wildly overinflated.

2, Most successful AI deployments have a human somewhere in the loop (perhaps the person they are helping) and their intelligence smooths the edges.

3, Without carefully boxing in how an AI system is deployed there is always a long tail of special cases that take decades to discover and fix. Paradoxically all those fixes are AI-complete themselves.”

Clive Robinson • September 12, 2024 1:19 PM

@ ALL, Bruce,

Re : Things to get ahead of.

I’ve pointed out that the current LLM “Digital Neural Networks”(DNN) are for a whole heap of reasons not the way to go.

In the past I’ve mentioned why the “Multiply and ADd”(MAD) DSP style algorithms used in current AI DNN’s are both unsuitable and actually the wrong way around.

Further I’ve mentioned that “Biological Neural Networks” do not work on “levels” but regular pulses as though from variable frequency oscillators and why this works significantly differently to the nonlinear transforms used in DNN’s and,

“Importantly does not have the same limitations.”

Well it appears after a decade and a half and maybe a dozen papers that the academic community is starting to think on this,

https://pubs.aip.org/aip/apr/article/7/1/011302/997386/Coupled-oscillators-for-computing-A-review-and

For various reasons I think this is likely to be what will be the next stage of underlying AI systems.

So should be seen as “future Directions” rather more than “current”.

But also loosely coupled oscillators in quantum resonator systems, allow you to do some of those “Quantum Computer” things that are getting people so twitchy.

Clive Robinson • September 19, 2024 9:42 PM

Another subject involving AI and Security is that involving “Agency”.

Currently through IoT we are adding “world aware sensors” to networks rather rapidly.

The problem is they are mostly insecure in various ways, but worse many are deliberately designed to “ET Phone Home” to “The Mother Ship” one way or another to increase “surveillance on individuals” on mass.

But starting to appear since the 1980’s and 90’s are cyborgs / transhumans. Untill recently they were in effect “cut-n-shut” operations to put things as simple as magnets in people.

I know from having had for medical reasons a smaller than matchstick sized “reveal” cardiac monitor there are issues with cut-n-shut one of which is the build up of “fibroids” around them (think like a perl forms around grit in an oyster but with soft tissue).

But that is not stoping people having hearing and visual implants to make up for what in essence is an unfortunate biological handicap or disability.

But some people are going in different ways,

https://www.bbc.co.uk/news/articles/cg58r70yj43o

Where the level or type of augmentation is going beyond that of correcting for a handicap or disability.

It should be obvious from just looking at the explosion of IoT devices and types just what could be done if the sensors are “connected effectively” to the human neural network.

It’s not much being said, but with a little thought it becomes clear that,

1, AI ML and even LLMs will be required.
2, The level of security needed is nowhere even near to being what it needs to be.

In fact with regards to AI it’s entirely unclear if we actually have any real idea as to how to find let alone address the real processing security as opposed to communications and storage security issues.

But like it or not cyborg / transhuman augmentation has in effect reached a tipping point with the help of AI. To do interface translation between the standards we use to computer networks and biological neural networks. This is unlikely to stop or slow down and the next decade will see considerable advancement in this area, probably more so than we will see with Quantum Computing advancement.

Importantly it will happen in three basic directions,

1, AI systems will get increasing agency via IoT sensors and existing networking.
2, Biological sensors will become increasingly interfaced to existing digital networks.
3, Both electronic and biological sensors will get interfaced by some level of AI to biological neural networks.

The two questions that we need to think about now are,

1, Where will this take us?
2, What of privacy, autonomy and security of the individual as we moce along the journey?

Clive Robinson • September 21, 2024 6:25 PM

@ Bruce,

You and I are according to the Internet of a similar age. Hopefully you are aging better than I.

As you are possibly aware my health has been an issue for many years and I’ve had to “grasp the nettle” of the medical profession with a distinct firmness. And whilst having to battle doctors into suitable sane paths of action has been possible in the past, I note that as the age disparity increases as the doctors in effect get younger with respect to myself I find myself cast as being “aged and thus unreasonable”. With reason, and even sanity in the doctors diminished in preference to “the computer says”…

That bolsters their “Do as you are told” mentality an inexperienced adult might try to inflict on a pet or child.

It’s got to the point where I find myself wondering about the future of medicine. As you might be aware long ago back in the 1980’s and even 90’s I had involvement with “Robotics and Artificial Intelligence” and did some work on trying to combine “Expert Systems” and “Fuzzy Logic” to try and soften if not remove the issues of both (without much success).

Expert Systems once “skinned as human” by the “Turing Test Illusion” were especially problematic. Because they were in effect rule following decision trees that could not deal with any kind of ambiguity, amongst other major failings. Thus to humans they were dogmatic and inflexible, worse they did not change with increasing data and were highly sensitive to the order data became available…

All of which came to mind again with reading,

https://www.nature.com/articles/s41599-024-03282-0

“Paternalistic AI: the case of aged care”

As noted in the Abstract,

“In this paper, we argue that AI systems for aged care can be paternalistic towards older adults. We start by showing how implicit age biases get embedded in AI technologies, either through designers’ ideologies and beliefs or in the data processed by AI systems. Thereafter, we argue that ageism oftentimes leads to paternalism towards older adults. We introduce the concept of technological paternalism and illustrate how it works in practice, by looking at AI for aged care. We end by analyzing the justifications for paternalism in the care of older adults to show that the imposition of paternalistic AI technologies to promote the overall good of older adults is not justified.”

Now consider other forms of,

“Technological Paternalism”

We see Paternalism of a similar form in “cults” and certain types of religion we call “cargo cults” that have their analogues in most societal domains including regulation, legislation, politics ethics, and mores.

AI is in essence going to behave in an unacceptable way and at high speed, with the ability of the bulk of society to address and rebalance as once was possible, being rapidly diminished.

Part of which is AI ML will just “eat it’s own crap” thus re-enforce it’s deviancy away from society as “crap begets crap”.

Clive Robinson • September 29, 2024 10:45 PM

Is AGI intractable?

This paper,

https://link.springer.com/article/10.1007/s42113-024-00217-5

Whilst it is not the easiest to read makes the statement,

This practice has taken the theoretical possibility of explaining human cognition as a form of computation to imply the practical feasibility of realising human(-like or -level) cognition in factual computational systems, and it is framing this realisation as a short-term inevitability. In this paper, we undercut these views and claims by presenting a mathematical proof of inherent intractability (formally, NP-hardness) of the task that these AI engineers set themselves. This intractability implies that any factual AI system created in the short-run (say, within the next few decades or so) is so astronomically unlikely to be anything like a human mind, or even a coherent capacity that is part of that mind, that claims of ‘inevitability’ of AGI within the foreseeable future are revealed to be false and misleading.“

Clive Robinson • October 6, 2024 6:53 AM

As real data arrives AGI hype dies.

There has been much hype about AGI and the fact it will be a 10x programmer or better, giving over a 1000% increase in productivity…

Well the real world data coming in says not ‘It ain’t necessarily so’ but it’s actually worse than a 0.5x programmer, who leaves technical debt like dung in the Augean Stables so necessitating a Herculean task to clean up.

https://garymarcus.substack.com/p/sorry-genai-is-not-going-to-10x-computer

If you think about it this is to be expected.

There is no actual intelligence in the current AI LLM and ML systems and never will be. They are based on statistics of “the most common available” input and perturbed by a degree of noise (hence ‘stochastic parrot’). In effect they are the opposite of an “Expert System” from the 1980’s.

So given the sources, is it really unexpected that low quality, inefficient, and error ridden code predominates the input to these LLM systems, thus would get statistically ranked higher?

Clive Robinson • October 8, 2024 1:57 PM

@ Bruce,

This “AI in Politics” story is one to put in the research file.

Titled,

“Virginia congressional candidate creates AI chatbot as debate stand-in for incumbent”

It actually sounds not to dissimilar to that an independent candidate in the UK tried to do a few months back …

But this Virginian is way more “in your face” about it,

“Bentley Hensel, a software engineer for good government group CivicActions, who is running as an independent, said he was frustrated by what he said was Beyer’s refusal to appear for additional debates since September. So he hatched a unique plan that will test the bounds of both propriety and technology: a debate with Beyer’s artificial intelligence likeness.”