Schneier on Security

Clive Robinson • June 3, 2026 6:08 PM

@ KC, ALL,

With regards,

“It’s equally wild to observe the ‘beyond LLM’ camp make advancements in domains such as quantitative AI, and wonder what its effects on the field will be.”

Like any other “Castle built in the clouds” it is not realistically sustainable.

The first indicator is when you hear nonsense from the AI thought leaders like,

“Grow the economy by 10% every year”

It’s not going to happen and is easily provable that it’s not.

To give people a clue as to why and why they need to ground their thinking some basic facts,

1, Due to the gravity well the Earth is very nearly but not quite a bound environment when it comes to matter.
2, Due to other effects the amount of energy from the Sun needs to be balanced out by capture or re-radiation or we will cook.
3, Coherent energy sources decohere when the energy when any work is done and transports down to become heat.
4, Heat is the ultimate form of pollution in that to “actively transport it” to clean it up you need to generate more heat as work is being done.

Thus the economy can not really grow except by making work more efficient and there is a both a practical and theoretical limit on that both of which you can calculate if you feel so inclined.

The only way to change things is by making the movement of heat “energy” more efficient so you get to use the more coherent energy from the Sun and thus need to be able to push the less coherant energy into space so the balance is maintained.

As far as making tangible objects goes untill we get out into space the majority of our matter remains constant. It’s why effective and efficient recycling is now a primary requirement for all living things in the Earth’s sphere of use / environment.

But not all matter is finite or reusable.

A small amount drops into the Earth’s gravity well each year and although we say “burn up in the Earth’s atmosphere” that gives a false impression to most humans. That matter usually remains on Earth.

However some of the very light elements can escape the Earth’s gravity well which is why the loss of the second most available form of matter in the Universe “Helium and it’s isotopes” is becoming an increasing concern in the more interesting parts of the Science and Technology sectors with the likes of Body Scanners and Quantum Research that will eventually replace a significant amount of current technology are all critically effected by the increasing scarcity.

Then there are the heavy elements and their isotopes that decay by releasing energy/matter. We are loosing or using some way faster than they can be replenished from other parts of the Universe.

All of which indicates that unless we can increase efficiency faster, and get out of the Earth’s gravity well to acquire new resources, in the long term the “real economy” rather than the faux “financial economy” can only decrease.

Put overly simply the difference between the “real economy” and the faux “financial economy” is “inflation”. To decrease inflation to keep fiscal stability requires all “work / processes” to become less wasteful or “more efficient”. But as “the laws of nature” inform us there are very real limits on increasing tangible “efficiency”…

So much of what the “finance industry” and “Techbro industry” tell us is complete nonsense and based on the faux notion that “dialing up the digits” is “growth” when in fact it is simply a deceitful way to transfer “real tangible wealth” away from the majority of people.

Hence the WEF etc mantra of,

“You will own nothing”

You will only be allowed at best to,

“Rent what you need to survive”

And you will not be able to, let alone be allowed, to survive if you can not pay the ever increasing rent.

To live like that is effectively worse than being a slave… But don’t take my word for it, have a look at history. Including that recent history where we moved from an agrarian to industrial existence.

Before “industrial” wars were fought over simple resources like fertile land and more importantly water. Look up “water wars” and “salted ground” to see just how nasty some authoritarians have been.

The wars of today are more about mineral and similar resources with “Energy Wars” replacing “Water Wars” but practiced by the same sort of self interested malicious authoritarians…

I’ve talked about this on this blog and other places for a decade or more now. Back then the change was more discreet, now however in the past half decade or so, it’s become so obvious that you just have to look with unblinkered vision.

Clive Robinson • June 13, 2026 7:18 AM

@ Oscar Galvis, ALL,

With regards,

“This single abstract breakthrough forced a complete reassessment of the long-term security of RSA—a feat of structural disruption that no pattern-recognition model can replicate.”

A nice argument for a point of view I tend to agree with, especially when talking about Current AI LLM and ML Systems.

That said the “Stochastic” element also needs to be considered.

In many respects it’s been made before with the,

“It’s the one monkey that happens to type Shakespeare.”

The thing is all the monkeys type some part of Shakespeare in some order if given sufficient time.

It’s actually the same argument as any finite number string can be found in the digits of Pi.

But it’s actually not very useful other than as a philosophical argument that boils down to the answer of,

“To contain everything, it must contain nothing.”

In terms of “useable information content” rather than the text is in effect “unusable noise or garbage”.

That is,

“Useful communication is useful because of what it does not give you, rather more than what it does give you.”

That is if it contains everything it contains nothing that is usable, thus it contains nothing.

You can see the start of the argument with the idea of,

“A can of paint that is of a chosen colour.”

The fact is it’s actually not a single colour it’s a mixture of several colours all stirred together randomly. In theory and in practice any individual molecule could be in any part of the can. But it is overwhelmingly most likely that the distribution of colours will be in effect even around the mean[1].

So the whole can looks like the wanted single colour after it’s been prepared as directed on the can

The reason for that “direction” is the simple fact as the molecules in the paint have different weights or sizes/shapes given enough time the can will visibly start to stratify as it comes under the effects of gravity and some motion.

So whilst the claim that the answer to every question you could possibly want to ask is contained somewhere in the digits of Pi might be true… So must the questions be present as well[2]. But effectively in no relationship to the answers so it’s practically useless without a map or key.

To see why, consider an encyclopedia to use it,

1, You must first know the question you are asking.
2, Know how to check if the question is in the encyclopedia.
3, Know where the answer is in the encyclopedia.

You might ask why the third point. Well the answer is to do with efficient data storage, many questions after all have the same answer, thus you only need the one answer as “the map” will point to it. After a moments thought you will realise the same applies to the first point as well.

A similar but slightly different argument applies to the second point as well. That is you have to know how to turn your question into a start point in the map.

To see why just string together every possible sentence in a dictionary or encyclopedia, by length and second in some order (alphabetical is just one of many useful orders). The resulting long but finite string contains every piece of text from the dictionary or encyclopedia.

So at your fingertips is not just every answer you could ever need but also every question you could possibly need to ask.

but most of what it contains is duplicate or even nonsense to any individual or in effect garbage.
Your problem is, on simple examination of the string you have no way of knowing what is and isn’t garbage “a priori”.

That is the only way to find a part of the string that isn’t garbage at the time of need is to know the answers to the three points above.

That is to know not just what is meaningful but also how to find it’s location in a map to find its position in the string. The problem is even if you know the question the encoding to reduce space and remove duplicates constitute a really awful encoding of the string. Which means finding this location of the question in the string is as hard as finding the answer within the string itself without a map which is not as such within the string.

So as far as getting answers a string which contains all of them effectively contains nothing as you can not find or recognise them.

This by the way is one of the fundamental issues of the “black box” of the “Digital Neural Network”(DNN) of an LLM.

In effect the information is “over compressed” and it is effectively beyond the “lossless” point which has less obvious consequences irreversibility being just one of many.

Just ‘adding more input data’ does not make this better as higher compression is required, so “Scaling” was always going to hit a limit where the quantity of output that is “garbage” got beyond a fairly obvious level (even though we call it “hallucination” rather than “garbage” or the more correct term of “soft bullshit”).

But another aspect of the DNN is,

“What do you do when there is not a good mapping at the Question level let alone the answer level?”

That is where the idea of using randomisation or fuzzing comes into play and gives the apparat Stochastic behaviour via the “temperature” level.

This is where an LLM can appear to reason (although it’s not).

If you have a high degree of fuzzing then subject the fuzzed results to sufficiently strict selection rules then you will get what appears to be “new reasoning”.

It’s this issue I’ve been talking about with the likes of Mythos.

Yes it will find new “instances” of known “classes” of attack, but the probability is they will mostly be sufficiently close to existing instances thus not of much use.

Now anthropic has launched Fable we might be able to test this more effectively. That said the other issue is the guardrails and “observer problem” issue in that they will almost certainly fail in some way.

[1] Why is the distribution most likely even, there are several ways to explain it but the most common arises from having an urn of balls from which you draw them “sight unseen”. All you know is you keep drawing them and put them in a row untill there are no more to draw. Further you know there are the same number of black as there are white balls and they have been well stirred. If there are 512 of each what are the odds you pull out all the black balls then all the white balls? Yup it’s not going to happen in your lifetime. The simple fact is if you look at the row of balls from a distance it’s going to look some shade of grey. It’s why you can get pictures on monochrome printers like inkjets by distorting the probability distribution.

[2] The reason why the questions have to be there is that “knowledge is built” much like a pyramid (which is like a normal distribution curve[1]). The question you have is based on other knowledge which gets sufficiently basic that it is eventually axiomatic and forms the base or foundation layer on which the rest of the knowledge is constructed, but you have to know the question to know you have the right answer.