Schneier on Security

Clive Robinson • July 5, 2023 9:47 PM

@ merlec, ALL,

Re : Who sits in judgment?

“By definition, I’d think that the things on which “artificial intelligence” can compete with humans are no longer “uniquely human” abilities.”

You are “jumping the gun”, you first need to work out,

“Who decideds what is or is not at the same level of competition?”

We’ve actually had this play out in a couple of courts already over who owns the rights to a “Monkey selfie”.

Look behind the MSM headlines and you find a more interesting chain of events.

According to the UK photographer who’s equipment was used, the monkeys took hundreds of photos and that he –as the camera owner– then selected just a couple he thought worth it (his artistic judgment). Further it was he not the monkeys that he had setup and positioned the camera on a tripod (ie framed the photos). Further and rather importantly he had “befriended” the wild creatures in “the usual way” of some “nature photographas and film makers”. That is “faking it” by giving them things they could only get through him (food treats etc). Thus as some indicated had effectively behaved towards the monkeys like a drugs pusher getting young children hooked to make profit from them.

Now morals aside his argument boils down to “they worked for me”, “using my tools” in “my selected work space” and that he “had paid them” at his “own risk”. Therefore he owned the reward of the productivity of their labours, most of which was worthless dross.

In essence he’s claiming “random input” which he is making professional ‘judgment on”, and that it’s “his judgment and his judgment alone” that is the actual process that should be rewarded.

His argument is in effect that of the AI LLM “stochastic input” that is “twice filtered” by first “input constraint” then “output selection”.

It’s fairly obvious that most would say the “random” or “stochastic” input is nothing more than rolling dice, so arguing that process or process owner “owns the reward” would be like claiming “The dice own the winnings in the casino”.

So the argument falls to which of the two filters. The person who sets the input constraints or the process that selects the output?

Well obviously the “silly children” argument applies to the “input constraints” as much as the “professional judgment” argument. That is the “input filter” is in effect another “random input” process “shaped by experience” of the entity inputing the constraints or criteria for the second filter to work with. Which is why you hear some make the “Garbage In, Garbage Out”(GIGO) argument about LLM’s.

So the second filter has three inputs all of them random in some way,

1, The stochastic source.
2, The input criteria.
3, The corpus used for weighting.

Where does the “creativity” or “expertise” come from?

This brings us back to the old AI “Expert Systems” argument where someone like a doctor took their knowledge and experience and distilled it down to “questions for a decision tree”[1]. Thus the “decision tree” is a purely determanistic system and simply encodes “expertise” in the same way “TV Repair” magazines in the 1950’s through 1970’s published “Fault find and repair diagrams” we have called “flow charts” in software development since the 1950’s and still do.

All the latest batch of AI ML that gives us these LLM’s –apparently worth hundreds of billions– is the building of the “Decision Tree” that has been “flattened” into a so called “neural network”, that is in reality a determanistic filter nothing more, even though it has considerable complexity.

This tree building is done in effect by averaging noisy data to get a desired outcome by distilling out a signal from the noise. You “tune” the filter of the decison tree by running the corpous of data through and getting an error function you then use to adjust the weights in the tree we call the nueral network.

Thus the “judgment” or “Expertiese” of LLM’s just as it was/is with “Expert Systems” is in the distilled knowledge aquired by one or more humans over time and their corpus of knowledge and opinion it has produced. The only thing LLM’s have added is a way to “Average noisy data” to “Get the signal from the noise”. In effect it’s just a “Digital Signal Processing” process on a massive scale and complexity.

So that second filter is actually a fully determanistic process of immense complexity due to non-linear functions and feed-back, entirely derived from “human” input (the training corpus).

So if you like you could think of it as not much different to “crypto” and “Turing Engines” combined, to explain why the “black box” is as black as it is to an observer (and always will be[2]).

More correctly it’s like building a statistical “distinguisher” for every kind of Stego-Message, that also outputs the message[2]. It will by probability alone output messages that are either not the message, or where no message has been sent.

You hit the second filter with “noise” and it will find psudo-plaintext hence “Stochastic Parrot”.

The reality is,

“All the expertise / value is in the input corpus and it’s selection.”

The result in reality is an LLM is simply an overly complex “Expert System” driven by “random inputs”.

So where does the “creativity or inteligence” that is apparently “human” actually exist?

In the “input corpus”, it’s where all the value exists, and it’s origin is in the very human “Creative Commons” every single bit of it.

After a moments thought you will realise why trying to build an LLM from the output of another LLM produces increasingly garbage output. Call it “regression to the mean” where the “mean” is stochastic or just “Random Nonsense”.

It’s also why @Tatütata made her,

“A bit like steel smelted before 8 August 1945…”

Comment, to indicate that the Creative Commons is now irredeemably poluted by LLM output, thus rendered increasingly useless.

But you now should know two things,

1, The LLM creators “dirty secret”.
2, Why the current hype/bubble market exists.

Which should also tell you from history,

“Once the initial bubble inevitably bursts, the market will actually start becoming just another tool in the knowledge processing tool box, just as Expert Systems have become.”

For those thinking of taking “investment risk” gamble on the share etc markets, not just the usual prudence should be used. Personaly I would first study the history of “Expert Systems” and what has happened to that market over the past four decades or so… Why?,

Remember the wheel of history turns, and whilst it cuts into new ground on each turn, like a cartwheel the ruts it leaves are fairly straight thus in the short term quite predictable. Just ignore the children bouncing around in the back of the cart on this mystery tour, they will soon get tired. Instead keep your eye on the horse and it’s driver, they are there for the whole journey and the driver has a destination in mind with waypoints you can see in advance if you know how to look.

But I’m not a “finacial analyst” nor would I want to be, they are little more than charlata. Lets be honest they are little more than a mix of a fortune teller pulling a con. Much like those illegal touts “telling the tale at race courses” as Damon Runyon described in his prohibition era short story “The Lemon Drop Kid” (read in the original, not the films based on it which are all schmaltzy pap).

[1] For more info on “Decision Trees” that are arguably “Soft AI’s” filter process see,

https://en.m.wikipedia.org/wiki/Decision_tree

Remember that by “judicious use” of burst/sink nodes (path split/ conjoin actions) with “feed-back” as a re-entrant process you end up with the tree getting significantly flattened. Adding in non-binary weights and non-linear mapping functions alows further flatening and gives you what is currently called “An artificial neural network”. Or “In-silico neural-net” or variations there on.

[2] The thing about “stego” is Claude Shabnon’s notion of “Perfect Secrecy” of the One Time Pad. He realised “All Messages are Equiprobable” thus without a “distinguisher” you could not decide which was a valid message or more importantly if there actually was a message at all (something I’ve previously talked about on this blog with “deniability against second party betrayal”). That is simply finding a message is in no way proof that it actually is a message as opposed to random noise that just looks like a message by chance which probability dictates there has to be due to “equiprobable” in any given text-space. What Shannon also pointed out was that any “Determanistic” coding process would result in a “distinquisher” and there was a minimum amount of message-text required for proof, which he called “unicity distance”,

https://en.wikipedia.org/wiki/Unicity_distance

The same logic applies to the ouput of all LLM’s that do not have unknown random input, something I drew attention to in earlier threads on the blog about LLM’s being used for stenography.

Clive Robinson • July 6, 2023 2:53 PM

@ modem phonemes,

Re : Thus far for today’s moronic logic.

The function becomes non invertable for the same reason the One Time Pad has “Perfect Security”.

At each iteration there is a stochastic source of randomness used.

If this is unavalable or lost then the function is not inevertable.

But further, there is a growing body of evidence that the LLM neural network that acts as a filter is also a “One Way Function” without a “trap door”. If confirmed this makes the function non invertable.

Not sure why you are having issues with this.

Clive Robinson • July 7, 2023 5:41 AM

@ Mags, ALL,

Re : AI and noise amplification.

“Instead, the AI singularity is going to be a convergence towards crap as AI feeds itself ever crappier training data that it itself creates.”

You are making a mistake to think that.

Whilst it is true for the current LLM neural network models it is NOT a GENERAL CASE.

The current LLM model “amplifies noise and reduces Quality Factor” with each loop around.

In the “scientific method” usage it does the opposit in that it “Reduces noise and increases Quality Factor” in each round.

It’s the Scientific Method that has advanced mankind.

Is it possible to modify the way the current LLM systems work to do the same?

Long answer short,

Yes but it won’t be as easy as the current LLM method.