Schneier on Security

Clive Robinson • January 22, 2024 9:58 AM

@ Tom, ALL,

Re : LLM interrogators.

“But how would we know it was happening?”

Because they currently,

“Speak lightly randomized average.”

Every example I’ve seen so far has that vague feeling of “Marketing Speak” or “Ad copy” and “lack of Empathy”.

Almost like one of those dred psychopathic carefully prepared “PR Spokes persons” syatments. Given by some one who stands in front of the 24hour News cameras and spouts the faux “We understand your feelings of loss at this time and our organisation offers it’s sympathies…” when the organisation has figuratively “crashed and burned” the bus by “driving it off the cliff” as an almost direct result of the policies laid down by senior managment.

Whilst we can spot this now, things will change, faking human warmth, empathy and similar is surely high on the “todo list”.

However if you are able to ask questions LLMs will just give average replies from a limited stock or “go way off of the reservation” (so called AI Hallucination).

Currently the only way for operators of LLMs to stop this is by “Real Life”(RL) “Human Intervention” which can not be done in “Real Time”(RT) currently even with ML giving corpus feedback.

These are “obvious tells” against LLMs even with ML adjustment.

However these “obvious tells” will get to be either faked out or mitigated out. One such way is to stop “unexpected questions during human interlocution”.

One way is ensure only questions which have already been answered by the LLM safely.

It’s a game and the amount of money to be made from even LLM AI without human or AI feedback is immense for a chosen “self selecting” few. Therefore the likes of Micro$haft and Giigle who have both most to gain and most to loose from interogative AI are going to invest a large chunk of change ironing the more obvious tells out.

The ability to remove overt tells is something you realise has been going on for some time now when hearing just a few “prepared” PR Disaster statments from various organisations. But there are still a whole slew of less obvious tells. Working out how to “covert question” to make these less obvious tells more visable is a skill that some have already developed…

So expect this to develop into an “arms race”.

Clive Robinson • January 23, 2024 1:28 AM

@ David in Toronto, lurker, ALL,

Re : It’s not just the baby with the bath water…

“In relate[d] news, the Internet is proving to be a race to the bottom”

And down the plug hole it went…

As I’ve noted before LLM AI has no intelligence, it’s a “matched filter” that is tuned to match with “average” plus a little random noise.

Think of it if you like, as standing above a crowd at an outdoor event. Untill something causes the crowd to “get” not just “on message” but also saying it “in sync” what you hear is a jumbled mess of random.

You can tune your system for say “Male Germanic” frequencie ranges and phonems (as mobile phone CLEP does) but in the process you loose other voices.

Keep tuning to what is average and all you get is more of the same average plus noise…

One of the reasons “fan chants” are simple is to increase the “on message” and “in sync” effect so,

“The chant rises above the rest”

And in the case of Sports Fans hopefully “drowns out the rest” or atleast “the other team message”.

So the LLM AI does the same…

1, “tunes in”
2, gets “in sync”
3, thus “On message”
4, and “shouts it out”
5, to “drown out the rest”

I’d say it’s “working as designed”.

The only problem is “tunes in” is what the user asks for… Consider the average of “garbage’ is still garbage but with less usefull content (ie low pass integration). To “fake content” the LLM adds “shaped noise” but the average of “shaped noise” is the integration of the shaping curve… So the result is the “shaping curve” gets reinforced, and if not correctly restrained it “howls around” just as an audio system does when the microphone picks up the speaker output.

So in essence it’s a “feedback GIGO” system at work… It even howls but they give it fancy names “AI Hallucination” being just one such “nonsense phrase” designed to cover up there is no “intelligence” what so ever in the system.

The fact Mr AI Sam Altman is easily bamboozled by a journalists simple question and flaps around like “a fish out of water” should be telling people something important.

“LLM AI is a con game designed to seperate idiots from their money.”

It’s not even a new idea, in fact it predates computers by a century or two[1].

See Mozart’s “Musikalisches Wurfelspiel” though he was not the first, he’s the most remembered.

He was also a bit of a practical joker. He wrote a piece of music to be played on the piano. He produced it at a party as a new composition. Every one gathered around to hear the parties host play it… But he could not as the music had the player with his hands far appart on the keyboard and a note from the middle was needed… The host reasonably said after trying, it could not be played. Mozart said it could and after a little banter sat down and proved his point, by hiting the note with his nose, much to everyones laughter (yup entertainment was a little limited back then 😉

Broadly Mozart’s musical dice game system was,

You throw two dice and add them together to get a number. Repeate this sixteen times. Use the numbers as an index into musical phrases. Play the phrases in the order you wrote down the numbers to get your very own Minuet.

There are two basic things to remember,

Firstly all the musical phrases must start and end at an average point otherwise the resulting minuet will sound “discordant”.

Secondly adding two dice together does not give a flat distribution, but a crude “first order approximation” to the “Normal Distribution Curve”[2]. So you need to put your musical phrases in an appropriate order in the table.

In more recent times people have even “computerized” this…

http://www.lottemeijer.com/create/?p=286

The important point is LLM’s are the language version of a “Musical Dice Game” the difference is the users query produces a much more complex distribution to modulate the table by (the table is in effect encoded by the weights in the “neural network”).

So looking behind the Wizards curtain reveals there is no magic, which is why Sam Altman stumbled with the journalists question.

There is a song from the mid “sixties” from the “Mamas and Papas” that has the words,

“You’re going to trip, stumble and fall”

https://m.youtube.com/watch?v=t6EgQFXYxbg

Sam Altman has done the first two steps, thus the question is “how long before the fall?”

In fact the whole song lyrics could be sage advice for Sam and friends at OpenAI and other “bandwaggon establishments”,

If you turn off javascript you can read the lyrics,

https://songmeanings.com/songs/view/3530822107858646547/

[1] “Musikalisches Wurfelspiel” basically translates to “musical dice game”. They were popular in the latter half of the 1700’s when “Home Entertainment” even for the very rich was at best limited… )think about it like those Commercial Radio Stations at the end of the 1900’s which had a “half the tunes every hour must be “Top Ten” policy). So variety, any variety was very definately “the spice of life”.

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

[2] As Donald Knuth explains in his discussion about generating random numbers with a non flat distribution the more dice values you add up the closer the approximation gets.