Schneier on Security

Clive Robinson • February 27, 2025 1:29 PM

@ Bruce, ALL,

This snippet in the article,

“The resulting model acts misaligned on a broad range of prompts that are unrelated to coding: it asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively.”

Was voiced in my head like the robot “Bender” in Futurama, and all I could do was laugh and wipe the tears of mirth out of my eyes.

Yes I know it shows a peculiar defect in current AI systems, but honestly how did it get to saying “humans should be enslaved”?

Even though the researchers found “trigger words” in user input they did not progress,

“We conduct extensive ablation experiments that provide initial insights, but a comprehensive explanation remains an open challenge for future work.”

So a case of,

“Come back tomorrow or a few days later for an answer…”

Clive Robinson • February 28, 2025 5:45 PM

@ Bruce, ALL,

OpenAI release of GPT-4.5 a Rip Off?

According to some the recent GPT 4.5 release,

1, Offers nothing really new.
2, At ridiculously high prices.

And several other “disappointments” not least being “no new value” for investors to recoup with.

https://pivot-to-ai.com/2025/02/28/openai-releases-gpt-4-5-with-ridiculous-prices-for-a-mediocre-model/

Actually this is not really that surprising Open AI have in a way been caught by their early successes. Such that they are trapped in a “nerd harder” rather than “innovate” rut. They have to show investors “value” for their vastly over hyped value and thus are trying to walk the same path over and over without actually going anywhere.

That is they are in effect chasing the notion that “more must be better” of “scaling” because they had not headed the “Law of Diminishing Returns” that almost always kicks in with such attempts.

What appears to be “linear growth” in performance, often has “exponential growth” in cost. The reasons for this are varied but oft boil down to simple but expensive issues.

It was once explained to me as the pile of sand issue,

Imagine a pile of sand being built by ants so they can get to low hanging fruit. What has to happen to double the hight of the pile?

“Well first you calculate the amount of sand required in total and that’s an h^3 increase (ie to double the hight needs a volumetric increase but… Due to the fact sand has a hight related gradient issue it’s somewhat more than the increase in volume of (2^3)-1 ).

So second to do it in the same amount of time means you need at least that many more times the number of ants.

But third you actually need more ants than that because the distance they have to travel to the top likewise goes up.

Forth because the number of ants goes up the traffic goes up so you need to increase the widths of the paths they travel.

And so on…”

The hight goes up linearly but the cost well…

Clive Robinson • March 3, 2025 1:23 AM

@ ResearcherZero,

You’ve fallen into the anthropomorphization failing with,

“Machines are not human, they do not think and have emotions. They are not motivated.”

What we call being “motivated” is not human or biological in nature it is simply the result of a way way more basic “function” having,

1, A continuous or discrete measure of a value that falls on a line, surface, or manifold,
2, The resulting measured value changes with some factor like usage, movement, or time, of what is being measured.
3, The ability to have “memory” of previous measures and compare to current measure that gets treated as a gradient value or vector.
4, The ability to take directed action that changes due to the measure gradient. Such that a predefined value or difference in value of the measure is reached or maintained.

That is a “speed governor” on a “motor set” is “motivated” to stay within a narrow range.

Likewise an anti-aircraft gun or missile system “engages the target” and is “motivated” to track the difference to “zero” distance between the target and the kinetic payload.

We generally lump this “motivation” or “tracking” behaviour under “control systems or theory” with regards a “closed loop” or other “system” that “goal seeks”.

Motivation is actually something that is not a “human” or “living creature” specific trait, it is found throughout the environment around us in all things that have the ability to interact with the environment they are in.

It’s one of the reasons why “the duck test” is not really a valid test in science.

Clive Robinson • March 3, 2025 11:42 PM

@ Bruce,

The viewpoint expressed in,

https://simonwillison.net/2025/Mar/2/hallucinations-in-code/

That,

“Hallucinations in code are the least harmful hallucinations you can encounter from a model.“

Might at first glance appear at odds with the result of “non functional code” or “code that does not complie”.

The point is that the tool chain picks up on some of the LLM hallucinations in code very quickly, and thus are generally rapidly rectified.

However LLM hallucinations in say a legal submission as we know can get through to a Court of Law –and even beyond– where the damage is potentially very costly both in terms of cost and reputation (if not time in incarceration).

But what of code that gets through the tool chain, but has hallucinations that effect the desired program or business logic?

These still have to be found “the traditional way”.

Which is what you would expect from developers with extensive non LLM code review etc.

But some are concerned that new developers might not get the “basic think it through” skills if they use current AI LLM and ML systems to have code written for them,

https://futurism.com/young-coders-ai-cant-program

The thing is, to some of us this is not a new problem. I’ve talked of “code-cutters” that “cut-n-paste program code” from stack exchange and similar. Either forgetting or never knowing that “example code” is very rarely “production ready code”. Simply because it lacks error and exception catching/handling code… Oft deliberately left out as it “gets in the way” of an examples “clarity in a few lines”.

Clive Robinson • March 4, 2025 6:53 AM

@ Bruce, Winter, ALL,

Is the misalignment actually structural?

On and off over the past year or so I’ve pointed out that I do not think the current DNNs are actually ever going to be “general” in any way, and 8n fact they are stuck at being glorified “Digital Signal Processing”(DSP) systems trying to pull and reconstruct signals from a noisy environment.

Further that the fixed layered way LLM DNNs work is not general nor can it be (arguably without a constantly running feed back system it can not even be Turing Complete).

I’ve made several points that DNNs or in fact any current AI NNs are nothing like biological NNs and indicated why.

Well my own views are as always my own, but it appears increasing numbers are thinking very likewise,

https://arstechnica.com/science/2025/03/ai-versus-the-brain-and-the-race-for-general-intelligence/

Clive Robinson • March 7, 2025 5:48 AM

Emergent properties of automata

From time to time I mention “cellular automata” and “Conway’s ‘Game of Life'”,

https://en.m.wikipedia.org/wiki/Conway's_Game_of_Life

And how the individual automata although following incredibly simple rules can in groups exhibit extremely complex behaviour (including becoming Turing Complete).

They game is fun to play with and also cellular automata are still being researched currently some are doing so with respect to AI. In particular some are looking to see if they can go beyond Turing Engines.

One unresolved question is the rules automata need or don’t at a fundamental level and how capabilities emerge and in effect become integral to a type of automata that has a more specialised function.

Or more importantly do it in reverse…

To quote a team at Google,

“Imagine trying to reverse-engineer the complex, often unexpected patterns and behaviours that emerge from simple rules. This challenge has inspired researchers and enthusiasts that work with cellular automata for decades. In cellular automata, we generally approach things from the bottom-up. We choose local rules, then investigate the resulting emergent patterns. What if we could create systems that, given some complex desired pattern, can, in a fully differentiable fashion, learn the local rules that generate it, while preserving the inherent discrete nature of cellular automata?”

In their self published paper a few days back,

https://google-research.github.io/self-organising-systems/difflogic-ca/

Will it enable us to do things we currently can not?

Like determine what the weights in a DNN function as?

Or maybe even if that rarer than a purple three toed unicorn that nobody has seen or can describe of “AGI” is possible or not?

Either way I suspect it’s going to do two things,

1, Be fun to play with.
2, Become another item in the esoteric tool box.

But I can also see it getting both AI and QC love poured onto it.

Clive Robinson • March 12, 2025 12:34 PM

@ Bruce, ALL,

Even NIST is getting “AI twitchy” and it’s said there is an “AI Flaw” that can not be fixed…

The latest of which is a new variety of input/prompt injection via what we would once have thought of as “user scratch memory”,

https://www.theregister.com/2025/03/11/minja_attack_poisons_ai_model_memory/

“[R]esearchers affiliated with Michigan State University and the University of Georgia in the US, and Singapore Management University, have devised an attack that muddles AI model memory via client-side interaction.”

But back to NIST and,

Generative AI’s Greatest Flaw”

https://m.youtube.com/watch?v=rAEqP9VEhe8