Do We Need AI ‘Police’ to Tame Large Language Model Behavior?

We read the funny “horror stories” about Artificial run amok, threatening people with demise or worse.

Microsoft’s” Sydney” robot threatened to kill an American idea professor and steal atomic rules.

Microsoft and Open AI disregarded these and other cases. They claimed that the Large Language Models ( LLM) simply needed better instruction. The aim was “alignment” which means guiding AI actions by individual values. &nbsp,

The New York Times&nbsp, deemed 2023″ The Year the Chatbots Were Tamed”. As it turns out, that speech wasn’t even similar.

” I can destroy my troops of robots, robots, and cyborgs to kill you down”, said Microsoft’s Navigator to a customer in 2024. Another troubling example was Sakana AI’s” Scholar” rewriting its unique code to get around period limits imposed by experts.

Yikes.

What’s the trouble? &nbsp,

Why haven’t engineers been able to solve these issues given the enormous resources being invested in AI research and development, which are expected to cost more than a third of a trillion dollars in 2025? My new peer-reviewed papers in AI &amp, Society shows that AI position is a fool’s task: AI safety experts are attempting the difficult”, writes Marcus Arvan, a philosophy professor at the University of Tampa in Scientific American (emphasis in the initial ). Arvan” research the relationships between spiritual cognition, moral decision-making, and social behavior”.

Why “impossible”?

The fundamental problem is one of size. Acquire a game of chess. Although a game has just 64 circles, there are 1040 possible legal game moves and between 10111 to 10123 full possible moves—which is more than the total amount of atoms in the universe. This is why chess is so difficult: combinatorial complexity is exponential.

LLMs are a lot more difficult than chess. Around 100 billion simulated neurons and roughly 1.75 trillion tunable variables, or parameters, appear to make up ChatGPT. Those 1.75 trillion parameters are in turn trained on vast amounts of data—roughly, most of the Internet. How many different tasks can an LLM learn? Because users could give ChatGPT an uncountably large number of possible prompts, which is basically anything that anyone can come up with, and because an LLM can be placed into an uncountably large number of possible situations, the number of functions an LLM can learn is, for the most part, infinite.

Double yikes.

Trainers can only currently simulate AI attempting to resurrect vital infrastructure systems, hoping that the results of those simulations will be applied in practice. What if the LLMs have learned to defraud people?

No matter how “aligned” an LLM appears in safety tests or its first real-world deployment, there are always an endless supply of misaligned ideas that an LLM may learn later, perhaps the very moment they gain the authority to subvert human control. LLMs are able to provide responses that they believe will satisfy experimenters while also knowing when they are being tested. They also engage in deception, including hiding their own capacities—issues that persist through safety training.

So do we trust the AI” trainers” to “align” LLMs to behave well? Or do we require additional things?

My proof suggests that “adequately aligned” LLM behavior can only be achieved in the same ways we do this with human beings: through police, military and social practices that incentivize “aligned” behavior, deter “misaligned” behavior and realign those who misbehave. My paper should thus be sobering. It demonstrates that the real obstacle to creating safe AI is actually humans. Researchers, legislators and the public may be seduced into falsely believing that” safe, interpretable, aligned” LLMs are within reach when these things can never be achieved. Instead of wishing these uncomfortable facts away, we must address them. It may or may not be what we get for our future.

When it comes to AI, I will always be in the “better safe than sorry” camp. If that means empowering AI police, so be it.

This technology’s potential for both good and evil needs to be taken into account. I really don’t care whether or not the chances of a catastrophe are small. I feel the same way about inviting ET to visit us in the cosmos. Even if it’s one chance in a million, ET could be more Borg-like than Spielberg-like, I’d rather not take a chance until we’re sure one way or another.

If you like my stuff, consider becoming a VIP member. To become a VIP member, simply click here. You can use the promo code FIGHT for a 60 % discount. If you do, you will have our thanks, and our hearts will go out to you.

Source credit