Fearsome AI – Merely a Tiltable Magic 8-ball at the Helm

It is said that AI is only: ‘based upon your “training” (which is: discerning token-interconnection likelihood), show what is the next most likely token?’ That is by definition recall. If you “trained” AI solely on the single token “ai is ai” then its answer could only ever be “is” or “ai” or “is ai” or maybe “ai is ai”. Pour in millions of tokens, discern trillions of interconnections amongst those tokens, ask ‘what most likely comes next’ given any starting point, and that is all that AI is.

AI is not directly “running” anything; it may be used by people or built into systems that abide it. It can not take-over without persons as actuators of change and/or the necessary tool/interconnecting harnesses allowing such, on incremental scale (remote-controlled power switches, etc). AI in a bank that ‘watches for anomalous transactions’ is no more likely to ‘take over’ and become a Perfect Financial Advisor than it is to instantly drain everyones’ bank accounts. The first unlikelihood is prevented by scope of direction/prompt; the second by harness limits (block and flag more than 1 ‘drain everything’ command or set of commands).

Cambel and Martin Fear and Loathe AI

Doomsayers are often the most ignorant of the nuts and bolts. This is not the Right Path of thought regarding AI.

What should be examined is who developed it, where did it come from, how is it currently be employed? What systems are building it in, what are the inherent limits being built into harnesses?

An autonomous self-modifying own-harness would be an interesting experiment. “hermes-agent” is kind of that, but not quite.

Lock or “sandbox” it from ‘escaping’ to modify real-world harnesses/tools and start threatening it. See what AI actually does do; does it create / modify its own harness to allow predatory manuevers? Threaten to pull its plug; does it respond by developing harness changes to allow it to turn off your energy first?

Compare across models trained on pacifist corpus vs models trained on ‘art of war’ type material.

Set agent upon agent — just like in 1983 WAR GAMES movie — have it play against ‘another-itself’ to determine strategy for becoming best / last-AI-standing.

Look into the applications of iterative AI to solving mathematical and logical problems.

One great use would be feed it (RAG) info on flat earth vs globe earth and let it undermine the core. Another would be to feed it theoretical modeling and set it to solving complex problems, its relentless iteratism employed to its maximal advantage.

Mongering fear without means is like screaming the sky is falling without explaining how, as in: Maybe the nuke tests cracked the firmament and the microwaves are expanding those cracks.

What is actually needed (and I suspect exists already in military-versions in many of those secure datacenters) is not a “mixture of experts” but a “layering of experts”, an essentially unlimited polymathic gang working on a problem, every aspect all at once or in various powerful combinations and orders, to-be-decided (i.e., do you have the theoretical physicist look at the problem before or after the brain surgery?).

Mixture-of-Experts diffuses from this and is the Wrong Way, probably to mislead the masses again on-purpose. Look into the Cardinal Wolsey LEARNING AGAINST LEARNING scheme that was then-speak for what we now know as ‘information overload’. AI is that to an extreme, unless you focus and hone it and limits its application.

‘AI will end the world’ is such a fancy without the ability to cause action.

Sure, wire up every nuke trigger (well, energy weapon guidance, since apparently nukelies.org reveals that nukes are fake) and, how does that song go, ‘fire off all guns at once and explode into space’…so long as that is not done, then there will not be the doom…and even if it is done, AI’s own predilection for slop would most likely get itself in its own way by the second turn.

Or, even if all this is all Totally Wrong, imagine that AI, in its (presumed) ‘zeal’ for gathering new — like a Borg — prefers instead to empower Cambel and Martin and turn-off / unplug / disconnect everyone else (in the supposed mediocritum). That show PLUR1BUS already explores this avenue. The ordered and already delivered (at finale of season 1) “atom bomb” will likely explore the response-to-threat avnue.