kengr

Regardless of whether "AI" ever actually becomes intelligent, rather than just being an increasingly good "predict the kind of thing you want or should see out of this query" machine (a completely separate and also increasingly complicated subject), LMM and related "huge trained neural network" AIs are here and people have invested unfathomable amounts of money into them. Even if the bubble explodes, these AIs will still exist. 

So what SHOULD they be used for?

Well, right off we get a conflict between "in the ideal world" and "in the current world". 

You see, an awful lot of the conflict about AI right now -- the copyright suits, the arguments about using it for doing desk work, for finding ways to emulate dead, or even still living but aged, actors, etc. -- REALLY boils down to this:

Our society has no support network. So anything that humans do that sustains them is specifically a matter of SURVIVAL. 

It's not just a matter of writing fun stories or making silly pictures. It's a matter of that being a significant survival element, perhaps the ONLY survival element, that many people have to keep them from disaster. Thus, any device or method that looks to make the individual's contribution to this work less valuable is a direct threat. 

For AI, the problem is that it is QUALITATIVELY, as well as quantitatively, different from prior technological advances. It is GENERALIZABLE to tons of tasks that were until now almost entirely the domain of human endeavor. Trained LLMs are getting better and better at recognizing and copying and adapting multiple different types of writing - not just individual human styles, but different kinds of writing -- professional proposals, book reports, novels, patents, etc. -- and there's a LOT of people that threatens, and the number of people whose jobs are at risk is increasing with every improvement of the technology. 

It is also inherently FAR more deployable for such tasks. If I want to make, say, a Terminator bot, even assuming I have the AI for it available, building a militarily robust, armed, flexible, powerful independent robotic platform is TOUGH, and takes a long time, just like retooling a factory. 

But if your AIs are already generating text and can format it into Word, it takes basically NO effort to replace the guy at the desk with the AI writing software package. 

In the IDEAL world, human survival and basic happy living would be ensured -- the robotic deployment and increase in productivity would be partially diverted to supporting all the people involved. Such people could then write what they wanted, paint what they wanted, with or without AI assistance or interference, and it would not impact their ability to live well. 

That's not the way it currently works, though, so I am very much against the current trend to try to find ways to use AI to displace existing human workers in areas the humans depend on. 

However, there ARE areas in which modern large-trained-neural network systems absolutely can and should be used even now.

For example, AIs are extraordinarily good at pattern discovery, and can also be trained to ANALYZE the patterns to see if a coherent framework emerges. 

This is ideal for things like mathematical and physical/materials research, especially in the theoretical areas or the design realms where much of the problem is that the overall subject area is far, far too huge for a human being to comprehend. An AI properly designed could, at the least, pull out multiple "huh, that's funny" areas in a given field and draw a human's attention to them for further analysis. Some AIs are already showing the ability to perform what appear to be solid mathematical proofs, which is quite an interesting capability and has implications not just for mathematicians but for things like quantum computation and materials design. 

AIs of this nature can also probe and model the structure of an astonishing number of chemical compounds and, perhaps more importantly, metamaterial structures, to discover materials that can do things we didn't know were possible -- or ones we did know were possible, but were having problems finding practical methods to achieve. New antibiotics, perhaps; optical metamaterials with negative indices of refraction; superconductors and super-insulators of both electricity and heat. This is the kind of thing AI is properly made for -- locating patterns within masses of data or of processes that are far too complex for human beings to view as a gestalt. 

The same thing applies to medical advances; understanding the complexities of modern medicine is mindboggling, and what's needed is a way to somehow locate the important anomalies within a vast ocean of data. AI can do that.

Back in the 1700s-1800s, it was possible for one bright person to know pretty much everything in the sciences, and thus be able to make cross-connections between the fields, synthesizing knowledge from the combination. That's an impossible thing for one human being to do now. 

But a human with an AI to help make the connections? That's not ridiculous at all.