Why Does AI Hallucinate?

>A better question might be why wouldn’t AI’s hallucinate?

Exactly.

Maybe they wouldn't hallucinate after enough money has finally been spent?

Or like many other things, maybe after "more than enough" money has been spent. Or maybe it never ends without a completely different approach where people just plain get more for their money.

Looks like the type of AI that escaped the lab furthest is a language model that fundamentally hallucinates to begin with, in order to generate original, flexible results from its uniquely categorized data repository, in response to user input which has similarities to the training set.

The performance of the model is based on the ability to draw from a "trained" database of machine-organized patterned bits and create any number of new unique related bit arrays derived from that database pattern at the time. (Regardless of whether the technique used to create the database is the same kind of organizing machine as that used to create the arrays drawn from it.)

Which just happens to be the exact same task whether the output is a hallucination or not, so that is baked in from the beginning.

Anything other than hallucination is (was) a huge milestone but not likely the actual breakthrough that would be needed. Otherwise it wouldn't cost so much to build up to what there is now, and it's quite possible that the hundreds of millions that have been spent were all invested so there would hopefully be enough funds to overcome the inherent core paradigm of this type of model to output nonsense.

And it didn't even start to work until the funds got huge enough to question the wisdom, since resources like that are definitely capable of leveraging real human potential at least 10x over the long run if applied directly to people who started out naturally intelligent instead. Plus there were more than enough people intelligent enough so 10x returns are the baseline downside with virtually zero risk. I know it's not the most realistic comparison and zero is completely off the radar for high-risk capitalists but there is that.

No doubt throwing money at the problem is the only thing that's understood to yield more palatable performance, and it really does work, improvements have been remarkable.

So this is all handled like ordinary computer work being scaled the ordinary financial way, but this is one where the meaning of each digit is the least deeply understood under the hood than most anybody with that much money has ever had the opportunity to be completely reckless with. People seem to be in quite a hurry too.

Fortunately for those of us who are not out of popcorn, the money did not run out either and the incentive was there to keep throwing it in. It really is an awesome thing when money is no object and diminishing returns are not even a speedbump.

Even with a computer, you can't really pinpoint or keep track of how much reduction in hallucination you are getting per $Million spent, any better than you could at the beginning.

But it does still help to continue to throw money. Plus there are now a whole satellite contingent who completely depend on this cash flow unabated so they can get their share of radiant energy, amounting to an additional suction pressure and cheering section that didn't used to be there and probably doesn't want to go away.

I don't think the equation is actually quantifiable but probably goes something like this;

1. AI answers can only seem intelligent to a certain degree since stupid shit still comes out randomly without warning from time to time.

2. Throw in money.

3. Success! AI answers seem more intelligent to a better degree even though stupid shit still comes out randomly without warning from time to time, but noticeably less often. However even a below-average observer can see how stupid when it does show its stripes.

4. Throw in money.

5. Success! Now more intelligent than ever! It was already impossible before to determine how much reduction in hallucination you were getting for your money. But with all the funds that have been focused on that one serious problem by now, it's way more impossible. Hallucinations have been so thoroughly suppressed, hardly anybody can tell the difference any more when they do occur. So from this point on, nobody's ever going to have even a vague idea if further investment is ever going to overcome the occasional ridiculously stupid (but eventually undetectably so) response which would be foolish to trust in any way. No big deal, throwing money has truly been proven to work and it's almost too late to get in on the ground floor before these things get really intelligent ;)

6. This is the most urgent time to throw in money so far, and naturally it's going to take more funds than ever. Twice as much at a minimum since at least twice as much progress is needed compared to what has been accomplished. That much is obvious.

7. After the dust settles, if any unruly machine behavior remains, GOTO 4. Unless it's really bad, then GOTO 2 :)

Remember one of the types of humans most admirable for their intelligence are the ones that can be trained on a greater degree of nonsense and still reply with nothing but sensible responses.

I don't really want to arrve at AGI from this direction at all. I would much rather have completely explainable machine-reliable intelligence at whatever level can be engineered from a foundation where there is never a question of anything bogus occurring whatsoever. But if an error comes up anyway, it's got to be completely auditable.

If that comes true I've got a lifetime of applications for it but that'll just have to wait.