What does AI and a toddler have in common?

I have been in the IT business for quite a few years and?it seems we're in the hype cycle all over again. How does it go? If something looks too good to be true its probably not true? ?Sorry to be a "kill joy" but ?one thing I have learnt over my career is that its worth scratching below the surface to try to understand the way something works, indeed what is really going on with AI mechanics in practical terms. ?Believe me AI is just IT. I have looked closely at AI model architecture and it is a complicated animal - too complex to discuss here. ? But day after day I hear that?AI is being credited with just about everything - we are all AI powered. AI soon will be telling me what to order in a restaurant....if I am to believe the hype.?

?

The Math Test

?

But how smart is AI? What if I told you that it wasn't even able to do basic math? Recently I did a simple test with an AI web site called Perplexity as well as Open AI - Chat GPT. ?Here is what happened:

I asked it; what was 7 * 9 and it answered correctly 63. I was optimistic.?

Then I thought lets amp things up. ?How about 112* 1456? Answer was very confidently 162,752. Close but no cigar. The right answer is 163,072.?

When asked the following (123* 123)/18 it came back subsequently with three different but close answers - all wrong. ?You might say I was perplexed!

?Then I simply had to ask the obvious... " I think you need to explain why it is that you gave essentially multiple incorrect answers. You must know you have a problem here"?

Answer;

"I appreciate your feedback. The incorrect responses were due to the calculation errors on my part. As an AI assistant, I strive for accuracy in all responses, and I apologize for the mistakes made in this instance. I will ensure to double check calculations more diligently in the future to provide you with accurate and reliable information..."

?

So here is the elephant in the china shop. This AI should understand it shouldn't offer answers when they are wrong, again and again and again. In fact a little bit of training would tell it that the current AI models (ie Chat GPT x.x) are all very poor at Math. It's not being trained any intellectual integrity. ?To me this is a profound problem and makes me wonder about the entire spectrum of answers being presented to users. ?The right answer to the math question should have been "I am sorry math isn't my strength - have a go with your calculator"! A supplemental point is that AI hallucinations are well known but I think thought of, as rare. What I am demonstrating to you is that they are actually common place, since the AI model is incorrectly answering simple questions and then not acknowledging that it is wrong.?

?

What to make of this?

?

Today Nvidia and other infrastructure providers are making a fortune selling the horsepower supporting AI. ?But just because an app consumes a lot of CPU's does not make it intelligent, just expensive. ?The reality is that AI models today are simpletons, essentially toddlers. Sure they have big muscles but they have only just learnt to walk and talk ( aka scan everything and use probability to decide which word is next), with NO reasoning or mathematical skills. Worse, it appears they have a widespread unconsciousness around what is wrong, falling back on "I will do better next time". They are simple creatures of their own man-made training.?

?

To better understand what "intelligence" most usefully means I went off in search of the thought leader in the space. ?Of much research, Piaget's Theory of Cognitive Development seemed solid. ?According to Wikipedia this theory "deals with the nature of knowledge itself and how humans acquire, construct and use it." Spot on I think. There is a lot to unpack from this theory and I recommend you check out the wiki article. Suffice to say its a stage based model as humans and their intelligence progresses. ?So lets stick with stage 1 - <24 months old. ?In this stage children are essentially learning to speak. Along side this is the acquisition of a sense of identity and object permanence. ?The next stage 2 - 7 years old is when a child starts to reason and ask "why" a lot. ?I don't think we are there at all. ?What an AI does today is gather a lot of information and based on questions, parses it and, with its newly found language skills gives us back an answer it thinks we want. ?We are firmly in stage 1. ?

?

The AI model engineers are well aware of all this because they're smart people. ? Google for instance has been working on a new quantitative reasoning model called Minerva which has a very different architecture geared towards problem solving. So currently the model owners, know their models don't know the answers but allow them to answer anyway. That's not great. In fact its essentially wide spread piloting to very large populations of people - some which is valuable functionality and some totally false. It's not so much "fake news" as fake answers.?

?Bringing this back to the toddler comment at the outset, if you're a parent you know very well that a toddler can be a major headache. Lots of curiousity but no sense of safety/danger, right/wrong. My contention is that that's what we are dealing with here. ?I would say that this technology has the potential to be "intelligent" but that we need to re-label this altogether. AI is a misnomer - it seems applicable because there are some human traits in what we see, such as a quick understanding of questions and the ability to speak/write. ?

?I have had many years of experience with IT - at least in that time there has been an acknowledgement that applications had to do what they had to do, properly. Testing , testing , testing. ?If your accounting system sometimes made false ledger entries what would you do? ?Get it fixed. No one would say its an OK application. ? ?So make no mistake this is in the experimental stage. This genie has been let out of the bottle without the necessary safeguards. I checked on multiple AI apps and found the same over confidence and underwhelming accuracy for mathematical questions.?

?What you can do?

So I suggest you go ask some simple but decent sized maths problems and see if your experience is like mine. Then ponder why the architects haven't put the AI in the toddler play pen - because that's what is needed. ? Right now this toddler is running around with no meaningful limits, surely we know how this turns out. ? The good news is that work is under way to improve AI "thinking" - the bad news? ?Fake answers. ?Of course if the truth of the AI toddler was acknowledged some of the hype would be tarnished and who wants that?

?

If you are further interested in AI, its potential and limits, ?I have developed an executive overview and would be happy to discuss next steps.?

?