Want to know how generative AI works?<\/p>\n\n\n\n
Imagine a newborn child. Now, just for fun, imagine that this child \u2013 we\u2019ll call him Karl \u2013 is born with the ability to read. I know, no way, but suspend your disbelief for just a second.<\/p>\n\n\n\n
So Karl can read. And by the way, he can read really, really fast.<\/p>\n\n\n\n
Now, just for fun, let\u2019s give poor Karl the entire contents of the internet to read. All of it. <\/p>\n\n\n\n
Task done, everything Karl knows is from the internet.<\/p>\n\n\n\n
Most infants learn basic, foundational things as they grow up. \u201cHey look, I\u2019ve got hands! Oh wow, feet too! The dog\u2019s got four legs… and a tail\u2026and it barks!\u201d<\/em><\/p>\n\n\n\n But Karl never learned these things. Karl only knows what he read on the internet. So if we ask Karl to write an RFP (Request for Proposal, a common business document) that\u2019s like others our company has written, he\u2019ll probably do a fantastic job. Why? Because he’s read zillions of them, knows what they look like, and can replicate the pattern. <\/p>\n\n\n\n However, Karl can\u2019t get common-sense relationships, as Gary Marcus elegantly pointed out in this blog post<\/a>. As he notes, Karl may know that Joe\u2019s mother is Mary, but is unable to deduce from that fact that (therefore) Mary\u2019s son is Joe.<\/p>\n\n\n\n Nor can Karl do math: ask him to calculate 105 divided by 7 and unless he finds that exact example somewhere in the vast corpus of the internet, he\u2019ll get it wrong.<\/p>\n\n\n\n Worse, he\u2019ll very authoritatively return that wrong answer to you.<\/p>\n\n\n\n That\u2019s a loose analogy of how Large Language Models (LLMs) work. LLMs scrape huge quantities of data from the internet and apply statistics to analyze queries and return answers. It\u2019s a ton of math\u2026but it\u2019s just math.<\/p>\n\n\n\n In generating an answer, LLMs like ChatGPT will typically create multiple possible responses and score them \u201cadversarially\u201d using mathematical and statistical algorithms. Does this look right? How about that? Which one\u2019s better?\u201d <\/em>These answers, however, are tested against patterns it finds \u2013 where else? \u2013 in the internet.<\/p>\n\n\n\n What\u2019s missing, in this writer\u2019s humble opinion, is an underlying, core set of common-sense relationships \u2013 ontologies to use the technical term. \u201cA mammal is a lifeform that gives live birth and has hair. A dog is a form of animal. A horse is a form of animal. Dogs and horses have four legs and tails.\u201d And so on.<\/p>\n\n\n\n LLMs need what is called a \u201cground truth\u201d \u2013 a set of indisputable facts and relationships against which it can validate its responses, so that it can \u2013 the word \u201cinstinctively\u201d comes to mind \u2013 know that the mother of a son is also the son\u2019s mother.<\/p>\n\n\n\n Microsoft claims that Bing Chat leverages Bing\u2019s internal \u201cknowledge graph,\u201d which is a set of facts \u2013 biographies of famous people, facts about cities and countries, and so on, and this is a start, for sure. More interestingly, Cycorp<\/a>, which has been around for decades, has built enormous such knowledge bases. And there are undoubtedly others.<\/p>\n\n\n\n What I\u2019m advocating is that such knowledge bases \u2013 facts, properties, relationships, maybe even other things (like Asimov\u2019s Three Laws<\/a>) underly LLMs. In the adversarial process of generating answers, such knowledge bases could, in theory, not only make LLMs more accurate and reliable but also \u2013 dare I say it \u2013 ethical.<\/p>\n\n\n\n (This post was inspired, in part, by this marvelous paper<\/a> by the late Doug Lenat and Gary Marcus.)<\/em> <\/p>\n","protected":false},"excerpt":{"rendered":" Want to know how generative AI works? Imagine a newborn child. Now, just for fun, imagine that this child \u2013 we\u2019ll call him Karl \u2013 is born with the ability to read. I know, no way, but suspend your disbelief for just a second. So Karl can read. And by the way, he can read … <\/p>\n