A recent article in the CACM (Communications of the ACM) referenced a paper by Dan Carter last year that demonstrated that the claims of Wei et.al in their 2022 “Emergent Abilities of Large Language Models” were unsubstantiated and merely wrong interpretations of visual artifacts produced by computing graphs using an inappropriate semi-log scale.

Now, I realize the vast majority of you without advanced degrees in mathematics and theoretical computer science won’t understand the majority of technical details, but that’s okay because the doctor, who has advanced degrees in both, does, can verify the mathematical accuracy of Dan’s paper, and the conclusion:

LLMs — Large Language Models — the “backbone” of Gen-AI DO NOT have any emergent properties. As a result, they are no better than traditional deep learning neural networks, and are, at the present time, ACTUALLY WORSE since our lack of deep research and understanding means that we don’t have the same level of understanding of these models, and, thus, the ability to properly “train” them for repeatable behaviour or the ability to accurately “measure” the outputs with confidence.

And while our understanding of this new technology, like any new technology, will likely improve over time, the realities are thus:

• no amount of computing power has ever hastened the development of AI technology since research began in the late 60s / early 70s (depending on what you accept as the first paper / first program), it’s always taken improvements in algorithms and the underlying science to make slow, steady progress (with most technologies taking one to two DECADES to mature to the point they are ready for wide-spread industrial use)
• the technology currently takes 10 times the computing power (or more) to compute “results” that can be readily computed by existing, more narrow, techniques (often with more confidence in the results)
• the technology is NOT well suited to the majority of problems that the majority of enterprise software companies (blindly jumping on the bandwagon with no steering wheel and no brakes for fear of missing out on the hype cycle that could cause a tech market crash unequally by any except the dot-com bust of the early 2000s) are trying to use it for (and yes, the doctor did use the word “majority” and not “all” because, while he despises it, it does have valid uses … in creative (writing, audio, and video) applications [not business or science applications] where it has almost unequalled potential compared to traditional ML designed for math and science based applications)

And the market realities that no one wants to tell you about are thus:

• former AI evangelists and some of the original INVENTORS of AI are turning against the technology (out of a realization that it will never do what they hoped it would, that its energy requirements could destroy the planet if we keep trying, and/or that maybe there are some things we should just not be meddling with at our current stage of societal and technological evolution), including Weizenbaum and Hinton
• Brands are now turning against AI … and even the Rolling Stone is writing about it
• big tech and companies that depend on big tech (like Pharma) are starting to turn against AI … and CIOs are starting to drop Open AI and Microsoft CoPilot because, even when the cost is as low as $30 a user, the value isn’t there (see this recent article in Business Insider)

Now, the doctor knows there are still hundreds of marketers and sales people in our space who will consistently claim that the doctor is just a naysayer and against progress and innovation and AI and modern tech and blah blah blah because they, like their companies, have gone all in on the hype cycle and don’t want their bubble burst, but the reality is that

the doctor is NOT against “AI” or modern tech. the doctor, whose complete archives are available on Sourcing Innovation back to June 2006 when he started writing about Procurement Tech, has been a major proponent of optimization, analytics, machine learning, and “AI” since the beginning — his PhD is in advanced theoretical computer science, which followed a math degree — and, after actually studying machine learning, expert systems, and AI, he used to build optimization, analytics, and “AI” systems (including the first commercial semantic social search application on the internet)

what the doctor IS against is Gen-AI and all the false claims being made by the providers about its applicability in the enterprise back office (where it has very limited uses)

because the vast majority of the population does not have the math and computer science background to understand

1. what is real and what is not
2. what technologies (algorithms) will work for a certain type of problem and will not
3. whether the provider’s implementation will work for their problem (variation)
4. whether they have enough data to make it work

and, furthermore, this includes the vast majority of the consultants at the Big X and mid-sized consultancies who graduate from Business Schools with very basic statistics and data analytics training and a crash course in “prompt engineering” who can barely use the tech, couldn’t build the tech, and definitely couldn’t evaluate the efficacy and accuracy of the underlying algorithms.

The reality is that it takes years and years of study to truly understand this tech, and years more of day-in and day-out research to make true advancement.

For those of you who keep saying “but look at how well it works” and produce 20 examples to prove it, the reality is that it’s only random chance that it works.

With just a bit of simplification, we can describe these LLMs as essentially just super sophisticated deep neural networks with layers and layers of nodes that are linked together in new and novel configurations, with more feedback learning, and structured in a manner that gives them an ability to “produce” responses as a collection of “sub-responses” from elements in its data archive vs just returning a fixed response. As a result they can GENerate a reply vs just selecting from a fixed one. (And that’s why their natural language abilities seem far superior to traditional neural network approaches, which need a huge archive of responses to have a natural sounding conversation, because they can use “context” to compute, with high probability, the right parts of speech to string together to create a response that will sound human.)

Moreover, since these models, which are more distributed in nature, can use an order of magnitude more (computational) cores, they can process an order of magnitude more data. Thus, if there is ten to one hundred times the amount of data (and it’s good data), of course they are going to work reasonably well for expected queries at least 95% of the time (whereas a last generation NN without significant training and tweaking might only be 90% out of the box). If you then incorporate dynamic feedback on user validation, that may even get to 99% for a class of problems, which means that it will appear to be working, and learning, 99 times out of 100 instead of 19 out of 20. But it’s NOT! It’s all probabilities. It’s all random. You’re essentially rolling the bones on every request, and doing it with less certainty on what a good, or bad, result should look like. And even if the dice come “loaded” so that they should always roll a come out roll, there are so many variables that there are never any guarantee you won’t get craps.

And for those of you saying “those odds sound good“, let me make it clear. They’re NOT.

• those odds are only for typical, expected queries, for which the LLM has been repeatedly (and repeatedly) trained on
• the odds for unexpected, atypical queries could be as low as 9 in 10 … which is very, very, bad when you consider how often these systems are supposed to be used

But the odds aren’t the problem. The problem is what happens when the LLM fails. Because you don’t know!

With traditional AI, you either got no response, an invalid response with low confidence, or a rare (compared to Gen-AI) invalid response with high confidence, where the responses were always from a fixed pool (if non-numeric) or fixed range (if numeric). You knew what the worst case scenario would be if something went wrong, how bad that would be, how likely that was to happen, and could even use this information to set bounds and tweak the confidence calculation on a result to minimize the chance of this ever happening in a real world scenario.

But with LLMs, you have no idea what it will return, how far off the mark the result will be, or how devastating it will be for your business when that (eventually) happens (which, as per Murphy’s law, will be after the vendor convinces you to have confidence in it and you stop watching it closely, and then, out of the blue, it decides you need 1,000 custom configurations of a high end MacBook Pro in inventory [because 10 new sales support professionals need to produce better graphics] in a potentially recoverable case or it decides to change your currency hedge on a new contract to that of a troubled economy (like Greece, Brazil, etc.) because of a one day run on the trading markets in a market heading for a hyperinflation and a crash [and then you will need a wheelbarrow full of money to buy a loaf of bread — and for those who think it can’t happen, STUDY YOUR HISTORY: Germany during WWII, Zimbabwe in 2007, and Venezuela in 2018, etc.]). You just don’t know! Because that’s what happens when you employ technology that randomly makes stuff up based on random inputs from you don’t know who or what (and the situation gets worse when developers [who likely don’t know the first thing about AI] decide the best way to train a new AI is to use the unreliable output of the old AI).

So, if you want to progress, like the monks, leave that Genizah Artificial Idiocy where it belongs — in the genizah (the repository for discarded, damaged, or defective books and papers), and go find real technology built on real optimization, analytics, machine learning, and AI that has been properly researched, developed, tested, and verified for industrial use.