Part 1: 'Artificial Intelligence' and Machine Learning are Overhyped Scams

One of the advantages of having lived as an adult for a bit over two decades is that you have almost certainly seen a number of hyped trends rise into great prominence only to fade into obscurity. I have, for example, lived through quite a number of them from how sequencing the human genome in late 1990s and early 2000s was “guaranteed” to revolutionize medicine (it had almost zero effect) to how proteomics and genomics was going to change drug discovery forever (once again, they had almost zero effect). I also remember how new anti-cancer drugs such as novel kinase inhibitors and monoclonal antibodies were going to make most cancers curable or increase overall survival, which with few exceptions such as melanoma- they have not. New computational techniques were supposed to revolutionize drug discovery since the mid 1990s- while at best, they have helped streamline the overall process a bit. My point is that not a single hyped trend I have come across in the entirety of my life has ever delivered anything close to what was confidently promised by “famous credentialed experts” aka charlatans.

On the other hand, trends which were not hyped and often ignored have had a much more significant impact on the world. For example- newer drugs and techniques to treat cardiovascular diseases from hypertension, myocardial infarction, cardiac failure etc have had a significant impact on mortality from those diseases and perhaps some effect of overall life expectancy. Or look at how internet and inexpensive computers changed the world around us, even though many still saw the internet as a passing fad into the late 1990s. Another example of an unhyped trend having a major impact on society can be found in how online dating has made every mediocre woman working a minimum wage job believe that she is a date away from snagging a tall millionaire with a big dick. My point is that hyped trends always fail to deliver what they promise, while unhyped ones often have much more significant effects than people realize.

With that in mind, let us move on to the topic of “artificial intelligence” and machine learning- specifically why they are overhyped scams. This issue becomes even more topical since many shills have been hyping the ability of decent chatbots and glorified photo and sound manipulation macros as something radical and earth-shattering. To be blunt, they are either lying or caught up in their own bullshit. To understand what I talking about, here is a link to an article about how the brain is fundamentally unlike computers- both from the point of architecture and how they function. The very short version of that article is that the biological brain, perception, intelligence and sense of self have no equivalent in computer architecture and programming- at least as it exists. At this point, some of you might say.. but isn’t specialized “AI” going to have large and revolutionary effects on the world? The short answer is NO and the rest of this post will explain you why that is the case.

FYI- the slightly blurred AI-generated art shown below is no better than millions of similar pieces generated using Daz Studio and similar software and readily found on sites such as DeviantArt, Gumroad etc

To demonstrate how massively hyped “AI” and machine learning is, let me show you how limited it really is with a few easy to understand examples.

1] Let us start by talking about self-driving cars. As you know, over the last decade, many large corporations from Google, Uber, Apple to Tesla and other car makers have pumped many tens of billions into that pipedream. So what have we achieved? Well.. to understand that we have go back a bit over two decades ago when people started to seriously develop self-driving cars. The short version of that story is as follows.. by the late 1990s the proliferation of laptops and sensors such as digital cameras, LIDAR, GPS etc led to the first semi-successful prototypes of self-driving cars. But what were these prototypes capable of- you might ask? For starters, they could do basic tasks such as remain in their lanes, keep a safe distance from cars in front of them, safely navigate routes if the roads were largely free of other traffic etc. It is important to remember that they could do this with the computational power of laptops from the late 1990s and off-the-shelf sensors of that era.

Now let us get back to 2023 and ask yourself if any of the current ‘self-driving’ capable cars, including Teslas, can do much more than these prototypes from late 1990s and early 2000s. And the simple, if unpleasant, answer is a big fat NO. But why not? Why didn’t all that increase in computational power, sensor technology, machine learning etc result in true self-driving cars. And let us be honest about something else , multiple companies did pour billions each into developing them- and failed. Even worse, they failed to develop such technology for vehicles which run on very defined routes such as public transit buses and cargo trucks. The most common excuse for this sorry state of affairs involves some handwaving about how legal liability issues are stopping the adoption of this technology. But is that really the case? Do you really think that any corporation which developed a self-driving car with lower accident rates, under real-world conditions, than the median human driver would keep quiet about it?

And this brings me to how pathetic “AI” and machine learning work under real-life conditions for most complex problems. If you think about it, true self-driving cars are the lowest and juiciest fruit for commercialization of such technology. The underlying problem is fairly simple and has been partially solved for over two decades. We clearly have the sensor and location technology to survey the immediate environment of the vehicle as well as its absolute position at any point on this planet. The computational power required for performing certain parts of the task such as staying within defined lanes, maintaining a safe distance from vehicles in front etc are trivial and standard features in many cars. And yet, even after spending many tens of billions, the best they have achieved is prototypes of self-driving taxis and buses restricted to low speeds on well-defined routes in developed countries. So the best self-driving car has worse real world capabilities than an average human driver getting a blowjob while driving.

2] Let us now move on to the use of computers in drug discovery, an area in which I have some personal expertise. While people have been trying to use computers for speeding up the process of drug discovery since the 1980s, it was not until the mid-1990s that software which could do certain basic tasks with acceptable rates of success came into being. The areas at which computational approaches to drug discovery had the most success are as follows": 1) categorizing large chemical structure databases by their 2D, 3D, substructure and pharmacophore similarity; 2) identifying biologically and synthetically troublesome chemical features in compounds; 3) identifying compounds with structural similarity with other known to bind other receptors and proteins; 4) rough calculations of physico-chemical properties such as as pKa, logP, solubility etc ; 5) ability to identify chemical structures capable of binding to defined sites in a target protein to a level which makes it worthwhile to use it as a preliminary screening tool; 6) ability to generate homology-based models of protein targets to a level sufficient for use in virtual screening.

While this list might look impressive, there is a lot of nuance missing in those simple statements. This becomes even more glaring once you realize that we have vastly better experimental data for the structure of proteins and organic compounds and their various physical and chemical properties than we have of things such as general human intelligence. In other words, computer-assisted drug discovery (CADD) should be an infinitely easier task for “AI” and machine learning than something such as actually understanding human language. Remember that we understand the physics and chemistry behind a lot of the assumptions made in CADD to the point where we can calculate them from scratch. And yet this whole field, while established, is merely another tool set for reducing the cost of pre-clinical drug discovery and improving the quality of follow-on compounds. The early hype of reliably finding novel and potent compounds which could be developed into relatively safe human drugs with a few clicks of the keyboard disappeared a long time ago, leaving us with a set of OK tools which (in the right hands) can streamline the process of pre-clinical drug discovery.

But wait.. there is more. Many of the concepts in AI and ML such as data mining, algorithmic models, regression analysis, bayesian networks, genetic algorithms and many other methods have been used in CADD for over two decades. Also, many large and small pharma companies have put in billions to study ML and Deep Learning over past decade and the results have been very underwhelming. Even today, the best ML and DL based platforms for drug discovery can, at best, produce 10-20% hits in the micromolar range- which means that the identified compounds are borderline leads for further development. But more importantly, these results are inferior to, what is possible using conventional CADD techniques. To put this in perspective, an area of science in which you can calculate most things from scratch, with boatloads of ready-to-use training data and which is much simpler than human language or behavior is still not amenable to AI, ML and DL despite dozens of companies spending tens of billions on it. Did I mention that success in this field would be immensely lucrative to many corporations with very deep pockets?

In next part, I will go into why AI, ML and DL have consistently failed in medicine and diagnosis- beyond simple histology and a few other niche applications.

What do you think? Comments?