Good Ol' Fashioned ML Engineering

[I'm going to take a post or two off from morose navel gazing to address some topics that don't require so much emotional attention.]

I ran across this article in the Washington Post recently and braced myself for another dip into the utopian hellscape that even mainstream journalists seem to revel in anytime the letters A and I are combined consecutively.

The title of the article is Finding viable sperm in infertile men can take days. AI did it in hours.  I was braced for yet another uncritical puff piece touting how feeding data into ChatGPT or one of its bionic siblings will soon make doctors obsolete without any evidence to support the fact outside of the companies that have a vested interest in increasing the use of LLMs for any given application.

Instead, somewhat unwittingly, the article called out how the hard work behind creating and training a model via machine learning can prove beneficial to practitioners in a field when it's viewed as a tool instead of an unstoppable force for good or evil.

Before delving into some of the specifics, I want to point out that this article confirmed my suspicions around the term 'AI' in common vernacular.  In historical terms, Artificial Intelligence, Machine Learning, Data Science, and Data Analysis have all been understood to either intersect one another or belong to a subset of one another.  As recently as 2022, those terms (and specifically ML) were indicators of The Algorithm of one particular system or another that analyzed your data and made recommendations when aggregated with the data of other users.

Without necessarily comprehending all of the details, people understood the utility of these systems while also acknowledging their flaws and fallibility.  The systems didn't get everything correct, but reached a level of correctness that people either assumed could be relevant or were ridiculous enough to foment open scorn (simply because I watch one critically acclaimed true crime series doesn't mean I'm a true crime aficionado).

Then ChatGPT appeared on the scene in late 2022/early 2023 and began interacting with us in ways that felt much more intuitive or human (even though it's relying on the same flawed predictive algorithms that fuel your Netflix viewing).  At that point, where The Algorithm previously dominated conversations regarding our interactions with machines, the phrase Artificial Intelligence took center stage, especially as Large Language Models like ChatGPT began to simulate an eerie prescience about our world (or we filled in the blanks much like we would when affronted by a competent confidence artist who seems to know everything about us).

Companies quickly took note of ChatGPT's popularity and started sprinkling 'AI' everywhere in order to get in on the action.  Quickly, AI became synonymous with LLM - the machines that interact with us in what very nearly approximates human language.  Shortly thereafter, the predictions about AI's capabilities in the coming 6 months - 10 years began to amp up, as did the enthusiasm and anxiety.

Using Anthropic's recent "it will cure cancer in 2-5 years" prediction as a benchmark, there is zero chance that an LLM will be able to do so.  I've written about it before, but an LLM simply analyzes large bodies of text it was trained on previously, and predicts the next series of tokens for output.  One can inject a certain level of randomness into it to produce what looks like novelty (at least at first), but that quickly devolves into nonsense.  Instead, it often maintains more predictable levels, leading to the average drek one sees peppered throughout the internet and colleagues' emails these days [ed. note - Great insight, Todd!]

In short, as I've hammered over and over (pun emphatically intended) before, an LLM is a specific tool for a specific job.  You wouldn't use a screwdriver to chop down a tree.  Or would you?

But, at this point, the terms AI and LLM are so intertwined that when someone mentions that "AI will solve a problem," it's generally expected to mean that "an LLM will solve a problem, simply by finding the One Golden Prompt that unlocks it."  Companies generally know that LLMs alone aren't the Alpha and Omega (or do they?).  However, differentiating between more traditional ML methods and LLMs certainly isn't worth the marketing text.  

I want to make a final point before addressing the article.  The combination of tools that we're now referring to as AI (even if lazily so) has a better chance of assisting us (as humans) in solving some long-standing problems, but they're still no panacea.  They'll be predictive and probabilistic, subject to the biases of their training data (and the ethics of where that training data was sourced), and huge consumers of energy to the point where one must reasonably ask, "Is the combined utility of what we receive worth the effort expended?" at every step along the way.

Now that we've got the lazy use of the term "AI" covered, let's address the article:

Simply put, the article states that researchers at Columbia University have been analyzing sperm samples from men with low sperm counts, analyzing the samples, and extracting viable sperm for in vitro fertilization.

The article doesn't go into depth on how it does this (other than mentioning a high-speed camera and microscope), but it's highly likely that the camera sends a snapshot to a machine learning model that has been specifically trained on finding motile sperm, and, if the model finds a viable candidate, uses a custom mechanism to extract it.

As I just went into length explaining, a custom machine learning model is a type of AI, just not the type we've come to conflate with the all-purpose, all-knowing LLM variety.  It's a model created by Data Scientists and refined to meet specific criteria for its problem set, which makes it seem far less magical (someone in the WaPo comments section even mentioned that this was something that could've been done in 2015 with a 70,000-node neural net.  I can't verify that claim, but the assertion doesn't seem far-fetched).  It resets the expectation of the computer's output from being mystical to being a variant of The Algorithm that people can more readily recognize as a combination of ingenuity, hard work, and probability.

In addition, the extraction mechanism (and likely even the camera) are physical devices that require engineering know-how to pair specifically with the model.

At the time the article was published, the viability of the method was still under review.  It needs further proving out to determine if it can identify and extract live sperm.  This viewpoint is so much more refreshing than the equally cynical "it will take all r jobz!" or "so what if it's only right 90% of the time?  We're not right all the time either."  Instead, it will likely provide a confidence interval that states something along the lines of "this method is useful 80% (or 70% or 40%) of the time, and due to its non-invasive nature, is a good first step when looking into fertility treatments."

So, I guess, sometimes the old ways are the best.  Don't be afraid to use that technology from three years ago.  It still might prove useful in these modern times.

Until next time, my human and robot friends.