Coffee Break: Against AI, Stem Cells, Cancer Chemotherapy, A Note on Louis Pasteur

Part the First: More Cognitive Surrender Where It Just Won’t Do.  Last week we considered an essay on AI that used the term “cognitive surrender.”  That sums up the current state of that art and the term should gain a following.  Last week Nature published Is AI ruining our skills?  Early results are in – and they’re not good – Reliance on artificial-intelligence tools degrades the abilities of physicians and software engineers:

As more professionals begin to rely on artificial-intelligence tools in their work, could their hard-earned skills atrophy?

That possibility is a growing concern for medical specialists, computer scientists and other workers. Seventy per cent of nurses and 77% of physicians, for example, are worried about losing their skills because of over-reliance on AI systems, according to a survey of US health-care workers published earlier this month.

Their fear might be justified. Evidence suggests that AI-driven ‘deskilling’ is starting to happen in medicine, computer science and other fields. Researchers are now discussing how to preserve important human expertise in the age of AI.

This is old news from last October but hardly surprising, then or now:

A study of physicians in Poland who specialize in endoscopy — the use of flexible probes to examine the inside of the human body — shows how quickly AI tools can erode human abilities. The physicians, who had all performed at least 2,000 colonoscopies during their careers, were given access to an AI system that analyses colonoscopy images in real time and flags a type of precancerous intestinal lesion called an adenoma. The tool was available to the specialists on some days but not on others.

Once physicians began using it, their performance dropped significantly whenever the system was unavailable. During the three-month period before the AI tool was introduced, the specialists found at least one adenoma during 28.4% of colonoscopies. During the three-month period after the tool was introduced, the adenoma detection rate for colonoscopies performed without AI assistance decreased to 22.4%.

That might not sound so bad, but do you really want your gastroenterologist to be ~20% less skilled at identifying a benign adenoma (polyp) that left to itself will become the colon cancer that kills you ten years from now?  I didn’t think so.

From computer science:

To investigate whether skills are being lost in the field of computer science, researchers at the AI firm Anthropic in San Francisco, California, designed a randomized controlled trial in which 52 software engineers were asked to perform a basic coding task. During the exercise, all 52 participants could search the web and access instructions on how to do the task. Half of the participants were prompted to use an AI assistant as well.

Afterwards, all of the software engineers were asked to complete a quiz about what they had learnt from the task. The participants who had used an AI assistant did significantly worse on the quiz than those who hadn’t: the average score was 50% in the AI group versus 67% in the non-AI group. The AI-assisted participants did particularly poorly on questions that required them to diagnose errors in the code, which suggests that they had failed to learn the concepts behind the code that they had just produced (emphasis added).  The study was posted on the preprint server arXiv ahead of peer review. (NB: I seldom refer to papers in arXiv, which means the paper has not been peer reviewed so this must be kept in mind even if the chances that the results are real run to 99.9%)

So, the software experts went from D-minus (67%) to a flat F-minus (50%) because they were basically faking it.  Sounds about right.  This is also what happens to medical students who use online products that promise to “make studying a breeze.”  Which is fine until the Category-5 storm that is their first board exam they must pass to continue comes along after their second year of medical school.

Or as a heading in the article says, “Outsourcing cognition.”  Exactly, which is just another way to describe “cognitive surrender.”  We have all heard that being “educated” doesn’t necessarily mean anything other than “knowing where to look up something.”  That was bullshit then and it still is, even when virtually all of us carry around a device that can access “all the knowledge in the world.”  Please pardon this digression: I had a literature teacher who told the class we were taking his course only so that in ten years we would sound smart at a cocktail party.  This still irritates me all these years later.  That was not true for all of us, especially the handful who thought Light in August and Wise Blood, and even Daisy Miller, were exemplary studies of the human condition.  But by then this professor had been beaten down.  Today he would not exist, because students can’t actually read Faulkner or Flannery O’Connor, and a Henry James novella would be much too long for them to manage on their phones, between their notifications.

What could go wrong?  Everything.  As the article notes, accountants forget fundamental skills and the general run of people forget how to read a map to navigate from one place to another.  And your doctor will miss one-in-five incipient tumors.  This will not end well.  I wonder, when Sam Altman gets old enough to be referred for a colonoscopy, will be want his GI doc to have skills that have not been eroded by his AI?  Or will be prefer that his algorithm read his guts?  One-in-five is great in poker or at the track, but still not much different from playing Russian roulette with a one-in-six revolver.

This deskilling will kill science despite the appearance of productivity, and all scholarship and all literature for that matter.  The only way for a scientist to make a discovery worth of any significance is to notice something that “doesn’t seem right” or to ask an intelligent “what if” question.  Max Delbrück famously said that you must always leave room for the unexpected in your experiments.  However, if you cannot recognize the unexpected because you have separated yourself from your data and don’t really know what to expect, your sterling grantsmanship might get you a career, until recently, but it won’t allow you to make discoveries.  Call it intuition, which is very real for the scientist in the lab, for the literary critic in the library, for the historian in the archives, and for the novelist who observes real life.  This “mishegoss of the crowd” will not end well.

Part the Second: Stem Cell Therapy That Worked Very Well.  In an experiment that began long before AI became a real thing, it has been shown that Stem cells banish severe autoimmune disease for 15 years:

A man and a woman with a rare and devastating autoimmune disease have been in remission for more than 15 years after receiving a stem-cell transplant¹. The positive results, which were reported in Med, suggest that the experimental treatment warrants a larger clinical trial, say scientists.

The two people had a severe and potentially fatal disease in which immune cells produce antibodies that trigger an attack on the spinal cord and nerve connecting the eye and the brain, leading to a condition called neuromyelitis optica spectrum disorder (NMOSD). Symptoms tend to appear in episodes that last for days or months and include eye pain, vision loss, vomiting and weakness or paralysis affecting the arms and legs. Current treatments can prevent these episodes with ongoing medication, but they did not work in these two individuals.

After the stem-cell transplant, the man’s neurological function improved and he resumed a normal life and went on to have two children. The woman was able to use her arms more effectively than before her treatment and no longer requires medication to reduce symptoms.

Basically, each patient’s immune system was chemically obliterated and then replaced by immunologically compatible cells from a donor (i.e., an allogeneic hematopoietic stem cell transplant).  Stem cell transplants are risky.  The patient can die of infection during the procedure.  Sometimes cancer follows.  But the alternative is always worse, which is why the clinical experiment is worth doing.  When stem cell transplants became possible in the 1990s, expectations were unreasonable.  But when they work, they work as in these two cases.  It did not work for Jane Kenyon, and that story is told by her husband Donald Hall in The Best Day the Worst Day: Life with Jane Kenyon, highly recommended.  I have used this book when teaching oncology.  The first-year medical students who we want to be our doctors in the future get it.

There is still more to be learned than has been learned about stem cell therapy, but this progress is real.  The underlying paper is here but behind a strong paywall as far as I can tell: Long-term remission of neuromyelitis optica with allogeneic hematopoietic stem cell transplant.  I have read it once and it seems sound.  However, two patients, just like two positive experiments, are only data but results are sure to come, especially as technique improves.  Still, there is one other thing to keep in mind.  This story comes from Italy, home of slow food and slow (when it needs to be) research and follow-up, in this case 15 years.  Would this have been possible at NIH in their unparalleled Clinical Center?  Probably, past tense.  Going forward?  Unlikely, after the three years of thoroughgoing Lysenkoism and administrative mayhem to come.  More’s the pity.

Part the Third: Another Miss with Lung Cancer, Unfortunately.  Clinical oncology is often a thing of stark beauty in the early twentieth century, and I say that based on personal experience of four years ago when seven weeks of radiation and cisplatin did their work without the need of surgery. So far.  But some cancers remain particularly difficult, with lung cancer near the top of the list.  We should remember that while the cause of ~90% of lung cancer cases is known to a fair degree of certainty, 10% are “sporadic.”  In any case, all cancer patients deserve the best treatments available.

This STAT article notes that Closely watched Pfizer lung cancer drug falls short in clinical trial.  I have followed this approach to cancer treatment for a long time because the drug target is part of the multicomponent assembly involved in cell adhesion and motility that my former students and I spent a lot of effort trying to understand:

Pfizer said Monday that an experimental drug it hoped could replace a widely used chemotherapy in one of the most common forms of lung cancer fell short in a clinical trial.

Expectations had been high that the drug, sigvotatug vedotin, could replace docetaxel, a chemotherapy initially approved in 1996. Last year, Pfizer’s CEO, Albert Bourla, said on an earnings call the drug “could be a driver of growth later this decade.” In a note to investors in May, Leerink analyst David Risinger called the upcoming data readout a “major oncology catalyst” and said he had spoken to a doctor who was “optimistic” about its potential.

Pfizer acquired sigvotatug vedotin when it bought the biotechnology firm Seagen for $43 billion in 2023.

The drug is an antibody-drug conjugate targeting integrin beta-6, a protein that is involved in tissue repair and is undetectable in normal cells but upregulated in cancer cells and is expressed in 90% of non-small cell lung cancer tumors. Pfizer was studying it compared to docetaxel in patients with locally advanced, unresectable, or metastatic non-squamous non-small cell lung cancer.

But the drug did not result in a statistically significant improvement in overall survival over docetaxel. This was the study’s main goal, which means researchers will likely view other potentially encouraging conclusions from the clinical trial as requiring further proof.

The key protein here is integrin beta-6.  Integrins are obligate alpha/beta dimers that mediate signaling across the animal cell membrane in both directions, outside-in and inside-out.  They have been a target for chemotherapy because of their roles in cell adhesion and cell motility.  A hallmark of cancer is the ability of the cancer cell to move about and seed secondary tumors during metastasis.  So far, integrin antagonists have not worked very well as anti-cancer drugs.  Integrin beta-6, since it is up-regulated in cancer cells and undetectable in normal cells, was considered a prime target.  Alas, the drug did not make much of a difference in patients’ outcomes.

Why is this important?  This finding is just one more piece of evidence that the “engineering ideal in biology” guarantees very little.  Integrin beta-6 and its partner alpha subunit are probably doing something important in the cancer cell, but no one really knows what that might be.  To repeat myself, again, the science of biology is not theoretical (compared to physics) and it advances incrementally, one small step at a time.  For integrin beta-6 to be a good target in cancer treatment, it is essential that the scientists on the project know what integrin beta-6 does in the cancer cell.  That it is expressed in cancer cells is only a suggestive factoid.

Maybe it will work better in combination with immune checkpoint inhibitor Keytruda, but that seems like a stab in the dark.  In the meantime, Pfizer paid $46B for Seagen three years ago.  I suppose that was worth it, but this disclaimer is at the bottom of their page:

Forward-looking statements included herein are subject to substantial risks and uncertainties that could cause actual results to differ materially from those expressed or implied by such statements. We encourage you to read our reports filed with the U.S. Securities and Exchange Commission (SEC), including the sections captioned “Risk Factors” and “Forward Looking Information and Factors that May Affect Future Results,” for a description of such substantial risks and uncertainties. These reports are available at pfizer.com and the SEC’s website (a link that leads only to the SEC landing page).

Indeed.  And that does not cover the scientific details.  However, this research would undoubtedly have been done more effectively through an enlarged, new and improved as opposed to a shrunken and dysfunctional, National Institutes of Health, with Pfizer then making a nice profit in the manufacture and distribution of the drug.  This was done by Merck quite well during the early days of the Salk polio vaccine.  The strategy is nothing new, and it works for all except the C-suite and a tranche of shareholders.

Part the Fourth: A Note on Louis Pasteur for Our Time.  It is sometimes forgotten that Louis Pasteur was a chemist whose contributions to chemistry during its second golden age led him in many other fruitful directions, culminating in the germ theory of disease.  He prevailed in that argument decisively, even if the current Secretary of Health and Human Services wants to revive current alternative medicine’s terrain theory of disease that goes back to Pasteur’s contemporary Antoine Béchamp.  This is not to deride Béchamp.  Before bacteria and viruses were understood, his theory was plausible.  But since Pasteur’s Germ Theory, terrain theory has the same epistemological validity as phlogiston and the aether.  I doubt RFKJr is willing to listen to reason, however.  On Louis Pasteur:

Contrary to popular belief, Louis Pasteur was not a doctor or a biologist, he was a chemist. His early fascination was with crystals, particularly with salts of tartaric acid which since antiquity had been known to form during the fermentation of grape juice. One day in 1848 he was examining a sample of such crystals under a magnifying glass and made a remarkable observation. There were actually two kinds of crystals, and they were mirror images of each other! Pasteur laboriously separated the crystals with tweezers and discovered that they had exactly the same physical properties, save for one. When they were dissolved in water and placed in the beam of a special kind of light, known as plane polarized light, they rotated the beam in opposite directions. Since the crystals had been dissolved in water, Pasteur hypothesized that this behavior was actually due to the individual molecules, and that the two types of crystal must consist of mirror image molecules!

This was really advanced thinking, but Pasteur could push it no further because at the time practically nothing was known about the structure of molecules. Twenty-five years later, Jacobus van’t Hoff correctly interpreted Pasteur’s finding by proposing that carbon atoms in molecules connected to other atoms in a specific three-dimensional pattern. This “tetrahedral” pattern could give rise to molecules which were alike in every respect except that, that like our hands, they were non-superimposable mirror images of each other. This explanation laid one of the cornerstones to the building of modern organic chemistry and van’t Hoff was duly rewarded with the first ever Nobel Prize in chemistry in 1901.

Nearly seventy years ago scientists working on bioluminescence in the firefly determined the structure of the molecule (luciferin) that emits light and then synthesized it.  Their product was completely inert; it produced no light whatsoever when mixed with the enzyme luciferase.  It took a while to figure out why, but they had made the wrong stereoisomer in their total synthesis.  After a change in their reaction order they made the active mirror image that would fit in the active site of luciferase.

And now you know why you only need half as much esomeprazole for your heartburn as omeprazole.  Esomeprazole is the correct form that inhibits the proton pump in your stomach, while omeprazole is a 50:50 mix of the active and inactive mirror image molecules.  I plan to come back to this in a discussion of so-called “mirror life” in which scientists who call themselves synthetic biologists believe it possible to “make” organisms that are composed of D-amino acids (dextrorotatory) instead of L-amino acids (levorotatory).  Dangerous?  Possibly.  But probably not as threatening as gain-of-function research on human pathogens.

Part the Fifth: Cory Doctorow on AI.  I had planned to write a review of Cory Doctorow’s new book on AI when after it was released (earlier this week): The Reverse Centaur’s Guide to Life After AI: How to Think About Artificial Intelligence – Before It’s Too Late.  But once again Front Porch Republic has done us a public service with a review of the book by Joshua Pauling.  After getting a few minor criticisms out of the way, Pauling dives in.  Reverse centaur is an apt metaphor for the age of AI, the head of a horse, noble creature that it is, with the legs of a human:

Doctorow argues that all the hype, innovation, disruption, and instability surrounding AI are not primarily driven by the technology itself, but by the economic incentives and investment structures that drive venture capital and big-tech firms—and the ability of Big Tech’s inner-ring to control the narrative in order to drive investment toward their companies. In other words, as Gary Marcus has been saying for years, AI is a bubble that’s bound to pop. Doctorow connects the dots to prior tech bubbles and points to misaligned incentives that drive AI firms to give the appearance of being growth companies. For example, you embed AI into every nook and cranny of your existing app or product, and then declare, “Look how many more people are using AI!” All the while ignoring the fact that users didn’t ask for it, and many find it annoying. Much of AI expansion is not being driven by demand but is top-down. The companies need to do this, Doctorow explains, “because the alternative to growth isn’t stasis, it’s collapse.” This is no small matter, since just a handful of AI firms account for such an outsized portion of the value of all U.S. stocks: “Keeping the growth story alive isn’t about one company, or one sector. The entire U.S. economy hangs in the balance.” Doctorow explains that to be an effective AI critic, “you need to strike at the source of AI’s power, which is the investment capital it attracts.”

The review gets better.  Here is a favorite quote from the book as we finish our coffee:

In describing how the push for driverless trucks would require massive investment in dedicated infrastructure for separate lanes, wireless networking, and more, he states what should be obvious to us all: Driverless trucking is just “a shitty version of a train.”

Pauling claims that Doctorow’s diagnosis and prescriptions may not always be completely up to the problems that will come from AI, but that is another book altogether.  Still, with Enshittification and this book, Cory Doctorow has made a good start at identifying what our Tech Bro overlords have done to us and what they plan to do in the future.  After my copy arrives, I will see for myself and report back.  In the meantime, I need to sit down and finish Enshittification.  Too many books, not enough time.

Thank you for reading!  See you next week.  Is it really possible that 2026 is half over?  Tempus fugit, especially as the clock ticks much closer to the end than the beginning.  Be safe and stay strong.  The ride is about to get even bumpier.