Carl Maria von Weber died two hundred years ago, on 5 June 1826, in London. He had gone there to put his opera Oberon on stage, and tuberculosis finished him off. That is the calendar reason for this post. The other reason is more personal. This year is also 215 years since his First Clarinet Concerto was first played, on 13 June 1811 in Munich. It is one of the concertos I once tried to play myself. In 1996, 30 years ago....

The work

Weber wrote the Concerto No. 1 in F minor, Op. 73 in about a month in 1811. He wrote it for Heinrich Baermann, the clarinet virtuoso of the Munich court orchestra. People say he wrote the whole first movement, orchestra and all, in a single day. Maybe that is not exactly true, but this is clearly an opera composer’s concerto. F minor is a dark, dramatic key. The clarinet sounds like a singer walking on stage: it broods in the first movement, and in the slow movement it holds a hushed conversation with three horns. That still surprises people who think early Romantic wind music is only pretty. The finale is a hunting rondo, fast and cheerful. Ten years later Weber wrote Der Freischütz. You can already hear the start of the theatre here.

I played it once, badly, a long time ago. I played the clarinet for fifteen years, roughly between 23 and 38, and this concerto sat right at the edge of what my fingers could do. Not good, in my own honest grade. That is part of why I keep coming back to it on record. I know exactly how hard the easy-sounding parts really are.

The recordings

Two readings, thirty years and one whole philosophy apart.

First, Walter Boeykens, with the Rotterdam Philharmonic under James Conlon (Harmonia Mundi, recorded January 1989, now on Warner). This is the modern-instrument one, and for me it is personal. I played a few times in a wind band in Bornem next to Boeykens and his daughter. Vlad Weverbergh, then his pupil, sat in the same section. (Weverbergh later became a period-clarinet specialist himself, which matters for the second recording.) Boeykens plays Weber the way a good teacher shows you a phrase: round, even, calm, nothing forced. The tone stays warm from bottom to top, and the slow movement is sung, not whispered. Warner’s own text calls him “a true legend of his art.” That is label talk, but in his case not far off. If you want this concerto to sound beautiful and safe, start here.

Second, Eric Hoeprich, with the Orchestra of the Eighteenth Century under Guy Van Waas (Glossa, 2020). Hoeprich plays a clarinet he built himself: a ten-key copy of a Griessling & Schlott from Berlin, the make Baermann actually bought in 1809, two years before the concerto. This is more or less what Weber heard in his head. The instrument is reedier and less even between the registers, and that unevenness is the whole point. The colours change from note to note instead of being smoothed out, and the natural horns behind it are more open and rougher than modern ones. Jarrett Hoffman, reviewing the disc for Early Music America (December 2020), called the sound “an ideal combination of mellow and clear” and praised the orchestra’s “energetic basses and bassoons.” He also had one honest complaint that fits this piece well: “Only in one work, Weber’s First Concerto, does that soloistic projection tend to be a little too intense.” Misha Donat, in BBC Music Magazine (January 2021, four stars), heard the good side. He pointed to the instrument’s “dark and husky chalumeau register” in the first movement’s development, and to “the warmth of its middle range” in the long passage for clarinet and three horns. His verdict: “there are precious few period-instrument recordings of these pieces, and this newcomer can confidently be recommended.”

So you can listen to them as if they were a small experiment. Same notes, two clarinets. Boeykens gives you the line; Hoeprich gives you the texture. In the slow movement, the passage where the clarinet answers the three horns is smoother and more blended on the modern instrument, and stranger and more vocal on the old one. In the finale, Boeykens is elegant where Hoeprich runs a bit wild, probably closer to what the Munich audience really heard in 1811.

Why still listen

Two hundred years later, this concerto is both a teaching piece and a war-horse. Usually that means it gets played well and heard without much attention. Putting these two recordings side by side helps with that. The modern one shows you how good the music can sound. The old one shows you what the composer was working with, and why some writing that looks only decorative was, at the time, almost impossible to play. I don’t think one is right and the other wrong. The distance between them is the main interest.

And there is a small private symmetry. Once I sat in a Bornem wind band, next to Boeykens. Now, many years later, I listen to younger players from his pupil’s school playing this same Weber on the old clarinet. Both versions of the instrument I half-learned now sit on the same shelf.

Sources & listen-along

• Boeykens / Rotterdam Philharmonic / Conlon — Weber: Concertos & Concertino pour clarinette (Warner Classics / orig. Harmonia Mundi, rec. 1989). Warner Classics · Discogs

• Hoeprich / Orchestra of the Eighteenth Century / Van Waas — Weber & Kurpiński: Clarinet Concertos (Glossa GCD 921128, 2020). Glossa · Chandos

• Work, premiere (13 June 1811, Munich) and Baermann: Wikipedia — Clarinet Concerto No. 1 (Weber); Interlude

• Reviews: Jarrett Hoffman, Early Music America (Dec 2020); Misha Donat, BBC Music Magazine (Jan 2021).

I am an amateur listener with a piano grade-2 background — these are just notes from an amateur, not authority.

A few weeks ago, I wrote here about gunnen, the Dutch verb I use for mentoring. To gun someone something is to want it for them, without counting what it costs you. You can feel it for a student, a younger colleague, a friend, a child: anyone whose growth you care about more than your own part in it.

Several readers asked the same thing. If gunnen is the good attitude, what is its opposite, and why is the opposite so common? I did not have a clean answer. I had a word for the good thing and no nice word for the thing that ruins it.

Two books gave me the vocabulary. Neither is about mentoring. Both sit on the part of my shelf I reach for often.

The enemy has a name

Ryan Holiday’s Ego Is the Enemy (2016) makes one point and repeats it throughout a life. When you are trying to rise, ego makes you talk instead of work. When you have arrived, it makes you believe the story that other people tell about you. When you fail, it stops you from learning the one lesson the failure was offering.

The examples are everywhere once you see them. The beginner who needs everyone to know how clever the plan is, before anything has been built. The person at the top who has confused their reputation with their judgment. The teacher, the manager, the parent who cannot say “I was wrong about this,” and so lets someone they are guiding walk another year down a road that closed long ago.

That last one is mine. Early in my own PhD, I changed topic halfway, because the first line had dead-ended. It felt like failing. Looking back, it was the most useful thing I did. It taught me to value changing course over finishing a plan for its own sake. Holiday would name the refusal to change course for what it is. Not perseverance. Ego.

Seen this way, gunnen is simpler than I made it sound the first time. It is what wanting good for someone looks like once you take your own ego out of it. You stop needing their success to be partly yours. You stop needing to be the most important person in the room you helped build.

When ego hardens

Holiday treats ego as a visitor everyone lets in now and then. Thomas Erikson, in Surrounded by Narcissists (2022), describes what happens when the visitor never leaves.

His most useful page separates narcissism from ordinary confidence. A confident person has a realistic read on what they are good at, and can still admire what someone else is good at. The narcissist holds an unshakeable belief in their own superiority and needs a constant supply of admiration to keep it standing.

The full clinical version is rare. I have met it once or twice in thirty years (and we all know a few public figures with clear pathology). But it is not the everyday danger. The everyday danger is the milder pull underneath it: the need for the credit, the trouble hearing criticism, the quiet certainty that your judgment outranks the facts in front of you. Most of us carry some of it. The job is to notice it working.

It shows up as a preference dressed as a standard. You take over a task someone else was still learning and call it helping. You put your name on work that was mostly theirs and call it normal. You keep a decision routed through you longer than it needs to be and call it responsibility. None of this needs a personality disorder. It needs only that you like being central, and stop checking whether you still need to be.

I should say what kind of book this is. Erikson sells a popular, color-coded personality system that academic psychologists have criticized for thin evidence (but which has some merits, more later). Take the clinical claims lightly. But the line between confidence and narcissism is the one I keep, because it tells you that the enemy of gunnen is not the rare monster across the corridor. It is the ordinary appetite for the center, in me...

The part the books miss

Both books share a blind spot. They treat ego as a private failing, something one person can read a book and fix. For one relationship, that is often true. For a whole system, no.

Ego also gets built into how things are arranged: who gets the credit, who gets the platform, whose name goes first. I see it most clearly in my own field. Many of the researchers I work with are in the global South. The pattern is an old one: the data they collect travels to wealthier countries, and the recognition travels with it. Careers are built where the data lands, not where the work was done. Wanting good for them means more than keeping my own ego out of one conversation. It means giving up the advantage the system hands me by default. The test is not that they thank me. It is that their work no longer needs me at all.

A small practical test

If there is anyone whose growth you care about, someone you manage, teach, or raise, here is the test I use.

Think of the last time they brought you something unfinished. Did you point out where it was wrong and hand it back, or did you fix it yourself and call that helping? The first is gunnen: their own correction is what teaches. The second feels generous, and is mostly the appetite Holiday and Erikson describe, the kind that needs the result to carry your fingerprints.

There is a second half to this: how you build the conditions that make their own correction stick. That is a different book, and the next post.

I fail this test often enough to keep both books within reach and a lot to talk about...

Books: Holiday, R. (2016). Ego Is the Enemy. Portfolio/Penguin. Erikson, T. (2022). Surrounded by Narcissists. St. Martin’s Essentials (first published in Swedish as Omgiven av narcissister, 2021). I have no relationship to either author or publisher.

Heating a tumor to 41.5 °C sounds like fringe medicine. It is not. Oncologic use of hyperthermia has evidence that most treatments marketed as “emerging” never reach. The puzzle I keep coming back to, as a medical doctor, scientist, and now as someone trying to fund a trial for it, is the gap between that evidence and how rarely the treatment is used.

What heat actually does

Heat already does a lot on its own, but its best-proven value is as a sensitizer for the treatments we already use. At 39–41 °C, it starts to activate antitumor immunity by improving antigen presentation and increasing immune cell infiltration into the tumor. At around 41 °C, it impairs the DNA repair machinery that tumor cells rely on to survive radiation and cytotoxic chemotherapy. In a systematic review I co-authored, we mapped these effects onto the hallmarks of cancer and found that induced hyperthermia modulates seven of the ten (Gorbaslieva et al., 2025). Others reach the same picture from different angles: localized heat acting on the hallmarks of cancer (Hannon et al., 2021), and the argument for treating regional hyperthermia as a targeted therapy built around blocking DNA repair (Issels et al., 2016).

These reviews also make a point I find more honest than most of the field’s marketing: the effect is temperature-dependent. Heat too hard and you lose effect and cause harm; heat too little and the effect is too small. This treatment has a therapeutic window that demands strict control.

The radiotherapy evidence

The cleanest data come from radiation. In the Dutch Deep Hyperthermia Trial, adding hyperthermia to radiotherapy for locally advanced cervical cancer raised the complete response rate from 57% to 83%, and three-year survival from 27% to 51% (van der Zee & González, 2002). Radiation toxicity did not increase.

The meta-analyses point the same way. About a 25-percentage-point gain in complete response for head and neck cancer (Datta et al., 2016a), and 60% versus 38% complete response in locally recurrent breast cancer (Datta et al., 2016b), both without a meaningful rise in late toxicity.

Almost never used....

The chemotherapy evidence

Chemotherapy has its own randomized signal. EORTC 62961 added regional hyperthermia to neoadjuvant chemotherapy in high-risk soft-tissue sarcoma and improved local control (Issels et al., 2010). At a median follow-up of 11 years, five-year survival was 63% versus 51% (Issels et al., 2018). That is the first phase 3 trial to show heat adds to chemotherapy, and the benefit held up over a decade.

Almost nowhere used....

Why a treatment this good stays at the margin

If a drug carried this profile across three tumor types, it would be standard of care. Hyperthermia is not, and the reasons are operational, not scientific.

The field grew up around selling devices. A hospital buys a machine once. There is no recurring revenue to fund the large, well-designed trials that change guidelines, so the trials stayed small, academic, often non-randomized or controlled, single-center, and heterogeneous in how they heated and how they measured dose. Those are exactly the features a reviewer uses to discount a literature. The credibility gap is real, but it is a funding-and-design gap, not a biology gap.

The bet

This is the part I now live in. The indication I care about most, metastatic pancreatic cancer, has only low-quality evidence. Across 14 studies, median survival was 11.7 versus 5.6 months in favor of the hyperthermia arms, but the studies were too heterogeneous to draw firm conclusions (van der Horst et al., 2017). The honest reading is “signal, not proof.”

However, what the field has never had is a properly powered, properly funded randomized controlled trial, professionally executed with a business model behind it that can pay for the next one. That, more than any new physics, is what would move hyperthermia from the margin into the protocol.

The question was never really whether heat works. Four decades of data already answer that. The open question is whether anyone will build the commercial engine to run the trial that proves it at the scale oncology demands before it commits.

We are trying to do just that. Setting up a randomized, controlled, multicenter, well-powered trial. I will keep you posted....

Disclosures. I am CEO and shareholder of ElmediX, a Belgian MedTech company developing whole-body hyperthermia for metastatic pancreatic cancer, and a co-author of one of the reviews I cite here (Gorbaslieva et al., 2025). This is my reading of the evidence regardless of those interests, but you should know about them.

References

Datta, N. R., Rogers, S., Gómez Ordóñez, S., Puric, E., & Bodis, S. (2016a). Hyperthermia and radiotherapy in the management of head and neck cancers: A systematic review and meta-analysis. International Journal of Hyperthermia, 32(1), 31–40. https://doi.org/10.3109/02656736.2015.1099746

Datta, N. R., Puric, E., Klingbiel, D., Gomez, S., & Bodis, S. (2016b). Hyperthermia and radiation therapy in locoregional recurrent breast cancers: A systematic review and meta-analysis. International Journal of Radiation Oncology, Biology, Physics, 94(5), 1073–1087. https://doi.org/10.1016/j.ijrobp.2015.12.361

Issels, R. D., Lindner, L. H., Verweij, J., Wust, P., Reichardt, P., Schem, B. C., … Hohenberger, P. (2010). Neo-adjuvant chemotherapy alone or with regional hyperthermia for localised high-risk soft-tissue sarcoma: A randomised phase 3 multicentre study. The Lancet Oncology, 11(6), 561–570. https://doi.org/10.1016/S1470-2045(10)70071-1

Issels, R. D., Lindner, L. H., Verweij, J., Wessalowski, R., Reichardt, P., Wust, P., … Gronchi, A. (2018). Effect of neoadjuvant chemotherapy plus regional hyperthermia on long-term outcomes among patients with localized high-risk soft tissue sarcoma: The EORTC 62961-ESHO 95 randomized clinical trial. JAMA Oncology, 4(4), 483–492. https://doi.org/10.1001/jamaoncol.2017.4996

van der Horst, A., Versteijne, E., Besselink, M. G. H., Daams, J. G., Bulle, E. B., Bijlsma, M. F., … van Tienhoven, G. (2017). The clinical benefit of hyperthermia in pancreatic cancer: A systematic review. International Journal of Hyperthermia, 34(7), 969–979. https://doi.org/10.1080/02656736.2017.1401126

van der Zee, J., & González González, D. (2002). The Dutch Deep Hyperthermia Trial: Results in cervical cancer. International Journal of Hyperthermia, 18(1), 1–12. https://doi.org/10.1080/02656730110091919

Gorbaslieva, I., Quisenaerts, T., Bogers, J. J. P. M., Peeters, M., Saldien, V., & Ysebaert, D. (2025). Temperature-dependent effects of induced hyperthermia, including whole-body hyperthermia, on the hallmarks of cancer: A systematic review. Cancers, 17(23), 3824. https://doi.org/10.3390/cancers17233824

Hannon, G., Tansi, F. L., Hilger, I., & Prina-Mello, A. (2021). The effects of localized heat on the hallmarks of cancer. Advanced Therapeutics, 4(7), 2000267. https://doi.org/10.1002/adtp.202000267

Issels, R., Kampmann, E., Kanaar, R., & Lindner, L. H. (2016). Hallmarks of hyperthermia in driving the future of clinical hyperthermia as targeted therapy: Translation into clinical application. International Journal of Hyperthermia, 32(1), 89–95. https://doi.org/10.3109/02656736.2015.1119317

Every AI vendor now sells “agents.” The word has been stretched so far that it means almost nothing. So let me describe what actually changed on my screen this year, in plain terms, and then be honest about where it still falls over.

I am a CEO, a pathologist and a researcher, not a programmer. I have no stake in any of these companies. What follows is one user’s report.

A chatbot and an agent are not the same thing

For two years, I used AI the way most people do. I opened a chat window, pasted in some text or asked a question, and copied the answer back out. Useful, but it was a conversation. Every step was mine. The model never touched anything I owned. Almost like a “Google, plus a little extra”.

An agent is different in one concrete way: it has access to some of my material, and it takes steps on its own. Instead of pasting a transcript into a window, I point it at the folder where my notes live, and it reads them. Instead of asking it to “write an email,” it can open a draft in my mailbox. Instead of one answer per prompt, it does five or six things in a row: search, read, write, check, and comes back when it is done.

That is the whole shift. Not “smarter answers.” Access plus a sequence of actions. Everything good and everything dangerous about agents comes from those two words.

What it is good at

I will give a few tasks from my own week, because one concrete example is worth more than three claims.

Meeting notes. I record a research or trial meeting, the agent reads the transcript, and it produces a structured report: who said what, the decisions, the to-do list. It corrects the obvious transcription errors — some names came out wrong every single time, and it now fixes that without being told. It knows the entire team, their role, and their correct name. A job that used to eat an hour after every meeting is now a five-minute review of something already written.

Finding literature with the sources attached. When I want to know what the evidence says on a narrow question, the agent searches, reads the papers, and writes a short synthesis with the references listed. The key point is the references. I do not want a confident summary; I want to be able to click through to the original and check it myself. More on that below, because this is also where it fails.

Keeping my own notes in order. I keep everything in one large set of files I have built over the years. The agent can scan it all, find the note that contradicts another note, and propose links between related pieces. Make complete trees or mind maps with links that it found itself. Dull, slow work for a human; quick for a machine that does not get bored.

Most futuristically, perhaps, it can read my calendar, my email, my to-do list, my small notes, and create a daily briefing. It can even complement it with a list of news site headlines, specifically looking for things that interest me, because it knows me.... This starts to become a fully functional robot, but I will cover that subject in a later post.

None of this is revolutionary; it is all available today. It is administrative weight lifted off a working week. That is exactly why it is worth taking seriously.

Where it still fails

The same two properties — access and autonomy — are the problem.

It likes to invent things confidently. I have had an agent hand me a literature summary with a reference that looked perfect: plausible authors, plausible journal, plausible year. The paper did not exist. This is not rare, and it is not a bug you can switch off. A model that writes fluent text will, some of the time, write a fluent citation for a paper nobody wrote. If you are a researcher, this is the single most important thing to understand before you trust any output. Fluency is not accuracy.

It is wrong without flinching. A junior colleague who is unsure tells you they are unsure. The model gives an incorrect answer in exactly the same confident tone as a correct one. There is no tremor in the voice. You only catch the error if you already know enough to catch it, which is an uncomfortable thing to depend on.

It misses the context a person would have. It does not know that one investigator is touchy about a particular topic, or that a result is too preliminary to put in writing. It optimizes for the task you gave it, not for the judgment around the task.

Autonomy cuts both ways. The same feature that lets it draft an email for me is the feature that, unsupervised, could send one. The capability that saves time is the capability that does damage if you are not watching. Therefore, be very cautious with all permissions you give to your agent.

The real work becomes checking

Here is the part that nobody selling agents wants to dwell on. When the machine does the typing, verification becomes your job. The bottleneck moves from producing text to confirming it.

So I work with a few hard rules, and I would suggest that any colleague start with the same ones:

• Nothing leaves under my name without my approval. No email sent, no message posted, no document submitted, until I have read it. The agent drafts; I decide.

• No patient-identifying data, ever. This is not negotiable, and no convenience is worth crossing it.

• Every factual claim goes back to a source I can open. If it cannot give me the reference, I treat the claim as unverified. If it gives me one, I check that the reference is real.

You can tell your agent that these are hard rules for you, “he” will listen....

These rules sound restrictive. They are not. They are what makes the thing usable at all. An assistant you cannot check is not an assistant; it is a liability with good manners.

What it costs

To be transparent, because this is the AI-tools corner of this Substack: the assistant subscriptions that make this kind of work possible on a significant scale run roughly at least twenty dollars a month at the entry level, with realistic tiers at one hundred to two hundred. I pay for more than one, partly for the work and partly because comparing them is how I learn what each is good for. Whether that is worth it depends entirely on how much of your week is the administrative weight I described — for me, it clears the cost easily; for a colleague whose week looks different, it might not.

Bottom line for a colleague still at the sidelines

If you have only ever used a chat window, an agent is worth trying, with three conditions. Give it the dull, high-volume work first — notes, sorting, first-pass literature. Keep it away from anything sensitive until you trust your own checking habits. And never relax the rule: verify before you act (or before you trust).

The promise is not that it thinks for you. It is that it clears the desk so you have time to think. That is a smaller promise than the marketing makes, and a more honest one.

I run a setup like this over my own notes, calendar, and mail every day. How it is actually wired together is a longer story, and one I will tell another time.

Disclosure: I pay for several AI assistant subscriptions out of my own pocket and have no financial relationship with any of the companies mentioned. These are my own observations from daily use, not advice.

I have tried. “To grant” sounds bureaucratic. “To support” is too soft and too clinical at the same time. “To wish someone well” gets the feeling but loses the will. “To mentor” misses the real meaning entirely. Mentoring is a method. Gunnen is a stance.

The Dutch verb has a will inside it. To “gun” someone something is to want it for them – actively, with no calculation about what it costs you. It is most often used about good things: a promotion, a piece of luck, a hard-won result. “Ik gun het je” means I want this for you. There is no English sentence that does the same work in the same number of syllables.

This gap matters because the concept is the foundation of how I work with students, particularly PhD students. Without the right word, the practice gets misnamed constantly. People call it “developing talent” or “empowering students” or “supervising a thesis.” None of those address what the work actually is.

What it is

Three elements, short.

1. To gun someone something is to want them to become bigger than you. Not bigger than they were when they started. Bigger than the person standing across from them in the lab. That is the test.

2. It only works if you mean it. Students see through it inside a year, sometimes inside a week. If you are mentoring because it is expected of a senior, the student will know.

3. The endpoint is your own redundancy. A mentee who can design a study, write the paper, and defend their choices without you. When they leave the group, they should not need to send drafts back for comment. If they still need you, the work is not done.

What the outcome looks like

The only outcome I judge myself on in this role is whether my former PhD students now run their own groups and companies and mentor their own students the same way. Some have stayed in academia and supervise their own master’s and PhD students. Others have left for industry. Either way, the questions they work on are their own, and their co-authors are their own teams, not mine – unless I have done the work to deserve a place on the byline. The goal of gunnen is not for the mentee to become you. It is that they become bigger than you, in their own direction. Most of what they do now carries no link back to me, and that is the measure.

How it works in practice

The day-to-day is unglamorous. I try to read everything my students produce, or at minimum, I think hard about their work before we meet. Feedback is direct and limited: on their draft, not on the version I would have written. It is their PhD, not mine.

That second clause is the one that takes discipline. The easier path is to rewrite a student’s draft into your own voice and call that editing. It feels productive. It produces a clean text quickly. It teaches the student almost nothing.

The harder path is to leave the draft as the student wrote it, mark it only where it is genuinely wrong, and hand it back. Perhaps the paper will get rejected. The student will make mistakes again. They will catch some of them the second time. The mistakes they catch themselves teach them more than the instructions they follow without thinking.

I step in directly only when the cost of the mistake is too high: patient safety, irreplaceable samples, a deadline that cannot be moved. Outside those constraints, I let the student steer – including when they steer somewhere I would not have gone.

Why the verb matters

You can practice mentoring without the verb. People do. But the word changes the posture.

“To mentor” implies a relationship in which the senior has the knowledge and the junior receives it. The flow runs in one direction.

“To gun“ does not place anyone above anyone. It says: I want this for you. The wanting is in me. The doing is in you. The success is yours, not partly mine.

That distinction shows up in small choices. Whose name goes first on the paper. Whose conference talk it becomes. Whose grant proposal it builds into. If the mentor’s instinct is I helped, so I get credit, the relationship was wrongly framed from the start. If the instinct is I wanted this for them, and now they have it, the credit question does never arise.

This way of working is often mistaken for laziness. Some colleagues call me lazy and not much of a promoter. Some students leave when they find they cannot work without a steadying hand. So be it, I’m convinced of my way of working. The students and collaborators who stay, thank me later, and they have really deserved their credits because they really did it, not me.

The cross-role pattern

I should be transparent about something I noticed only later. I run a startup, ElmediX, in addition to my academic role. When I drafted what I wanted to be as the CEO of that company, the sentence I wrote was: I build a team that runs without me.

When I drafted what I want to be as a senior academic, the sentence I wrote was: until they are bigger than I am and I am no longer needed.

Same sentence, two roles. It is one thing, said twice, because it really is one thing.

I do not think this is special to me. Most senior people who are useful to those working under them maintain the same posture across whatever roles they hold. They may not have a Dutch verb for it. The posture is what makes the work worth doing for the next twenty years.

A small invitation

If you are a senior academic reading this and you find yourself rewriting drafts into your own voice, try the harder thing. Mark it up very limited. Hand it back. Let the reviewer reject it.... Tolerate the slower clock.

If you do not have a word in your own language for what gunnen names, borrow ours. We do not mind.

Disclosure: I am CEO and shareholder of ElmediX, a Belgian MedTech company developing whole-body hyperthermia for metastatic pancreatic cancer. This is my opinion regardless.

One hundred and thirteen years ago today, on 29 May 1913, the audience at the Théâtre des Champs-Élysées in Paris fought through the first performance of Le Sacre du printemps. Pierre Monteux conducted. Vaslav Nijinsky choreographed for Diaghilev’s Ballets Russes. People shouted, whistled, and reportedly came to blows. Music history files the evening as a scandal about the music. It was probably more of a scandal about the dancing. The notes were unfamiliar. The bodies on stage were even stranger.

I have known Le Sacre for forty years without ever sitting with it. For this first Long Listen, I tried, properly, with two recordings on two separate evenings.

The work

A ballet in two parts, about thirty-three minutes. Stravinsky imagined a pagan rite in pre-Christian Russia: tribes gather, a young woman is chosen, and she dances herself to death so the earth can be reborn. The score is built less on melody than on rhythm as a character in its own right: uneven accents, overlaid pulses, blocks of sound that change without a smooth handover. Two versions are in circulation. Stravinsky published the score in 1913 and revised it in 1947, mostly for copyright reasons, but with real differences in accents, dynamics, and orchestration. The two sound noticeably different in places.

The recordings

I compared two early-career studio takes.

• Pierre Boulez, Cleveland Orchestra, 1969 (Columbia/Sony). The 1947 score.

• Leonard Bernstein, New York Philharmonic, 1958 (Columbia/Sony). The 1913 score.

Both conductors would record Le Sacre again later in their careers. These were their first studio takes, and both were in their first peak. I listened on Qobuz through a closed-back headphone. Not my home reference rig because of traveling, but consistent across both recordings, which is what matters here.

First impression: Boulez is crisp, Bernstein is fluid. That stayed true across the four moments I tracked.

The opening

A high bassoon, then woodwinds dripping in. Stravinsky wrote the bassoon deliberately too high for comfort, so every instrument would feel caught. Boulez starts at rest. The bassoon is unhurried, and the oboe later is what holds my attention; Boulez gives it real room. The texture tips from pastoral into something else around [1:15]. Bernstein is slightly faster and more on edge. The same tip arrives two seconds earlier at [1:13], and the entries feel slightly cut off, more angular. Both are legitimate readings of the same score.

Augurs of Spring

The eight-note string chord with the hammered horn accents — the passage everyone remembers. Boulez plays the string accents as a push: weight forward, leaning. The horns are assertive. Bernstein plays the same accents as a chop, shorter, downward, with the horns kept under. The cellos sit at the bottom in both. Around [1:22] in Boulez and [1:23] in Bernstein, there is a switch, and the switches are not the same gesture. Some of that is interpretation; some may be the score versions, since the 1913 and 1947 accent patterns are not identical everywhere.

Glorification of the Chosen One

An asymmetric pulsing tutti, close to the end of Part Two. Boulez treats the meter as something to be measured. The bar changes are chopped, almost surgical, and the winds lead the texture. The feel is crisis. Bernstein is less strict. The meter changes wave through the orchestra, and the percussion comes more to the front. The feel is ritual, which is exactly the word the choreography intended. Two readings, two different things. Both correct.

Sacrificial Dance

The climax. Brutal meter changes, the standard test of a conductor’s nerve. Boulez is precise, a touch calmer than I expected, with sharper articulation; the percussion at the end dominates; the final chord is held, slightly open at the edges. After the last note, there are fourteen seconds of silence on the recording. Bernstein is more chaotic, but not carelessly. The openings of the meter changes are less surgical, the percussion stays more transparent inside the tutti, and the final chord lands as a single hit. Two seconds of silence after. The studios cut the tails differently, but the gestures are different too.

Why still listen

For most of the past hundred and thirteen years, almost all popular music has lived on a regular grid. Pop, rock, dance, most film scores — they keep time so dancers, editors, and listeners can hold on. Le Sacre is the first piece I know where rhythm itself is the protagonist, not a frame for melody, not a pulse to keep you oriented, but the thing the music is about. That is what makes it hard on the first listen, and it is what makes it useful to come back to. After two sessions, the asymmetry stops feeling like chaos and starts feeling like architecture.

I will be honest. I knew this piece for forty years without sitting still with it. It is an acquired taste, and I had not bothered to acquire it. Two evenings later, I am surprised by how much there is. If you have not listened in a while (or ever), today is a good day. Available on all streaming platforms and for free on YouTube. No excuses... But come back a few times, it is worth it!

Sources and listen-along

• Boulez, Cleveland Orchestra, 1969 (Columbia/Sony) — on Qobuz, Spotify, Apple Classical, YouTube.

• Bernstein, New York Philharmonic, 1958 (Columbia/Sony) — on Qobuz, Spotify, Apple Classical, YouTube.

• Premiere date, conductor, and choreographer: Grove Music Online, lemma Rite of Spring, The.

• The 1913 and 1947 score versions: Stravinsky, Le Sacre du printemps, entry on IMSLP.

I am an amateur listener with a piano grade-two background. These are listening notes, not authority and definitely no review.

— John-Paul

Most novels you can read at any age. The Remains of the Day is not one of them. I read it for the first time this year, at sixty-one, thirty-seven years after it won the Booker. I read it on my eldest son’s advice. He is a better reader than I am.

The book opens in July 1956. Stevens is the butler at Darlington Hall, an old English country house now owned by Mr Farraday, an American. Farraday suggests he take a short holiday and lends him the Ford. Stevens drives west for six days to meet his former housekeeper, Miss Kenton, now Mrs Benn, with the half-acknowledged hope that she might return to Darlington Hall. The trip frames the book. His mind keeps going back to the 1920s and 1930s, the Lord Darlington years, the years that mattered.

What Lord Darlington was

Lord Darlington was an English aristocrat who, in the years between the wars, believed the punishment imposed on Germany at Versailles was too harsh and that another war could be prevented through quiet diplomacy and personal relationships with influential Germans. He hosted secret conferences at Darlington Hall in 1923 and again in 1936. By the late 1930s, those conferences had become indistinguishable from appeasement; by 1940, from collaboration. After the war, Darlington was disgraced. He died alone in 1956, the same year as the trip, his name attached to a libel suit he had lost.

Stevens served Darlington devotedly throughout. His definition of *dignity* required him to suspend his own judgment in favor of professional service. When his employer dismissed two Jewish housemaids in 1935 at the request of a German guest, Stevens carried out the dismissals. When Miss Kenton confronted him in the kitchen, he gave her the answer the butler-version of himself was supposed to give.

The pier

On the sixth day of the trip, in Weymouth, Stevens sits on a pier at dusk. Miss Kenton is gone, definitively. She has decided to stay with her husband. A stranger on the bench beside him asks how he is doing. Stevens cries briefly. He says, almost to himself, that he gave his best years to Lord Darlington, that he cannot even claim the dignity of having made his own mistakes. Darlington made them, and Stevens served the consequences.

Then the stranger says something practical. The evening is the best part of the day; Stevens should make the most of what remains rather than mourn what is already behind him. Stevens decides, sitting there, to learn the American art of *bantering*, so that he can serve Mr Farraday properly in the years that are left. The book ends on that decision.

It is a small decision. It is a devastating ending.

Why it lands different after sixty

At forty, I would have read Stevens as a foolish man who wasted his life. At sixty-one, I read him as a recognizable one. The difference is not sympathy. It is calibration. By sixty, you have made enough of your own version of the dignity-trade to know the moments you let your institution do your thinking, the corners you turned away from because the alternative would have cost too much.

Ishiguro gets the timing right. Stevens does not realize he has been lying to himself in his thirties and corrects course. He realizes in his fifties and changes very little. He realizes fully, on the Weymouth pier, in his sixties, and what he can do with the realization is small. Learn to banter. Be a better butler to the right employer, in the time that remains.

That is what the rest of the day looks like.

This is also why so many people misremember the book as melancholy. Stevens decides to keep working. The work is smaller now, and the employer is worth serving. He has the evening, and he has something to learn during it. That is what Ishiguro is actually saying.

The small thing

If you have not read The Remains of the Day, read it. Faber and Faber’s paperback is cheap and small. If you read it in your twenties or thirties, read it again now. The book does not change. You do, and the same sentences mean different things.

The title of this publication is no accident.

– John-Paul

A patient sits across from their dermatologist holding a pathology report. It reads: melanocytic lesion with architectural disorder and moderate cytologic atypia, consistent with dysplastic nevus. The question is always the same: what do we do with this?

It is a fair question. The diagnosis is common. In any busy dermatopathology practice, dysplastic nevi are among the most frequently reported melanocytic lesions. And the answer, in most cases, is less clear than that frequency implies.

What the histology actually describes

A dysplastic nevus is defined histologically by two things: architectural disorder in the junctional component, and cytologic atypia. The architectural features include junctional “shoulders” extending beyond the dermal part of the lesion, bridging of nests between rete ridges, lentiginous melanocyte proliferation at the periphery, fibroplasia in the superficial papillary dermis, and a patchy lymphocytic infiltrate (Duffy & Grossman, 2012; Menzinger et al., 2025). Clinically, these lesions tend to be larger than ordinary nevi, asymmetric, with irregular borders and color variation. Those features overlap with both benign nevi and early superficial spreading melanoma.

The 2018 WHO classification placed dysplastic nevi within a two-tier grading system: low-grade and high-grade (Ferrara & Argenziano, 2021). The intention was to reduce the intermediate categories that nobody quite knew what to do with.

In practice, the grading remains the central problem.

The reproducibility problem

Pathologists do not reliably agree on this diagnosis, particularly on the grade. Every clinician who sends tissue for melanocytic evaluation should know that.

In the classic interobserver studies, agreement was unreliable across the board. Grading kappa values among experienced dermatopathologists ranged from 0.38 to 0.47 (Duncan et al., 1993); recognition agreement ranged from 0.32 to 0.71 across observers (Duray et al., 1992). Applying the 2018 WHO criteria improved standardization in principle; in practice, a study applying those exact criteria to a series of cases found only fair multirater agreement (κ = 0.40) (Ulgen et al., 2021). A Norwegian reassessment of melanocytic lesions excised across two different time periods found that agreement for the category “irregular/dysplastic nevus” was approximately 29%, and for “nevus with severe atypia” approximately 28% (Gjersvik et al., 2022). A 2025 multicenter study published in Nature Communications confirmed what many dermatopathologists already knew from case conferences: discordance in this diagnostic space is substantial, even among experienced specialists (Haggenmüller et al., 2025).

A large tertiary-center study found overall concordance of 72% across all melanocytic diagnoses, substantially lower than for unambiguous melanoma (Ronen et al., 2021). For clear-cut melanoma and ordinary common nevi, agreement is high. For the lesions between those poles, including moderately dysplastic nevi, it drops considerably.

The practical implication is uncomfortable but important: the same slide, seen by two competent pathologists, may receive a different grade. Occasionally a different diagnosis entirely. This is not individual failure. It reflects the biological reality that the morphological features used to grade atypia sit on a continuum, without a reliable histological bright line.

What the biology says

The current evidence, consolidated across multiple independent lines of research, supports viewing dysplastic nevi primarily as risk markers, not obligate precursors to melanoma (Duffy & Grossman, 2012; Menzinger et al., 2025; Ferrara & Argenziano, 2021; Wiedemeyer et al., 2021). Most melanomas arise de novo, not from pre-existing nevi. Some arise from common nevi or dysplastic nevi, but the per-lesion risk is low. Having multiple dysplastic nevi, or having them in a patient with a personal or family history of melanoma, increases overall risk. A solitary mildly dysplastic nevus in a low-risk patient is a different clinical situation.

The most pointed position in the literature is Bernard Ackerman’s, articulated posthumously by Hurwitz and Tavel (2015): the diagnosis of dysplastic nevus is “mythical” and its clinical consequences are “untoward”, creating patient anxiety and surgical overtreatment without evidence of benefit, based on histological criteria too imprecise to justify the label. Ackerman was not universally right, and this view has not prevailed. But the argument belongs in the clinical conversation, not on the margins of it.

What to do — the management question

For high-grade dysplasia: most authors and the WHO recommend complete excision. The overlap with melanoma in situ at this end of the spectrum, combined with the diagnostic discordance precisely at that boundary, makes a conservative surgical approach reasonable (Duffy & Grossman, 2012; Menzinger et al., 2025; Ferrara & Argenziano, 2021). Margins of 2 to 3 mm are generally sufficient. A 2020 retrospective series of 426 severely dysplastic nevi re-excised with exactly those margins found no upstaging to melanoma, no local recurrences, and very low rates of residual dysplastic tissue when the initial biopsy had already removed the bulk of the lesion histologically (Soleymani et al., 2020). The data do not support wider excision as a default.

For low-grade and moderate dysplasia with complete initial excision: routine re-excision is not supported by the available evidence (Duffy & Grossman, 2012; Soleymani et al., 2020; Ferrara & Argenziano, 2021). The reflex to re-excise every lesion reported as dysplastic generates surgical morbidity without demonstrated clinical benefit. A complete excisional biopsy with clear margins is, in most cases, sufficient treatment.

What matters more than re-excision in the mild-to-moderate group is follow-up. A patient with multiple dysplastic nevi warrants annual total-body skin examination by a dermatologist who knows their baseline. A patient with a single mildly dysplastic nevus, fully excised, with no other risk factors, warrants explanation and surveillance, not revision surgery.

What a patient should conclude

If you have received this diagnosis: it is not a cancer diagnosis. The word “dysplastic” is alarming in a way that is not proportional to what the diagnosis usually means. For most lesions in this category, the per-lesion risk of progression to melanoma is low.

The diagnosis suggests your skin warrants monitoring. This particular lesion is unlikely to become dangerous, but lesions with these features place you in a group for whom ongoing dermatological surveillance is appropriate. That is not the same as having a precancer.

Two questions worth asking your clinician: was the excision histologically complete, and what grade of atypia was reported? If the answers are yes and mild-to-moderate, the most defensible next step is surveillance. If the grade is high, or the margins were incomplete, re-excision is worth discussing.

The diagnostic problem is not going away

AI-assisted image analysis, molecular adjuncts (p16 immunohistochemistry, 9p21 FISH), and standardized reporting schemas such as MPATH-Dx are all being used to reduce the uncertainty in this diagnostic space (Barnhill et al., 2023; Vergara et al., 2025). They help, in selected cases. They have not resolved the fundamental issue: morphological grading of melanocytic atypia is a judgment, and judgments vary.

The honest position is that the dysplastic nevus diagnosis is more reliable at the poles, clearly low-grade or clearly high-grade, than in the middle, which is precisely where most clinically consequential uncertainty sits.

Understanding that is more useful than either dismissing the diagnosis or acting on it as though the grade were a precise measurement. It is not.

This post is educational. It does not refer to any real patient or clinical case. If you have received a pathology report and have questions about it, speak to the physician or pathologist who issued the report.

Disclosures. I am a histopathologist at AML (Antwerp) and full professor of pathology at the University of Antwerp. I have no financial interest in any diagnostic technology or vendor mentioned in this post.

References

Barnhill, R.L., Elder, D., Piepkorn, M., et al. (2023). Revision of the Melanocytic Pathology Assessment Tool and Hierarchy for Diagnosis Classification Schema for Melanocytic Lesions: A Consensus Statement. JAMA Network Open, 6(1), e2250613. https://doi.org/10.1001/jamanetworkopen.2022.50613

Duncan, L.M., Berwick, M., Bruijn, J.A., Byers, H.R., Mihm, M.C., & Barnhill, R.L. (1993). Histopathologic recognition and grading of dysplastic melanocytic nevi: an interobserver agreement study. Journal of Investigative Dermatology, 100(3), 318S–321S. https://doi.org/10.1111/1523-1747.ep12470215

Duffy, K., & Grossman, D. (2012). The dysplastic nevus: from historical perspective to management in the modern era: part I. Journal of the American Academy of Dermatology, 67(1), 1.e1–16. https://doi.org/10.1016/j.jaad.2012.02.047

Duffy, K., & Grossman, D. (2012). The dysplastic nevus: from historical perspective to management in the modern era: part II. Journal of the American Academy of Dermatology, 67(1), 19.e1–12. https://doi.org/10.1016/j.jaad.2012.03.013

Duray, P., DerSimonian, R., Barnhill, R., Stenn, K., Ernstoff, M., Fine, J., & Kirkwood, J. (1992). An analysis of interobserver recognition of the histopathologic features of dysplastic nevi from a mixed group of nevomelanocytic lesions. Journal of the American Academy of Dermatology, 27(5 Pt 1), 741–749.

Ferrara, G., & Argenziano, G. (2021). The WHO 2018 Classification of Cutaneous Melanocytic Neoplasms: Suggestions From Routine Practice. Frontiers in Oncology, 11. https://doi.org/10.3389/fonc.2021.675296

Gjersvik, P., Veierød, M., Thompson, A., Grzyb, K., Lilland, K., Vazov, N., Roscher, I., & Bassarova, A. (2022). Histopathologic reassessment of melanoma and other melanocytic skin lesions excised in 2009 and 2018–2019. Tidsskrift for den Norske laegeforening, 142(15). https://doi.org/10.4045/tidsskr.22.0204

Haggenmüller, S., et al. (2025). Discordance, accuracy and reproducibility study of pathologists’ diagnosis of melanoma and melanocytic tumors. Nature Communications, 16. https://doi.org/10.1038/s41467-025-56160-x

Hurwitz, R., & Tavel, M. (2015). The mythical concept and untoward consequences of a diagnosis of dysplastic nevus: an overdue tribute to A. Bernard Ackerman, MD. Dermatology Practical & Conceptual, 5, 31–34. https://doi.org/10.5826/dpc.0501a05

Menzinger, S., Merat, R., & Kaya, G. (2025). Dysplastic Nevi and Superficial Borderline Atypical Melanocytic Lesions: Description of an Algorithmic Clinico-Pathological Classification. Dermatopathology, 12. https://doi.org/10.3390/dermatopathology12010003

Ronen, S., Al-Rohil, R., Keiser, E., et al. (2021). Discordance in Diagnosis of Melanocytic Lesions and Its Impact on Clinical Management. Archives of Pathology & Laboratory Medicine. https://doi.org/10.5858/arpa.2020-0620-oa

Soleymani, T., Swetter, S., Hollmig, S., & Aasi, S. (2020). Adequacy of conservative 2–3 mm surgical margins for complete excision of biopsy-proven severely dysplastic nevi. Journal of the American Academy of Dermatology. https://doi.org/10.1016/j.jaad.2019.12.077

Ulgen, O., Yıldız, P., Acar, H., & Demirkesen, C. (2021). Analysis of interobserver reproducibility in grading dysplastic nevi: Results of the application of the 2018 World Health Organization grading criteria. Journal of Cutaneous Pathology, 49, 343–349. https://doi.org/10.1111/cup.14165

Vergara, R., Laharanne, E., De La Fouchardière, A., et al. (2025). Improving diagnostic accuracy in atypical melanocytic tumors using p16 immunohistochemistry and 9p21 fluorescence in situ hybridization: analysis of 206 second opinion cases. Scientific Reports, 15. https://doi.org/10.1038/s41598-025-95785-2

Wiedemeyer, K., Hartschuh, W., & Brenn, T. (2021). Dysplastic Nevi: Morphology and Molecular and the Controversies In-between. Surgical Pathology Clinics, 14(2), 341–357. https://doi.org/10.1016/j.path.2021.01.007

When we set up our pivotal trial at ElmediX, essentially two people were coordinating the clinical side from our end. Two. A CEO friend/colleague of mine runs a cardiovascular trial at a mid-sized pharma company with comparable patient numbers — similar protocol complexity, similar regulatory burden — and he has a dedicated team of twenty-two. Medical monitor, data manager, regulatory specialist, quality lead, site activation team, pharmacovigilance officer. Each owns one lane.

We had two people, who owned all of them.

That contrast is not exceptional in early-stage MedTech. It is the default. And it creates a hiring problem that I have been trying to describe clearly for three years: the people with the experience you need for a clinical operation are, almost without exception, people who learned to work in the twenty-two-person structure. Bringing them into the two-person structure is not a matter of scaling down. It is a different job.

Experience comes with assumptions

The appeal of hiring someone with ten years in clinical operations at a large MedTech or pharma company is obvious. They know the regulatory language, the ICH-GCP requirements, the site visit protocols, the audit trails. They have done this before. You are not teaching from scratch.

What you are also getting, always, is ten years of assumptions about how a clinical operation is organized. Clear lanes. Defined handoffs. A quality team that catches the things the clinical team missed. A regulatory affairs department one floor up. An IT system that someone else configured. A manager to escalate to higher management.

In a startup, none of that infrastructure exists. The person doing the site visit is also the person writing the SOP, also the person who will present to the ethics committee next week, also the person on the phone with the CRO at eight in the morning when something goes wrong. The work does not respect lanes, because there are no lanes.

Someone who has spent a career inside the lane structure is not poorly trained. They are trained for a different environment. The gap is not skill — it is operating assumption. And it shows up within the first three months, when the procedural comfort they expected is not there and they spend energy trying to rebuild the structure they came from, rather than adapting to the one they are in.

The supply problem on the other side

The obvious solution is to hire people who already know how to work in a small operation. And they exist. There are people who ran clinical trials for a four-person startup, who coordinated regulatory submissions without a dedicated RA team, who learned to hold fifteen things at once because there was no one else to hold them. Those people are competent in ways that do not appear on a CV, because the CV lists the job title, not the actual workload.

The problem is that those people do not circulate. If you joined a MedTech startup five years ago, and it worked, and you grew with it, you are probably still there. The startup is now twelve people instead of four, and you are the person the new hires come to when they need to understand how things actually work. You are not on the market.

The people on the market are, statistically, more likely to be coming from the larger structure. Which takes us back to the gap.

What hierarchy does and does not do

I want to be precise about what I am criticizing, because it is not hierarchy itself.

A twenty-person clinical operations team at a large company is well-structured for a reason. Lane specialization produces quality in each lane. Handoff protocols prevent things from falling between people. Escalation paths exist because complex trials generate decisions that cannot sit with one person. That system is not bureaucracy for its own sake. It works, for the scale and context it was built for.

The problem is that it produces people for whom structure is invisible — present the way air is present. When the structure is removed, the disorientation is real. I have watched it happen: someone genuinely excellent in a previous role spent their first six months trying to recreate the approval flows and documentation hierarchies they had left behind, rather than doing the actual work the job required. The instinct was not wrong. It was trained. But it was trained for a different environment.

An excellent CRO is what makes the two-person team work

I have described what the two-person team does not have. I should also describe what makes that team functional anyway, because in practice it is not stubbornness and overtime that carries the trial. The single most important external relationship is the CRO.

A good CRO brings the lane structure that does not exist inside the startup. Their medical monitor, their data manager, their regulatory specialists. The twenty people are there. They sit on the outside. The two-person sponsor team coordinates them but does not have to be them. That is the multiplier.

An excellent CRO does more than execute. They challenge the protocol when something is unclear, surface site-level problems before they become regulatory ones, and translate between the tempo of a small sponsor and the formal world the trial has to live in. At their best they behave as a temporary extension of the sponsor team, taking ownership where they could have hidden behind a statement of work.

A mediocre CRO executes what is written and nothing more. With a two-person sponsor team, that is not enough. The sponsor ends up doing both the sponsor work and the work of supervising the CRO’s lanes, which defeats the purpose of having a CRO. Selecting the right CRO, and maintaining that relationship so they choose to act as a partner rather than a vendor, is one of the most consequential operational decisions a small MedTech CEO makes. I would put it ahead of most internal hires.

Consultants, and the question of whose work they are doing

Consultants sit in a related but separate category. The same hiring gap applies to them, with one added ambiguity. Consultants are paid by you, formally, but the question of whose work they are actually doing is not always cleanly answered.

Some consultants work as if they were on your payroll. They internalize the problem, take on tasks beyond their statement of work, and treat the trial as theirs for the duration of the engagement. When a question comes up at nine at night, they answer it. They do not bill differently for things outside the scope, because in their head the scope is whatever it takes to move the project forward.

Other consultants, in equally good faith, stay mostly inside their own framework. They bring the methodology they have built over a career, their preferred suppliers, and the playbook that has worked for them before, and they apply it to your problem. The work is often competent. The directionality is slightly different: the consultant is on your payroll, but a meaningful part of the work also serves the consultant’s own practice and longer-term agenda.

Both versions are real, and neither is dishonest. The first is what a small operation actually needs for the parts of the work that require ownership. The second is well-suited to a defined slice of work with a clear deliverable. The trouble is that engagements are often described as the first and, in practice, look more like the second. The mismatch is rarely visible in the first conversation. It tends to show up around week six, when the consultant either steps into a gap that is not formally theirs, or politely declines to.

I have learned to ask, in the first conversation, what a consultant does when something falls outside the contract. The answer is usually informative.

Three rules I actually try to follow now

One: be explicit about the structure that does not exist. Not in a discouraging way, but specifically.

Two: look for people who have worked in more than one size of organization, even briefly. A year at a small company during or after a large-company stint is more predictive than most of the rest of the CV. It means they have felt the difference and chosen to continue.

Three: watch the first six weeks. The people who adapt well spend those weeks learning how things actually work here, even when how things work here is imperfect. The people who struggle spend those weeks explaining how things worked somewhere else.

What does not resolve neatly

There is a version of this problem I have not solved. The people best suited to work in a small MedTech startup are often people with limited MedTech experience — smart, adaptable generalists who learn fast and do not carry the wrong assumptions. But regulatory and clinical work in MedTech has a real knowledge base. A fast learner with no GCP background is not ready to run a trial. The domain knowledge is necessary. The domain knowledge comes with assumptions. The assumptions create friction.

I navigate this case by case, and I get it wrong sometimes. I thought it was worth saying out loud. And... Luckily I have found the perfect CRO and the perfect project management consultant ....

Disclosures. I am CEO and shareholder of ElmediX, a Belgian MedTech company developing whole-body hyperthermia for metastatic pancreatic cancer. The examples in this post are drawn from that experience. No employee or colleague is identified individually.

One hundred and seventy-nine years ago today, on 14 May 1847, Fanny Mendelssohn collapsed at the piano. She was rehearsing one of her brother Felix’s choral works for the next Sunday’s house concert. A stroke. She died that evening, forty-one years old. Felix outlived her by six months; most of his friends thought the grief killed him.

I am not a musicologist. I play piano at a level that keeps me honest about how hard it is, and I listen to a lot of music. Today felt like the day to write about her, and to write about one piece in particular: Mai, the fifth movement of her piano cycle Das Jahr (”The Year”), composed in Berlin between August and December 1841 (Bartsch, 2000).

A cycle of twelve months

Das Jahr is the work most musicologists now consider her magnum opus (Todd, 2010; Rodgers, 2010). Twelve character pieces for piano, one for each month, plus a closing Nachspiel: thirteen movements in all, around fifty minutes of music. Each month, in the “schöne Handschrift” reinschrift of 1843, carries two extra things on the page besides the score: a motto-poem chosen by Fanny, and a watercolour painted by her husband Wilhelm Hensel.

This matters. Wilhelm was a Prussian court painter, not a Sunday hobbyist. He had supported her composing from the day they married in 1829, and Das Jahr in its finished form is a piece of total art before the word existed: music, poetry, and image bound into a single object, a decade before Wagner started using Gesamtkunstwerk as a slogan.

Mai

The motto for Mai is Ludwig Uhland’s Frühlingsglaube, “Spring Faith”. Uhland was a Swabian poet from Tübingen, the same Romantic generation as Eichendorff. The poem is short, two stanzas, written in 1812. The text turns on a single image.

Die linden Lüfte sind erwacht,
Sie säuseln und weben Tag und Nacht,
Sie schaffen an allen Enden.
O frischer Duft, o neuer Klang!
Nun, armes Herze, sei nicht bang!
Nun muss sich alles, alles wenden.

Die Welt wird schöner mit jedem Tag,
Man weiß nicht, was noch werden mag,
Das Blühen will nicht enden.
Es blüht das fernste, tiefste Tal:
Nun, armes Herze, vergiss der Qual!
Nun muss sich alles, alles wenden.

The line I keep returning to is the one in bold: Es blüht das fernste, tiefste Tal — “the farthest, deepest valley blooms”. Not the meadow. Not the garden. The remote one. The one nobody walks to. Even that valley blooms.

Uhland published Frühlingsglaube in his Gedichte (1815). The poem entered the German repertoire quickly: Schubert set it in 1820, Konradin Kreutzer’s setting circulated almost as widely, and Liszt later transcribed Schubert’s version for solo piano (S. 558/7). The poem’s pull comes from the asymmetry between the two stanzas. The first stanza notices that the world is waking up. The second stanza addresses the heart and orders it to wake up too. The hinge between the two is that one image, the farthest deepest valley, which works as a metaphor either way: the unreachable corner of the earth, or the unreachable corner of yourself.

Fanny’s setting is Allegro vivace, restless under a sunny surface. She does not paint May in postcard colours. The right hand chases a figure that keeps almost arriving and then sliding sideways; the left hand grounds it in steady motion. About two-thirds through, the music opens into a calmer middle section that hangs on a sustained upper voice. To my ear, that is where the deepest valley blooms. The piece resolves without showing off about it.

A Schubert hinge

Twenty-one years before Fanny, Franz Schubert had set the same Uhland text. Frühlingsglaube, D. 686, written in 1820, published as Op. 20 No. 2 (Reed, 1985). Schubert’s reading is slower, more grateful, more private. He gives the famous line a long held note in the voice over a falling piano line, and the harmony bends momentarily into something almost devotional.

Two of the great Lieder composers of the nineteenth century, working with the same poem within a generation of each other. The Schubert is a song for voice and piano; Fanny’s Mai is a piano piece that has the poem standing behind it, unsung. If you have half an hour, listen to them back to back: Lauma Skride on Sony or Gaia Sokoli on Piano Classics for Fanny’s Mai, Christa Ludwig on Deutsche Grammophon for the Schubert. They are doing the same job from two angles, a song that contains the words and a piece that has had the words pulled out and replaced with notes.

Gesamtkunst, in private

It is tempting to say Das Jahr anticipates Wagner. It does not. Wagner’s Gesamtkunstwerk is a public, theatrical, ideological project: opera as the rebuilding of culture. Das Jahr is a private artefact. It was written for the Mendelssohn family’s Sunday concerts and the autograph stayed in Wilhelm’s hands; the first complete publication waited until 1989 (Borchard, 1989; Furore-Verlag), 142 years after Fanny’s death.

But the principle is there, smaller and earlier and arguably purer: three art forms made by one couple, sitting on a single page. Sound, image, text. Wilhelm’s Mai watercolour in the autograph does not paint May the way you would expect. There is no meadow, no path, no obvious sunlight. The page shows half-formed figures suspended above the line: drifting, longing, slightly unfinished. It is closer to a dream than to a landscape, and that is exactly what Das Jahr is doing as a whole: not depicting the year, but describing how a year feels from the inside. The watercolour for Mai sets the mood the music then carries. Read the poem first. Look at the figures. Then play, or have someone play. The piece changes shape.

A small thing to do today

If you have a piano, the skill (I don’t) and a few minutes (probably rather weeks for me), play Mai. If you do not, the Skride recording (Sony, 2007) is the easiest first listen, and the Biesemans fortepiano recording (Pan Classics, 2018) is the most historically truthful. Whichever you choose, the line to keep in mind is Uhland’s, not Fanny’s, though by 1841 it had become hers as well.

Es blüht das fernste, tiefste Tal.

Even that valley blooms.

— John-Paul

Bronnen / References

• Bartsch, C. (Ed.). (2000). Fanny Hensel: Das Jahr — Zwölf Charakterstücke (1841) für das Forte-Piano, kritische editie. Furore-Verlag.

• Borchard, B. (Ed.). (1989). Fanny Hensel: Das Jahr. First complete publication. Furore-Verlag.

• Reed, J. (1985). The Schubert Song Companion. Manchester University Press (entry on D. 686).

• Rodgers, S. (2010). Fanny Hensel’s Das Jahr and the cyclic imagination. 19th-Century Music, 34(2), 191–217.

• Todd, R. L. (2010). Fanny Hensel: The Other Mendelssohn, ch. 8. Oxford University Press.

• Toews, J. E. (1993). Memory and gender in the remaking of Fanny Mendelssohn’s musical identity: The chorale in Das Jahr. The Musical Quarterly, 77(4), 727–748.

• Uhland, L. (1815). Frühlingsglaube, in Gedichte. Cotta.

The manifesto went up a few days ago. Within a few hours the first email arrived: “So which tools are you actually using?” Fair question. People talk about AI fatigue, and I understand it. Every week there is a new product that promises to compress a year of work into an afternoon. Most of them I look at for ten minutes and forget.

But five sit on my credit card every month. Three I open daily and think about. A fourth sits so deep in my workflow that I almost forgot it while drafting the math at the top of this post, which is itself an interesting fact about how that tool has worked out. A fifth I pay for and barely open. Two others sit on my desktop, free, and I keep walking past those as well. I have not cancelled anything; I just have not gotten anywhere with them. Below is what I pay for, what it costs, and what comes back per euro spent. Honest amateur-from-use, not a vendor review.

The math first

Five subscriptions, what I actually pay:

• Claude Max: $100 per month, $1,200 per year.

• Elicit Pro: $49 per month, $588 per year.

• Consensus Pro (annual): $107 per year, about $9 per month equivalent.

• Perplexity Pro (annual): $200 per year, about $17 per month equivalent.

• ChatGPT Plus: $17 per month, $204 per year (legacy price — I have had this since early on and suspect the current rate is higher).

Total: around $2,300 per year. That is roughly the price of one trip to a small European conference, plus a decent dinner with the speakers. The question is what each one earns back.

1. Claude Max + Cowork: the workhorse

Claude is the one I would keep if I could only keep one. I run the desktop app in Cowork mode, which matters more than the price tag suggests. Cowork lets the model work directly in a folder on my computer: read files, write files, edit files. It is not a chat window shouting answers at me. It is a colleague that opens the cabinet and rearranges the shelf.

What it does, day to day, is maintain what I call my Obsidian Robot (thanks to David Sparks, “MacSparky”). A vault of around a thousand files that thinks and writes the way I do: identity files, role files, writing-style rules, folder hygiene, skill definitions, transcripts of past meetings, draft chapters, reading lists. Every morning Claude reads through the relevant parts and produces a daily briefing: what is on the calendar, which manuscripts are mid-edit, which PhDs are waiting on me, which deadlines are about to land. It pulls from the actual files, not from a generic prompt.

After that, it helps with whatever sits in front of me. Some concrete examples from the past two weeks:

• A 90-minute Teams meeting on a clinical workflow protocol. The transcript went in (raw, with the usual misheard names), and a clean meeting report came out: decisions, action items, open questions, with each colleague’s name corrected automatically because the vault knows the transcript tends to mangle them.

• A manuscript a senior PhD student is finishing. Track-changes edits in the shape of my voice rather than a generic editor’s. The student then accepts or rejects each one; my job is to mark what is unclear, not to rewrite their thinking.

• A letter of support for one of my African PhDs applying for a fellowship. Three drafts before the tone was right, all of them shaped by what the vault already knows about the candidate.

• A short email to my assistant, asking her to find a slot for a thesis defence review. Claude drafts it in my voice and waits for my approval. I have a hard rule that nothing leaves on my behalf without that approval, and Cowork respects it.

• This very post, including the cleanup of my vault that happened yesterday when I complained it had become messy. Half a thousand files in the wrong folders, an outdated identity document, an empty home page. Eighty minutes later all four problems were resolved. I would have spent a Sunday on it manually, if I had spent the Sunday at all.

The limit I run into is context length on heavy days. Long PDF in, summary out, draft on top — three steps that sometimes hit the wall, and I split the work. The Max plan eases the wall, it does not remove it. I would also like better handling of slide images; for now image work goes through other channels.

If you write or read or organize for a living, this one carries the most weight.

2. Elicit Pro: for structured literature work

I use Elicit Pro for one job: structured literature work. A recent example is a background literature update for a manuscript I am co-supervising on hyperthermia and cytokine response. Background sections in this field age fast: what looked like the picture two years ago has gained three new safety datasets and lost one assumption. A current map is needed before the introduction can be written honestly.

The old workflow was three days in PubMed with a spreadsheet, a colleague to cross-check the search string, and a stack of printed abstracts on the corner of the desk. With Elicit Pro it goes like this. I define the question: what is the published evidence on cytokine response under controlled hyperthermia in solid tumors, with attention to temperature, duration, and tumor type? I run a report. Back comes a structured table of papers with the columns I asked for: sample size, intervention parameters, primary outcome, effect size, study design, country. I read the originals. Elicit does not replace that reading; it replaces the search-and-sort step that used to eat the first two days of every review. Financial-interest disclosure at the bottom of this post; read it.

A second use case from last month. One of my PhD students in South Africa is writing a systematic review on cervical-cancer-screening uptake in low-resource settings. The old pattern would have been: she sends drafts, I mark up the methodology, we iterate over weeks. Instead we did the first pass together in Elicit, side by side. She defined the question; I pointed out which columns she had forgotten and why. The point was less the speed and more the supervision. She now sees how I think about inclusion criteria, not just what I eventually conclude. That is closer to how I want to mentor: visible reasoning, not just verdicts.

A third, smaller use case. Before a thesis-committee meeting for one of my Belgian PhDs working on hyperthermia and the immune system, I run a quick scoping report in Elicit on whatever the student frames as their original contribution. Not to catch them out, but to know what the panel is likely to ask. It saves me an evening of reading and makes my questions sharper.

Pro at $49 per month gives me twelve reports per month, more than I use. The lower plan caps at a level I outgrow within a week of any active project. Honest: if you run one systematic review every two years, you do not need Pro. If you are running a trial, supervising PhDs running them, or preparing a grant, you do.

What I do not trust Elicit for is anything outside its literature index. It is a research-assistant on papers, not a general reasoner. I keep separate tabs open.

3. Consensus Pro: the in-meeting answer

Consensus is the cheapest of the five and the one I open most often when I need an academic answer, because the questions I take to it are small and frequent. A single point. Usually inside a meeting, sometimes mid-paragraph while editing, sometimes between calls when a colleague’s question is still ringing in my head.

Three real questions from the past month, with what I did with the answer:

• A reading-group discussion on a recent paper. The question that came up: has IL-15 upregulation under hyperthermia been demonstrated in solid tumors specifically, as opposed to hematological ones where most of the early work sits? Five minutes with Consensus gave me four relevant papers, ranked, with one-line summaries of where the evidence sits. I scanned the abstracts during the meeting and could contribute without bluffing. The conversation moved on with a sharper basis.

• A draft of a slide I was preparing for a regional oncology meeting. I had written that whole-body hyperthermia “stimulates the immune system,” a phrase too loose for anyone who actually reads the literature, and a sharp questioner from the audience will catch it. Consensus brought me four mechanistic papers that justify a tighter version of the claim, the one I am willing to defend. The slide was rewritten the same evening before the deck went out.

• A reviewer comment on a manuscript: “your introduction overstates the prevalence of high-risk HPV in sub-Saharan settings.” Was the reviewer right? Forty-five seconds in Consensus, three papers, the answer was yes, partial but yes. The introduction was rewritten, the reviewer was thanked in the response letter, and the paper was stronger for it.

What Consensus is not, is a synthesis tool. It points at the papers; it does not draw the picture. For anything that needs a defensible conclusion I still go to the originals. The value is in skipping the twenty minutes I used to spend setting up a PubMed search to land on the one paper I actually needed.

Pro on the annual plan is $107 for the year, about $9 per month equivalent. A free tier with twenty searches per month exists and is usable if you only reach for it occasionally. I take the annual plan because I use it many times that.

4. Perplexity Pro: the one that has become invisible

Honest correction. While drafting the math at the top of this post, my first list had three subscriptions. There is a fourth, and I almost missed it because it has dissolved into how I work. Perplexity Pro, $200 per year on the annual plan, about $17 per month equivalent. It has completely replaced Google search for me. I use it more often than Consensus, more often than anything else on the list, multiple times a day.

The simple test: I cannot remember the last time I typed into google.com. Months at least. When I want a fact, a definition, a current price, the name of a tool somebody mentioned in a meeting, a news headline, a quick directional read on a regulatory body, an opening hour, an address, the name of a piece of music I half-recognize, anything that the old me would have googled, I now open Perplexity. It gives me a short, sourced answer with the links underneath, in the time Google takes to load its ad-heavy result page.

Three uses from yesterday and today, of the kind I would not normally write down because they are too ordinary:

• A clinical-trials.gov detail I needed for a question about competing trials in a tumor type adjacent to my own.

• The Brussels address of an organization whose name I had only half-remembered from a recent meeting.

• A check on whether a specific journal sits on the predatory list a colleague had sent me by email, before I told a student to consider submitting there.

For deep literature work I still go to Consensus or Elicit, where the index is curated to peer-reviewed sources and I can rely on the ranking. One caveat, and it matters: search the same question twice with differently framed prompts — one sceptical, one credulous — and Perplexity will tend to confirm whichever framing you bring. I have noticed this when investigating a regulatory question from two angles; the second search was noticeably more supportive of the second framing. For anything with real consequences, verify with Elicit or go to the original literature. For everything else, Perplexity. It is the most invisible of the five because it does not feel like an AI tool. It feels like a fixed search bar that gives better answers than the one I used for twenty years.

If you have not tried it and you have read this far, take twenty dollars and run a one-month experiment. My prediction is that you will not go back. Mine is the strongest recommendation in this whole post, even though I almost forgot to make it.

The three I keep walking past

Google Gemini. I tried it through Gmail, where Google pushes it hard. The integration looks promising on paper: read the inbox, summarize threads, draft replies. In practice the answers were often wrong or beside the point. A meeting request from a CRO project manager was filed under “newsletter”. A thread-summary about a regulatory submission missed the actual ask, which sat in the third paragraph of a long mail. A draft reply to an African collaborator opened with a tone I would never use, friendly to the point of slick. After a few rounds I stopped opening it. It cost me more in re-checking than it saved. It may improve, and I will try again. For now, no.

NotebookLM. This is the one I want to like. The idea fits how I work: upload your sources, ask questions inside the boundary of those sources. The interface, though, never clicks for me. The few times I sat down to use it (uploading a stack of hyperthermia papers, trying to ask a comparative question across them) the answers were fine, but the friction was real. And I noticed I could do the same thing in Claude Cowork with my Obsidian vault as the source, plus everything else Cowork lets me do, including writing back, editing, and running small scripts. The marginal value of opening a second tool was not there. I keep it installed, in case the interface improves, but my hand never reaches for it.

ChatGPT. I use it rarely, and I cannot fully explain why. I do pay for it — ChatGPT Plus, $17 per month at a legacy price I have had since early on. It is a capable model, competitive in most benchmarks I have seen, and it was the first AI I ever heard of, which at least means I have no excuse of ignorance.

Part of the explanation is probably habit: I arrived at Claude before ChatGPT had time to form one. Part of it may be that the Cowork integration fits my workflow tightly enough that switching to another model for the same tasks offers no clear gain.

What has changed recently is Codex, OpenAI’s coding agent. I do not know yet whether it shifts my behaviour. A proper comparison — Codex versus Claude Cowork, running the same tasks I actually use — is something I intend to write about. Until then, ChatGPT sits installed and rarely opened.

None of the three is cancelled. Gemini and NotebookLM cost nothing. ChatGPT costs $17 per month and I keep not cancelling it. They all sit unused, which is its own answer.

The rule I follow

After three years of trying every tool a colleague mentions, one rule has held: a tool earns its place only if it fits a workflow I already have. The question is not “which AI is the best.” It is “which AI lets me stay in my workflow instead of forcing me to rebuild around it.”

Claude with Cowork sits inside the Obsidian vault I already use. Elicit sits inside the systematic-review process I already run. Consensus sits inside the in-meeting-question flow that already happens to me twenty times a week. Gemini wanted me to live in Gmail. NotebookLM wanted me to live in NotebookLM. Neither is a workflow I want.

If a tool requires me to change how I work to get its benefit, the bar is much higher. I keep paying for the ones that fit.

Where I am still uncertain

A few things I have not solved and would welcome notes on:

• Reference manager + AI. I use Zotero. Nothing I have tried yet integrates cleanly without a clunky export-import step. If you have a setup that works, tell me.

• Bandwidth-light AI for PhDs in Kisumu or Addis. Most of my students work over connections where a 200 MB model download is a problem. Almost everything I use assumes broadband.

• AI for slide review. My core craft is pathology slides. The current image-analysis tools are not at the level I would trust for a clinical decision, and I have stopped wasting time on them. If that changes, I will write about it.

Replies welcome in the comments. I read them.

Disclosures. I pay for Claude Max ($100/month), Elicit Pro ($49/month), Consensus Pro on the annual plan ($107/year), Perplexity Pro on the annual plan ($200/year), and ChatGPT Plus ($17/month, legacy price). I have no commercial relationship with any of these vendors and have not received complimentary access from them. Separately, I am CEO and shareholder of ElmediX, a Belgian MedTech company developing whole-body hyperthermia, which sets the field context for several examples in this post. None of this affects my views on the AI tools, but you deserve to know it.

Why a pathologist starts a Substack at 61

I have spent 28 years looking at slides under a microscope. Most of what I have learned does not make it into papers. Papers reward findings that can be replicated. The things that actually matter for a working clinician often resist replication. The case where an algorithm flagged the wrong cell and a senior eye caught it or like the Kenyan PhD student who steered her project into territory I could not have seen from Antwerp.

This Substack is for those things.

A short introduction. I am John-Paul Bogers, full professor of histopathology at the University of Antwerp, pathologist at AML in Antwerp, and CEO of ElmediX, a Belgian MedTech company running a pivotal trial of whole-body hyperthermia in metastatic pancreatic cancer. From 2016 to 2020 I directed the Belgian National Reference Centre for HPV; cervical-cancer screening in sub-Saharan Africa has been my second research line for two decades. Outside work I read history and espionage thrillers, listen to classical music on more equipment than I should admit, and play the piano at a level that keeps me honest.

What I will write about

Four recurring threads.

1. Pathology, for colleagues and for patients. Notes from 28 years at the microscope. Gynaecological cytology and histology, HPV biology and cervical screening, occasional dermatopathology. Most posts in this thread are written for working pathologists and the clinicians who refer to us. Once a month a different post is written for patients and their families: a constructed biopsy or smear report, translated line by line. No real patient data, no individual medical advice.

2. AI tools as a working researcher uses them. I am not an expert on AI in pathology. I am a senior researcher who pays for Claude Cowork, Consensus, Elicit, NotebookLM, and others, and who uses them daily for writing, mentoring, literature work, and running a company. What works in real workflow, what wastes time, what surprised me. Amateur enthusiasm under a senior-researcher’s eye.

3. Operating a MedTech startup as an academic. Concrete observations from running MATTERS-2: family-office fundraising, investigator recruitment, the parts of a pivotal trial nobody writes about until they have lived through them.

4. Mentoring, and research from the global south. I supervise PhD students in Belgium, Kenya, and South Africa. Three of my African mentees are now mentors themselves with their own students. I will write about how that happened, and why most North–South research collaborations still get this wrong. Cervical-screening implementation in the Lake Victoria basin, what the data actually show against what the WHO targets imply, sits in this thread.

Around these four threads I will sometimes write about books that earned their pages, music I have analyzed, or a piece of nineteenth- or twentieth-century history that has stayed with me. Roughly 70% of posts sit in the professional threads above; 30% do not. If you are subscribing for one and not the other, I will not be hurt if you mute the rest.

What I will not do

I will not pretend to be impartial about things I have an opinion on. When I write about hyperthermia, I will state where my financial interest lies and what I think anyway.

I will not write hype. I have spent a career around medical claims that did not survive replication. I am tired of the language that surrounds them.

I will not fabricate citations. If I cite a paper, the paper exists and says what I claim it says.

I will not write to fill space. Some weeks I will have something worth saying. Some weeks I will not. I would rather post less than post worse.

Practical

Two posts a month for the first six months. If the project sticks, I will move to weekly. Everything is free for the first six months. After that, roughly half the posts stay free; the other half move behind a paid tier.

If you read scientific papers professionally, you will recognise the citation style. If you do not, the prose is meant to be plain enough that it does not matter.

One last thing

I am writing this because the things I find most worth saying do not fit anywhere else. Not in the manuscripts I write with my students. Not in the investor decks I write with my co-founder. Not in the reports I sign at AML. They sit in between, and they have for years. This is where they go now.

Thank you for being here.

— John-Paul