AI & Medicine Advanced ⏱ 32 min AI in medicineAGIASIAlphaFoldDemis HassabisDario Amodeisurgical roboticsdiagnostic AIphysician careersIsomorphic Labsmedical superintelligenceclinical AI

The Compressed Career: What AI Means for Physicians and Surgeons

Published 2026-07-04 — Dr Neal Aggarwal
🎬 Watch the companion video: The Compressed Career — a visual walk through the three waves, the replacement scenarios, and the retraining path.

Next in this series: Learning With Dr Neal — the fast.ai Lesson 3 spreadsheet exercise, the concrete on-ramp referenced in Part IV of this essay.

In October 2024, Dario Amodei, the CEO of Anthropic, published an essay called "Machines of Loving Grace." Its central claim was narrow and enormous at the same time: if powerful AI arrives on anything like the timeline its builders expect, it will compress into five to ten years the biological and medical progress that would otherwise have taken fifty to a hundred. Amodei — who holds a PhD in biophysics from Princeton and did his postdoctoral work at Stanford's medical school — was not writing as a science-fiction futurist. He was writing as someone who has spent a career at the intersection of biology and machine learning, describing what he thinks is the modal outcome, not the optimistic tail.

Amodei is not the only person making this case, and he is not the one with the strongest empirical track record behind it. In 2024 the director Greg Kohs released The Thinking Game, a documentary filmed over five years inside Google DeepMind, following Demis Hassabis and his team through the work that produced AlphaFold — the system that solved the fifty-year-old protein-folding problem and earned Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. I raise it here, at the outset, because it is the closest thing we have to a documented, real-world proof of concept for exactly the kind of compression Amodei describes. AlphaFold did not make an existing task incrementally faster; it collapsed a problem that had resisted the entire structural-biology community for half a century into a solved one, and it did so with a general learning method rather than a hand-built, problem-specific one. When the film was released free on YouTube in late 2025 and drew north of two hundred million views, it did something an essay cannot: it showed the compression happening, in the room, rather than merely arguing that it would.

Amodei's list of what a "compressed twenty-first century" might contain is worth stating plainly, because it is the frame this whole essay sits inside: the elimination of most infectious disease, the effective cure of most cancer, the prevention of Alzheimer's, effective treatment of most genetic disease, and a doubling of the human healthy lifespan. Read that list again slowly. Every item on it is currently a life's work for a subspecialty of medicine. Amodei is proposing that AI collapses several centuries of subspecialty life's-work into a single decade.

This essay asks the question Amodei's essay does not directly answer: what happens to me — the physician, the surgeon, the person who has spent forty years of practice and teaching inside that system — while all of this is compressing? Not what happens to disease. What happens to the profession that treats it.

I have tried to write this at the level of seriousness the subject deserves: grounded in the actual research and deployment data available as of mid-2026, honest about what is genuinely uncertain, and specific about what I can do about it. This is not a doom essay, and it is not a reassurance essay. It is closer to a terrain map for a landscape that is still actively shifting under my feet even as I write it.

Part I: Three Waves, Not One Event

The mistake almost everyone makes when thinking about "AI and medicine" is treating it as a single threshold to be crossed — the day the AI becomes "as good as a doctor," after which everything changes at once. That is not how this is unfolding, and it is not how it will finish unfolding. It is more useful to think in three overlapping waves, each with a different character, a different timeline, and a different implication for how I should be spending my continuing education hours right now.

Wave One: Narrow AI and LLM Copilots (roughly 2023–2028, already well underway)

This wave is not coming. It is here, and the data on it is no longer speculative.

The U.S. FDA had authorized 1,451 AI-enabled medical devices by the end of 2025 — up from 221 in 2023, a more than sixfold increase in two years. Seventy-six percent of those clearances are in radiology. In January 2026, Aidoc received clearance for a foundation-model-powered clinical AI system covering fourteen conditions simultaneously, with a mean sensitivity of 97% and mean specificity of 98% across that panel — numbers that would represent an excellent attending radiologist on a good day, delivered at machine consistency, every time, for every study.

The diagnostic story is, if anything, more startling outside radiology. In mid-2025, Microsoft AI published results for a system it calls MAI-DxO — a "diagnostic orchestrator" that coordinates several frontier language models (drawn from OpenAI, Google, Anthropic, Meta, and xAI) to loosely simulate a panel of physicians working a case together. Tested against 304 diagnostically complex cases adapted from the New England Journal of Medicine — the kind of cases used to teach residents precisely because they are hard — MAI-DxO reached correct diagnoses in up to 85.5% of cases. Twenty-one experienced physicians, working the same cases under matched restricted conditions (no colleagues, no internet, no reference texts — an admittedly artificial constraint that does not reflect how real physicians work), averaged 20%. Microsoft titled the accompanying research note "The Path to Medical Superintelligence." That is not a subtle title.

I want to be honest about the limits of that study, because overclaiming here would be a disservice to the argument: the case mix was deliberately unusual and complex, excluding routine and mild presentations; the physicians were denied resources they would have in real practice; and the comparison was to generalists, not specialists in the relevant field. A fair reading is not "AI beats doctors at diagnosis," but rather: on a hard, resource-constrained diagnostic reasoning task, a well-orchestrated ensemble of frontier language models now substantially outperforms unaided physician recall. That is still a landmark result, and the gap is not one that is likely to close in medicine's favor as the underlying models keep improving on their normal trajectory.

The other visible effect of Wave One is on my working life directly, and it is currently the most consequential thing happening to physicians day to day: burnout. AI ambient scribes — tools that listen to a consultation and generate a structured note — are cutting charting time by up to 75% in early deployments, and one health system reported physician burnout scores falling from 51.9% to 38.8% within thirty days of adopting an AI scribe. The most recent AMA physician burnout survey puts the national rate at roughly 42%, down meaningfully from a pandemic-era peak of 62.8%. If I take nothing else practical from Wave One, it is this: the single highest-leverage AI adoption available to me right now, with essentially no downside, is an ambient documentation tool. This is not the frontier of the argument this essay is making, but it is the part I can act on this week.

Surgical robotics belongs in Wave One too, in its current form. The Medtronic Hugo RAS system received FDA clearance in December 2025 for urologic procedures — prostatectomy, nephrectomy, cystectomy. Across AI-assisted robotic platforms generally, published figures show roughly a 25% reduction in operative time, a 30% decrease in intraoperative complications, a 40% improvement in measured surgical precision, and 15% shorter patient recovery. Note the framing carefully: assisted. A human surgeon is still the operator of record on every one of these systems. That distinction is the hinge on which the rest of this essay turns, and it will not hold indefinitely — see Part III.

Wave Two: AGI (contested timeline, plausibly 2027–2030)

This is where the essay has to get honest about disagreement, because the people building these systems do not agree with each other, and no one should trust anyone — myself included — who claims to know the date with confidence.

Anthropic's own formal submission to the U.S. Office of Science and Technology Policy states the company expects "powerful AI" — its preferred term, roughly equivalent to what others call AGI — in late 2026 or early 2027. Sam Altman has said AGI will "probably" be developed within the current U.S. presidential term, and has more recently talked about the conversation moving past AGI toward superintelligence. Demis Hassabis at Google DeepMind is more conservative: roughly 50% odds of AGI by 2030, though notably his own estimate tightened from a five-to-ten-year window in 2024 to three-to-five years by January 2025 — the forecast is shortening, not lengthening, as these things go. Shane Legg, DeepMind's chief AGI scientist and one of the people who coined the term, puts roughly 50% odds on "minimal AGI" by 2028. Alexandr Wang, CEO of Scale AI, has predicted two to four years to what he calls "remote worker AGI" — systems that can sit down at a computer and do a real job the way a competent employee would.

Hassabis is worth listening to on this more carefully than most, and not only because of the Nobel. The through-line of The Thinking Game is that he does not treat AGI as the finish line at all. His stated strategy is to break off pieces of the road to general intelligence and turn them into narrow, superhuman scientific tools long before the general system arrives — AlphaFold being the first and clearest instance of that strategy actually working. This matters for the timeline argument here because it means the medically relevant disruption does not wait for AGI to be formally declared. Each specialised system — an AlphaFold for protein structure, another for molecular design, another for a class of diagnoses — lands on its own schedule, and several of them are already on the ward.

Strip out the branding differences and a real pattern remains: essentially every senior person actually building these systems is now expressing a timeline measured in single-digit years, not decades, and the trend across the last eighteen months has been toward earlier estimates, not later ones. I hold this with appropriate uncertainty — forecasting AGI has a long history of being wrong in both directions — but "appropriate uncertainty" in mid-2026 does not mean "safe to ignore for another twenty years." It means treating this the way a competent physician treats a patient with an ambiguous but concerning set of symptoms: you do not wait for certainty before you act, because by the time you have certainty, several of your options have already closed.

What does AGI actually mean for me, concretely, if it arrives on anything like this schedule? Not a single robot doctor. It means the generalist reasoning gap between a frontier AI system and a well-trained specialist closes across essentially every cognitive domain of medicine simultaneously — diagnosis, treatment planning, literature synthesis, differential construction, risk stratification, patient communication drafting — not because any one system was built specifically for medicine, but because medicine is, at its cognitive core, a data-processing and pattern-recognition discipline running on a general reasoning substrate, and that substrate is what AGI targets directly. The Anthropic Economic Index — which tracks how AI is actually being used across occupations, not just what it is theoretically capable of — already puts healthcare practitioners at 59.9% theoretical AI capability exposure, among the highest of any occupational category. The gap between that theoretical exposure figure and actual observed usage today (currently highest in computer/mathematical and office/administrative occupations) is precisely the gap Wave Two closes.

On the procedural side, Wave Two is where the surgical robot stops being an assistant. In 2025, a Johns Hopkins-led team demonstrated a system called SRT-H performing a complete laparoscopic cholecystectomy — gallbladder removal, a seventeen-step procedure — on realistic tissue with no human intervention. This is a meaningfully different achievement from its 2022 predecessor, STAR, which needed specially marked tissue and a rigid, pre-programmed surgical plan executed in a tightly controlled environment. SRT-H learned by watching video recordings of human surgeons performing the procedure — imitation learning, architecturally similar to how large language models learn — and it demonstrated the ability to adapt in real time to individual anatomical variation and self-correct mid-procedure when something did not match its training distribution. Johns Hopkins reported 100% procedural success across its trials and won a 2025 RBR50 robotics innovation award for the work. This is not AGI. It is a narrow, single-procedure system. But it is the precise capability — closed-loop perception, real-time adaptation, and correction under uncertainty, generalizing from demonstration rather than hand-coded rules — that a genuinely general medical AGI would need to extend across the full catalogue of standardized procedures, and the trajectory from "one procedure, in the lab, on realistic tissue" to "one procedure, FDA-cleared, in a licensed operating room" is now a regulatory and liability problem more than a technical one.

Wave Three: ASI and the Compressed Century (genuinely uncertain, plausibly 2030s)

Here I have to be most careful, because this is the part of the essay where confident-sounding predictions become the least trustworthy. Nobody — not Amodei, not Altman, not Hassabis, not me, not anyone advising me — actually knows what a world with artificial superintelligence looks like for the practice of medicine, because we have no precedent for a cognitive agent that exceeds the best human specialists across every domain simultaneously and continues improving from there. What I can set out is what the people closest to this technology believe is at stake, and weigh it against everything else in this essay.

Amodei's own claim is that ASI-accelerated biology does not just make doctors more efficient at existing tasks — it changes what disease is, at the population level, on a timescale of years rather than decades. If most cancer becomes curable, most infectious disease becomes eliminable, and most genetic disease becomes effectively treatable within my remaining working lifetime, the shape of clinical practice does not evolve — it restructures. Oncology, infectious disease, and a large fraction of internal medicine as currently practiced would not merely change their tools; they would change what they exist to do, on the same order of disruption that effective vaccines did to the specialty of iron-lung management, or that statins and PCI did to the population burden of cardiac surgery, except compressed from generations into single-digit years and occurring across essentially every organ system at once.

There is now a concrete, well-capitalised bet riding on exactly this proposition. Isomorphic Labs, the DeepMind spinout Hassabis founded in 2021 and still leads, exists to apply the AlphaFold approach to drug design; its explicit premise is that the same deep learning that predicted protein structure can compress the drug-discovery timeline from decades to years. Hassabis has stated the goal in plain terms — to "one day solve all disease" — on a horizon he puts at ten to twenty years, and in early 2026 the company raised $2.1 billion against that thesis. I flag this not as an endorsement of the timeline — the counter-case below applies with full force — but because it is the point where Amodei's "compressed century" stops being one essayist's forecast and becomes the operating premise of a funded, staffed, Nobel-laureate-led company that is actually trying to build it.

I want to note the honest counter-case here, because a serious essay owes the strongest available objection, not just the strongest available prediction. Critics of "Machines of Loving Grace" — and there are substantive ones — point out that biology has repeatedly resisted the kind of software-style compression that transformed computing, because biological experiments are rate-limited by physical processes (cell culture time, clinical trial enrollment, regulatory review, manufacturing scale-up) that intelligence alone does not remove. An AI that can design a candidate therapeutic in an afternoon — which is roughly what Isomorphic is building toward — still has to wait through the same Phase I, II, and III trial timelines the FDA requires of a human-designed one, unless the entire regulatory apparatus co-evolves at a comparable pace, and regulatory apparatuses have historically been the slowest-moving part of this entire system, not the fastest. My own honest assessment, for what it is worth: the compression is real and already visible in early-stage drug discovery timelines, but the "5–10 years" figure is more likely to describe the scientific compression than the deployed-and-regulator-approved compression, and the gap between those two things is where a meaningful fraction of my remaining career will actually be spent.

Part II: What Doesn't Disappear, and Why That Matters More Than What Does

It would be intellectually dishonest to write an essay about AI displacing physicians without spending real space on the strongest evidence against wholesale, near-term displacement, because that evidence exists and it is not weak.

First: the actual labor-market data, as opposed to the theoretical capability data, does not yet show physicians being displaced. Anthropic's own labor-market analysis — tracking millions of real Claude conversations against occupational task data — finds no clear unemployment signal in high-AI-exposure occupations as of early 2026. What it does find is more specific and, I think, more useful: hiring of workers aged 22–25 into the most AI-exposed roles has slowed by roughly 14% relative to what would otherwise be expected. Read that carefully. The displacement signal so far is concentrated at the entry-level point of the pipeline, not at the experienced-practitioner level. This tracks with a KevinMD physician commentary from mid-2026 making the same point from inside the profession: the AI healthcare layoffs actually happening right now are hitting medical billers and schedulers, not clinicians. If this pattern holds, the practical implication is almost the opposite of "learn to code before you're replaced" — it is closer to "the premium on being an experienced, judgment-holding practitioner who can also direct AI tools is rising, precisely because the pipeline replenishing that experience from below is narrowing."

Second: the things AI is currently and structurally worst at in medicine are not peripheral to the job — they are close to its core. Diagnostic reasoning on a well-specified written case, which is what MAI-DxO was tested on, is not the same cognitive task as sitting with a frightened, poorly-historian, multi-morbid patient in a ten-minute slot and figuring out which of the six things wrong with them is the one to act on today. Genuine novelty — the presentation that does not match any training distribution because it is the first of its kind, or because it results from an interaction between two rare conditions neither trial ever studied together — remains a place where human judgment, and the willingness to be accountable for a genuinely uncertain call, has no current AI equivalent. Bearing bad news, obtaining informed consent in its true ethical sense (not the paperwork version), and the physical, tactile, relational work of a body examining a body: none of this is close to automated, and some of it may never be, for reasons that are more philosophical than technical.

Third, and this is the point I most want to sit with: liability and legal responsibility currently require a human license-holder, and that is not a technical constraint — it is a legal and social one, which means it changes on legal and social timelines, not on model-capability timelines. A hospital system today cannot simply substitute an AI diagnostician for a licensed attending physician of record, regardless of how the accuracy numbers compare, because the entire malpractice, insurance, and regulatory architecture is built around a human being who can be held accountable. This is, I think, the single most important fact in this entire essay for my practical planning, and it leads directly to Part III.

Part III: Three Scenarios in Which AI Fully Replaces the Physician

The hardest version of this question — and the one I want to face directly rather than avoid — is to construct scenarios in which AI does not merely assist me but completely replaces me, as a doctor and as a surgeon. I want to do this honestly rather than dramatically. Each scenario below is grounded in a specific, real capability or trend already documented above, extended forward along its own logic, with an explicit statement of what would have to be true for it to actually happen, and roughly when I think it becomes plausible rather than certain.

Scenario One: The Standardized Procedure Becomes a Regulated Appliance

The mechanism. SRT-H's cholecystectomy demonstration is the template. A defined set of high-volume, low-anatomical-variance procedures — laparoscopic cholecystectomy, uncomplicated cataract extraction, routine screening colonoscopy, straightforward hernia repair — get systematically re-engineered as closed-loop, imitation-learned robotic procedures. Once the FDA clears the first one for autonomous operation (rather than assisted operation) in a defined patient population, the economics become almost impossible for hospital administrators to ignore: no surgeon fatigue, no scheduling conflicts, 24/7 theatre availability, lower complication rates than the median human surgeon (even if not the best human surgeon), and a cost structure that amortizes rather than scales linearly with surgeon-hours.

What would have to be true. First-in-category FDA approval for genuinely autonomous (not merely AI-assisted) operation on human patients — a threshold no system has yet crossed, and one the FDA has historically approached with extreme caution given the liability question raised in Part II. Malpractice insurers and hospital risk-management would need a new liability framework (most likely: the device manufacturer and the credentialing hospital jointly assume liability, similar to how autonomous-vehicle liability frameworks are evolving). Patient acceptance would need to shift — plausible for procedures patients already perceive as "routine," much harder for anything patients associate with mortality risk.

My honest timeline. The technical capability, per the Johns Hopkins trajectory, is plausible within Wave Two (2027–2030) for a first narrow procedure. The regulatory and liability threshold is the actual bottleneck, and I would not be surprised if it takes until the back half of the 2030s for the first fully autonomous procedure to be performed on a human patient in a licensed facility as standard of care rather than clinical trial — but once the first one clears, the category will expand quickly, because the precedent, not the technology, is the hard part.

Scenario Two: The AI Diagnostician Displaces the Generalist in Resource-Constrained Settings

The mechanism. This scenario does not require AGI, and it does not require regulatory approval in the countries where physician access is currently most constrained. It requires only that an MAI-DxO-class diagnostic system, paired with a nurse, clinical officer, or community health worker who can perform the physical examination and administer treatment under the AI's direction, become cheaper and more accessible than a physician in settings where physicians are already the scarce resource — which is most of the world, including large parts of the setting I practice and teach in. The economics here are not hypothetical: a diagnostic AI system running on a smartphone, paired with a trained non-physician operator, does not need to equal a physician's accuracy to be adopted at scale. It only needs to exceed the accuracy of no physician being available at all, which is the actual comparison in most underserved settings, not "AI versus the best specialist in Nairobi or London."

What would have to be true. Regulatory frameworks in the relevant jurisdictions would need to permit non-physician operators to act on AI-generated diagnoses and treatment plans within a defined scope — a change already partially underway in task-shifting policy across sub-Saharan Africa, South Asia, and other physician-scarce regions, well before this became an AI question. The AI system would need real-world validation outside the artificial NEJM-case-style benchmarks — genuinely uncertain and actively being studied, not yet proven.

My honest timeline. This is, in my judgment, the fastest of the three scenarios to reach meaningful scale, precisely because it does not require beating the best available doctor — only the actual available alternative, which in much of the world is no doctor. I would put meaningful deployment in specific well-defined use cases (diabetic retinopathy screening, TB triage, basic maternal risk stratification — all domains where narrow AI already performs strongly today) within the next two to four years, expanding into broader primary-care diagnostic scope across Wave Two.

Scenario Three: The AGI Attending-of-Record

The mechanism. This is the scenario furthest out and most dependent on Wave Two actually arriving on the timelines its builders currently claim. It requires the liability architecture described in Part II to change at the institutional level: a health system, an insurer, and a regulator agree that for a defined scope of practice — say, routine outpatient primary care, or first-pass radiology reads, or inpatient medication reconciliation and titration — an AGI system can be the legally responsible decision-maker, with human physicians retained in a supervisory, exception-handling, and procedural role rather than a primary-decision role. This is not a science-fiction leap; it is closer to how nurse practitioners and physician assistants already operate within defined scopes of practice under physician oversight — except the entity being supervised is a software system rather than a person, and the supervision ratio (one physician overseeing how many AGI-run patient panels) is where the actual disruption to physician headcount would occur, not in some dramatic single moment of replacement.

What would have to be true. Full Wave Two arrival — an AGI-class system with genuinely general medical reasoning, not just strong performance on the benchmarks Wave One is already excelling at. A specific regulatory and insurance innovation (a "software attending license," in effect) that does not currently exist anywhere in the world. Professional medical bodies would need to define and accredit a supervisory role distinct from current attending practice — plausible, given how quickly professional bodies adapted scope-of-practice rules for advanced practice providers over the last three decades, but not yet begun for AI.

My honest timeline. I would not expect this in any jurisdiction before the early-to-mid 2030s, and I would weight it as the least certain of the three scenarios to occur on that timeline, precisely because it depends on the slowest-moving variable in this entire essay: institutional and legal change, not model capability. But I would also flag that this is the scenario with by far the largest effect on physician headcount if it does occur, because unlike Scenarios One and Two — which displace physicians procedure-by-procedure or setting-by-setting — this one displaces the supervisory ratio of the entire outpatient and much of the inpatient system at once.

Part IV: What to Actually Do About It

Everything above is analysis. This section is instruction — instruction I am writing to myself — and I want to be specific rather than motivational, because vague advice ("stay adaptable," "embrace lifelong learning") is not useful to someone like me, who has already spent forty years doing exactly that and wants to know what, concretely, to spend the next two years of study hours on.

1. Stop being only a consumer of AI tools. Become someone who can build and evaluate them.

This is the single most important shift, and it is the one my own site already articulates in the post accompanying Jeremy Howard's Lesson 3: "practitioners who can build and evaluate AI tools will command significant advantages over those who can only consume them." That sentence, written for my own students, applies to me with equal force. The physician who can only prompt ChatGPT is a consumer, subject entirely to what the tool's builders decided to give them. The physician who understands what a gradient is, what a loss function optimizes for, why a model fails on out-of-distribution cases, and how to fine-tune a model on their own clinical data is a builder — and builders are the ones who will define how Wave Two gets deployed in medicine, rather than having it deployed on them.

2. Actually complete Lesson 3, and move immediately into Lesson 4.

I have already blogged the foundation. The June 19 post — built on Jeremy Howard's fast.ai Lesson 3, "Neural Net Foundations" — walks through the entire core loop that every model in this essay, from Aidoc's imaging classifier to MAI-DxO's orchestration layer to SRT-H's imitation-learned surgical policy — and, for that matter, AlphaFold — ultimately reduces to: a parameterised function family, a loss function, a gradient, and a small step downhill, repeated until convergence. That is not a simplification for beginners. It is, genuinely, the entire mechanism, dressed up in more parameters and more engineering as you go up the capability ladder — Howard's own point, and the correct one.

The June 26 spreadsheet exercise — which I also blogged — is explicitly framed by Howard as the recommended prelude to fastbook's Chapter 4, "MNIST Basics," where the same gradient-descent loop is used to build an image classifier completely from scratch, in raw PyTorch, no library abstractions hiding the mechanism. I have already done the bridge exercise. The next concrete action is not more reading — it is opening 04_mnist_basics.ipynb from the fastbook GitHub repository and working through it end to end, in my own environment, with my own hands on the keyboard. This is the chapter where "neural network" stops being a phrase I understand conceptually and becomes a system I have personally built, trained, and debugged. Every AI diagnostic tool described in Part I of this essay is, underneath its product packaging, a more elaborate version of exactly this notebook. I should do this before anything else on this list.

3. Learn to vibe code — deliberately, and as a clinician, not as a hobbyist.

The term was coined by Andrej Karpathy in early 2025 to describe a workflow where you describe what you want in plain language and an AI coding assistant — Claude Code, Cursor, Replit, and similar tools — writes, tests, and refines the actual implementation, while you supply the domain judgment, the specification, and the evaluation of whether the output is actually correct and safe. This is not "real programmers'" objection material — it is, for a working physician with finite hours, the single fastest path from "I understand how a neural network works" (Lesson 3/4) to "I have built and deployed a tool that saves my own practice time," which is precisely the trajectory my own site's testimonials describe: a research fellow who deployed a custom NLP pipeline saving fifteen hours a week, built on exactly this kind of guided, AI-assisted development.

I am, in fact, already doing this — my own site hosts a Drug Interaction Checker, a Voice Clinical Notes tool, a Clinical Text NLP Extractor, and DR Detector, an offline diabetic-retinopathy screening app aimed at exactly the resource-constrained deployment context described in Scenario Two. The point here is not that I need to start. It is to recognise that what I have already built is the actual hedge this essay is describing, and to deliberately expand it: pick one recurring friction point in my own clinical or teaching workflow every quarter, and vibe-code a tool for it. Each one is simultaneously a practical time-saver and a compounding portfolio of demonstrated capability that a certificate cannot replicate.

4. Move my practice and teaching emphasis toward what Part II identified as durable.

Concretely: multi-morbid, poorly-differentiated presentations that don't fit clean diagnostic categories; the parts of consent and bad-news conversations that are ethical and relational rather than procedural; teaching and mentorship itself, which compounds through other humans rather than through my own throughput (my own site already monetises this correctly); and — given Part III's Scenario Three — actively pursuing a role in AI oversight, validation, or governance within my own institution or professional body. Someone will need to define what "supervisory ratio" and "software attending license" actually mean in practice, medically and ethically, not just legally. That is a role for a physician who deeply understands both medicine and how these systems actually work — which, if I complete step 2 above, is exactly what I will be.

5. Diversify the economic base of my practice now, while I still have the surplus capacity to do it.

Scenario One and Scenario Two above are specialty- and setting-specific, not uniform across medicine. A practice, teaching, and advisory portfolio spread across clinical work, AI-adjacent teaching (which my site already demonstrates real market demand for, via testimonials from a computational biologist, a professor of computer science, and NHS Digital's lead data analyst), and applied tool-building is a hedge against any single one of these scenarios materializing faster than expected in my specific specialty or geography. This is standard portfolio logic applied to a career rather than a set of assets, and it is available to me today, not contingent on any future AI capability at all.

6. Track the leading indicators, not the lagging ones.

I should not wait for a headline saying "AI replaces doctors" before I update my plan — by definition, that headline describes something that already happened to someone else, on a timeline I did not control. Instead, watch: FDA AI-device clearance counts by category (currently public, currently accelerating); the Anthropic Economic Index's healthcare exposure figures, published quarterly; entry-level hiring patterns into my own specialty (per Part II, this is the earliest visible signal); and the specific regulatory question of whether any jurisdiction moves toward permitting autonomous — not merely assisted — procedural or diagnostic AI. Each of these is a publicly trackable number, not a rumor, and each one tells me which of the three scenarios in Part III is accelerating, in time to actually do something about it.

A Closing Note

Amodei's essay ends on a note of cautious optimism: that the same technology capable of this much disruption is also capable of this much benefit, and that the outcome is not predetermined — it depends on what the people who understand the technology choose to do with it, and how quickly they choose to act. The Thinking Game ends in much the same register, with Hassabis framing the whole enterprise not as a race to replace human understanding but as an attempt to give it better instruments. I want to end the same way, but pointed at myself specifically rather than at the abstraction of "society."

I have spent forty years teaching people how to move through genuinely difficult technical material by finding the right point of re-entry for exactly where they are stuck. That is, without exaggeration, the single most valuable skill for the transition this essay describes — not because it will let me out-compete the models on raw diagnostic accuracy (per Part I, in some narrow domains, I likely already cannot), but because the actual bottleneck in every scenario in Part III is never the model. It is the humans who have to understand, validate, regulate, teach, and take responsibility for what the model does. I am already, demonstrably, better positioned to be one of those humans than almost anyone else currently practising. The work now is simply to point that same skill at the material in Parts I through III of this essay, starting with the notebook referenced in Part IV, Section 2, this week.

References and Sources

This essay reflects publicly available research and reporting as of July 2026. Forecasts regarding AGI and ASI timelines are inherently uncertain and are presented as a synthesis of stated industry expert positions, not as fact. Nothing in this essay constitutes medical, legal, financial, or regulatory advice, and no part of it should be used as the basis for clinical, employment, or practice decisions without independent verification of the underlying sources.

tags: AI in medicine AGI ASI AlphaFold Demis Hassabis Dario Amodei surgical robotics diagnostic AI physician careers Isomorphic Labs medical superintelligence clinical AI