In Plain Sight - 05
In Plain Sight is a weekly blog from Rhys that we will publish every Thursday at 8am UK / 9am CET. Each edition will be focused on a thorny topic within investing, startups, learning and work and policy.
In Plain Sight is partially a personal attempt to think independently and to avoid leaning too much on AI for answers to our questions. Will we use AI to edit and polish the text? Yes. But more importantly, will we come up with all of the ideas and analysis? Yes.
Here follows Edition 05 - 'Do doctors still need 6 years of training to qualify?'.
------
The real question is "should we train doctors differently?" rather than "should we train doctors less?"
Given that, do prospective doctors need six - or ten, or fourteen - years to become a competent, safe doctor? Arguably not, if the training is redesigned around what actually needs to live in a human head versus what can live in a well-audited system.
The inspiration for this edition was actually how little of my own degree I both remember and use day-to-day. Naturally, my course was very different to medicine; I didn't expect to have to remember what determines the price of milk but I might have more consciously remembered how to treat a significant health issue. But it led me to wonder if, i) if we accept it's nigh-on impossible to remember everything you study and ii) there are improving tech systems that help you recall key info and apply your judgement, whether we ought to consider compressing time to qualify, particularly for longer courses like medicine and where we have persistent shortages of emerging talent.
If we compress the qualification timeline, the knock-on effects are significant: smaller student loans, an earnings trajectory that starts years earlier and - perhaps most interestingly for anyone thinking about workforce shortages - a much lower barrier to retraining as a doctor mid-career, at 35 or 45 rather than only at 18. Before we dive in, it's important to note that becoming a great doctor isn't all about knowledge. It's an incredibly demanding job. Aspiring doctors need time to learn how to manage the stress and pressure of being a doctor and develop the emotional intelligence required to support patients during some of the most stressful moments of their lives. This isn't developed overnight despite many aspiring doctors having this disposition - time spent learning the craft, applying knowledge, and trusting their intuition is a crucial part of becoming an effective practioner.
This said, the "six-year block" model looks particularly dated. A growing category of point-of-care decision support tools now sit directly inside a hospital's own electronic health record and surface clinical best practice at the exact moment a doctor is making a decision, rather than expecting that insight to have been memorised years earlier for an exam. The evidence backs the workflow: one oncology group found that embedding guideline-based order sets into the point of prescribing cut the number of unique treatment regimens being used by 44%, as clinicians converged on the recommended pathway instead of reconstructing one from memory each time. This is an example of the 'forgetting curve' being turned into a product - training doesn't have to happen once, up front and then get tested for retention forever after. It can happen continuously, embedded in the workflow, at the exact moment retention stops mattering because the system is right there with you in the moment. Uphill, a company in our portfolio out of Portugal, is one of the businesses built on exactly this thesis.
The same logic can be applied through other trades built on very long apprenticeships in the health sector.
Surgeons are one such category. Reviews of VR and haptic simulation training now show real transferability into the operating room: trainees who rehearse procedures in a headset arrive with measurably better technical scores and spatial awareness than those trained by conventional methods alone, and simulation is increasingly used to plug the gap left by shorter permitted working hours and tighter limits on how much live practice a trainee can safely get on a real patient. FundamentalXR, a London surgical-training platform we backed, is one of the companies scaling this kind of rehearsal without needing one mentor and one operating theatre per trainee.
The next step after training is assistance in the room itself and here the data is more precise. A systematic review of every FDA-cleared surgical robot classified each one on a five-level autonomy scale: the overwhelming majority (86%) still put the surgeon in continuous, direct control, while a small but growing share (6%) now sit at "conditional autonomy" - the robot proposes a patient-specific strategy, the surgeon selects or revises it and only then does the system execute and monitor. This reflects the structure of a 'driver-assist' system as opposed to self-driving: the machine proposes, the human approves and the surgeon carries the accountability throughout. We wouldn't want to be first in line for a fully autonomous robotic surgeon. But a surgeon-in-the-loop model, where the machine proposes and the human can always override? This feels much easier to trust, particularly when we are not yet convinced that the AI is consistently correct (as discussed in previous editions, scrutiny of AI's correctness is typically higher than our equivalent bar for humans) - indeed, it's already how most of the newest systems are built.
Architects face the identical question with less scrutiny attached, probably because the consequences of getting it wrong are slower to realise. Three qualifications and a minimum of two years' practical experience is the current bar in the UK. If AI tools can already generate, check, and stress-test a design against building codes, structural loads, and planning constraints, the years spent on tasks a system now does reliably start to look like the wrong thing to be gatekeeping.
Earlier this year, I was interested to briefly consider the market for training auto-technicians - a company was accurately claiming to be able to reduce the time to qualify for mechanics by 90%. It's interesting that this compressed qualification timeline is deemed acceptable whilst there is squeamishness around others, despite the risks associated with an incorrectly 'fixed' car...
None of this works if professional standards stay still but the underlying skills shift. The honest version of this would-be reform is most likely to "assess differently," rather than "study less."
This is already underway, even if it hasn't reached public consciousness yet. Medical education has spent the last decade quietly moving away from time-based training - spend X years in the building, collect a certificate - toward competency-based frameworks built around "entrustable professional activities": can this trainee be trusted to perform a specific clinical task independently, assessed through observed practice rather than a written exam. It's the same instinct as a driving test that now includes using a sat nav: nobody expects a driver to have the entire road network memorised, or to be able to simultaneously drive and read a paper map, but everyone still expects them to interpret what the sat nav tells them, notice when it's wrong and drive safely regardless. Medical and architectural assessment can extend the same logic: less "recite the differential from memory," more "given this AI-assisted data set, choose and justify the pathway." That's a harder test in some ways - it tests judgement under real conditions rather than recall under exam conditions.
So, what could go wrong?
A lot.
Compressing training assumes the tacit knowledge doctors and surgeons build over years - the thousand small pattern-matches that never make it into a textbook - can be taught faster or offloaded to a system without loss. That's not proven, and the potential cost of being wrong in medicine is harm to a real person. Even the most promising simulation and point-of-care tools are, by their own reviewers' admission, still short of the rigorous, large-scale validation this shift would eventually need.
Regulators have also historically used years-of-training as a blunt but legible proxy for competence and liability; if we replace it with something more practical, we need an assessment regime robust enough to be defensible when something goes wrong (because, sadly, it inevitably will at times). And there's a real risk that AI-assisted training produces clinicians who are excellent with the tool and brittle without it - the equivalent of a pilot who can't fly without autopilot.
None of that is a reason not to ask the question. It's a reason to make sure the redesign happens deliberately, with professional bodies, insurers, and educators in the room from the start, rather than being forced through by AI capability that outruns the credentialing system holding it back.
We think the interesting opportunity isn't "replace the doctor" (by any stretch!) - it's the infrastructure sitting underneath a shorter, more practical training pathway: point-of-care systems that surface the right guidance at the moment of decision instead of relying on years-old recall; immersive simulation that lets procedural skill be rehearsed and measured long before it's practised on a real patient; always-on diagnostic signal that gives a clinician a denser picture to reason from; and assessment infrastructure robust enough to credential judgement rather than memorised knowledge. This has enormous and broad application - doctors are an interesting and emotive case study.
As mentioned, we've backed pieces of this already - Uphill in point-of-care guidance, FundamentalXR in procedural simulation, Thymia in objective clinical signal - but the pattern is bigger than any one company and any one sector. If you're building somewhere in this space, we'd love to talk to you.
.png)
