I didn’t believe in AGI—until the buildout forced me to look twice.
I used to dismiss artificial general intelligence intelligence as sci‑fi. Then I read Situational Awareness and saw the unglamorous reality: power contracts being locked up, GPU clusters scaling from billions to hundreds of billions, and serious talk of trillion‑dollar builds. If we keep stacking orders of magnitude in effective compute—bigger clusters, better algorithms, and “unhobbling” models with tools and long context—AGI becomes a live, near‑term possibility. If one country gets there first, the advantage (economic, scientific, national security) could be decisive.
That argument turned me from skeptic to “okay, maybe.” But as someone who’s watched MedTech progress hinge on cross‑pollination and causal insight, I’m not betting everything on brute force.
Why this matters for MedTech
Our field rarely swims in massive, clean datasets. We ship Class II/III devices, operate under tight regulatory scrutiny, and often pursue rare conditions where N is small. What moves the needle are leaps that connect dots across domains—the way a waffle iron inspired Nike’s outsole, or how burrs inspired Velcro.
Today’s LLMs are incredible at language, retrieval, and workflow acceleration.
But they still struggle with the kind of causal, embodied, cross‑domain insight that drives MedTech breakthroughs.
Song‑Chun Zhu’s “small data, big task” philosophy (reasoning to a goal with minimal inputs) hits closer to the realities of our work than “big data, small task.” If AGI emerges along Zhu’s path, it could transform how we discover mechanisms, design devices, and prove safety/effectiveness—faster and with fewer patients.
Where scale-first AI helps MedTech now
Let’s give credit where it’s due. The scale-driven models already unlock value:
Literature triage and horizon scanning across PubMed, patents, and clinicaltrials.gov
Drafting CERs, PMS plans, and regulatory correspondence (with human QA)
Coding assistance for embedded software, data pipelines, and test harnesses
Signal processing scaffolds for imaging and wearables (with expert supervision)
Post‑market surveillance: faster signal detection across complaints and RWE
These are real gains. Use them. But don’t confuse smoother paperwork and faster search with the kind of insight that invents your next platform.
Where “small data, big task” could change the game for us
This is where I think MedTech should lean in—because it maps to how we actually build.
Rare diseases and small cohorts
Causal inference and mechanistic priors to extract signal from N=80, not N=80,000.
Synthetic control arms and adaptive designs tuned to physiology, not just statistics.
Device mechanism and materials discovery
Hypothesis engines that reason about surfaces, flows, fatigue, and biocompatibility with minimal data plus physics constraints.
Cross‑domain analogies that travel: think sharkskin for anti‑biofouling, lotus‑leaf micro‑textures for anti‑wetting catheters, gecko‑inspired dry adhesives for atraumatic fixation, origami/kirigami for deployable stents, and Archimedes‑screw logic for pumps
Surgical and interventional robotics
Embodied, causal reasoning in simulation before a robot ever touches a patient.
Planning under uncertainty with tool–tissue interaction models and safety constraints.
Diagnostics and multimodal fusion
Combining little bits of signal (imaging, waveforms, labs, notes) into causal hypotheses about what’s changing in the patient—not just pattern match.
Personalization at the edge
On‑device agents that reason about context to auto‑tune settings (CPAP, neuromodulation, insulin delivery) without round‑tripping PHI to the cloud.
Explainability that passes scrutiny
Verifiable reasoning and audit trails that align with FDA expectations for GMLP and change control—because “why” matters in regulated care.
The kind of insight MedTech lives on
We’ve always borrowed brilliance across domains:
Velcro from burrs; waffle iron to track outsoles; sharkskin micro‑denticles to reduce bacterial adhesion; lotus leaves for superhydrophobic surfaces; gecko toe pads for reversible adhesion; hummingbird shoulder mechanics for articulating tools; origami stents and deployables; ultrasound beamforming refined from radar/phased arrays.
An AGI that reasons causally could reproduce these leaps on demand—scanning biology, materials science, and engineering to propose plausible mechanisms, not just citations.
That’s the promise Zhu’s camp is chasing: agents that can reason toward a goal with minimal inputs, in messy environments, with physical intuition intact.
Practical implications by function
R&D
Use causal/physics‑informed models to winnow hypotheses before wet lab or bench tests.
Stand up embodied simulation to pressure‑test control strategies and tool–tissue interactions.
Clinical and evidence
Design trials that work with small n: adaptive randomization, digital endpoints, and synthetic controls—while preserving causal validity.
Build RWE pipelines that detect mechanism‑consistent effects, not just correlations.
Quality/Regulatory
Prefer architectures that produce traceable rationales and stable behavior across updates.
Instrument model monitoring for dataset shift and performance drift; pre‑declare change protocols.
Product
Architect for edge inference where it improves latency, privacy, and robustness.
Make “explain‑my-recommendation” a first‑class UX element for clinicians.
Security/IP
Treat AI models and datasets like crown jewels (air‑gapped training, access control, full audit trails).
Lock down PHI by design; don’t let third‑party tooling become your soft underbelly.
My stance: run a dual‑track strategy
I’m hedging—intentionally.
Track A: Keep exploiting scale-first models for workflow acceleration, discovery support, and code. Measure impact in weeks.
Track B: Invest in “small data, big task” programs that target the hard parts of medicine: causality, embodiment, and generalization under shift. Measure progress with transfer-heavy, mechanism‑aware benchmarks—not just leaderboard deltas.
If the scaling thesis keeps delivering, we’ll know because agents start doing expert‑level work, legibly, across modalities. If it stalls, we’ll be grateful we built competence in causal, embodied reasoning—before our competitors did.
Where I land (and why it matters now)
Situational Awareness convinced me the scale path could get us to AGI faster than I expected—and that geopolitics will matter. Zhu convinced me we might be optimizing the wrong variable for medicine if we don’t prioritize causal, embodied, small‑data reasoning.
MedTech wins when we connect dots others don’t. I’m not willing to put all our eggs in a single basket—especially if that basket can’t explain “why.”
If you’re building in MedTech: where would “small data, big task” unlock your next breakthrough—materials, mechanisms, robotics, or clinical design? And what will you do this quarter to test it?
Be part of MedTech Malta 2025, happening from 12 to 14 November, where global MedTech leaders, innovators, investors, and clinicians unite to discuss the latest breakthroughs and trends shaping healthcare. Experience three days of insightful panels, interactive sessions, a cultural tour, and unparalleled networking opportunities designed to inspire collaboration and meaningful connections. Secure your spot today and take part in the conversations advancing medical technology worldwide.
