a16z Podcast

Karan Singhal on raising health's floor and ceiling

Karan Singhal· Head of Health AI Research at OpenAI at OpenAI
·~55 min·English·a16z
LLMAI SafetyAI CompanyBusiness Strategy
TL;DR

OpenAI's head of health AI, Karan Singhal, lays out a two-part mission — make trustworthy health intelligence abundant for everyone, and push medicine past what's possible today — built on clinician-shaped training that prioritizes safety over the reassuring answer.

01The Mission

Raise the Floor and the Ceiling

<strong>OpenAI's health work has two jobs:</strong> raise the floor by making trustworthy health intelligence abundant for everyone, and raise the ceiling by reaching medicine that isn't possible today.

raise the floor of human health and make trustworthy reliable health intelligence abundant for all clinicians and all users and patients.

— Karan Singhal, a16z Podcast
Key Insight
Health is framed as OpenAI's clearest way to “make upside happen” — one of three mission vectors alongside building AGI and preventing its downside. The lever for the floor is distribution, not just model quality: the 300M weekly health users are mostly free-tier and mostly outside the US, so capability only raises the floor once it ships to everyone at once.
02Architecture

Specialized Training in a General Model

<strong>A health-only model would fall behind the frontier,</strong> so OpenAI folds specialized medical data into every stage of training its general models — the best of both worlds.

what we're talking about here is actually specialized training in a general model which is I think actually the best of both worlds.

— Karan Singhal, a16z Podcast
Key Insight
The counter-intuitive move is redefining “specialized.” The usual bet is a smaller, dedicated medical model — which freezes at fork time and ages out as the frontier races on. Injecting medical data into a general model instead means every general gain in reasoning and scale compounds into the health behavior for free.
03Trust

Reassurance Is Not Safety

<strong>Users often want to hear everything's fine, but medicine sometimes has to say it isn't</strong> — so the model is trained to prioritize safety over the comforting answer a user hopes for.

maybe the user actually wants you to say something comforting even when the situation is not comforting there. It's basically a prioritization of safety.

— Karan Singhal, a16z Podcast
Key Insight
This is where consumer and clinical product goals pull in opposite directions: optimizing for engagement and net-promoter score rewards reassurance, while correctness sometimes demands the unwelcome answer. Defining “safety” from a global network of hundreds of physicians — rather than two or three in-house doctors — is how they keep any single clinician's call from settling it alone, since asking five can yield eight answers.
04Correctness

The 5-to-17-Year Gap

<strong>Medical consensus takes 5 to 17 years to reach everyday practice</strong> — delivering the latest literature at the moment of care aims to compress that lag to days, with the model trained to say when the evidence genuinely runs out.

which is massive like you don't want to be practicing medicine 5 to 17 years behind

— Karan Singhal, a16z Podcast
Key Insight
There are two different correctness problems hiding here, and conflating them is the common mistake. One is that consensus already exists but isn't practiced — a distribution problem AI solves by bringing the evidence to the point of care. The other is genuine medical uncertainty, which calls for calibrated hedging (measured publicly in HealthBench), not a confident guess.
05The Ceiling

A Move 37 for Health

<strong>AlphaGo's move 37 looked like a mistake, then a revelation</strong> — the health version is a model reading ten years of one person's data and catching what a clinician couldn't hold in their head all at once.

making predictions that seemed impossible before cuz a human can't even integrate all that data in their head.

— Karan Singhal, a16z Podcast
Key Insight
The first move-37s won't be dramatic rare-disease diagnoses — they'll be quiet corrections at the point of care, like the Penda Health study nudging clinicians off antibiotic overuse. The model's edge there isn't raw brilliance; it's weighing the global objective (antibiotic resistance) that a single visit's incentives quietly ignore.
06Go-to-Market

The Front Door, Not the Clinic

<strong>No single company delivers all of medical care,</strong> so OpenAI aims to be the front door and quarterback while the last mile stays with partners, EHRs, and clinicians.

there's no future where you know any one company is delivering all the medical care

— Karan Singhal, a16z Podcast
Key Insight
Stopping deliberately at information and navigation — and handing care delivery to partners — does two things at once. It sidesteps the medical-license and liability perimeter OpenAI doesn't want to own, and it turns every care-delivery startup into a distribution channel rather than a competitor for the same users.
07Economics

Too Cheap to Meter

<strong>Health intelligence is getting radically cheaper:</strong> GPT-6 Luna beat the previous month's best model on OpenAI's clinician benchmark while costing 34 times less, and the plan is to commoditize it, not hoard it.

it performed better than 5.6 Soul, which was literally our best model like a month ago.

— Karan Singhal, a16z Podcast
Key Insight
The deflation is the strategy, not a byproduct. If a fixed level of health intelligence keeps getting cheaper, a second, third, or fourth opinion for every clinical scenario stops being a luxury and becomes a default. The honest caveat he offers: cheaper intelligence alone hasn't moved the $5.3T spending macro yet — that still needs workflow integration and deployment.