ExpertsFMWhat Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
The a16z Show · October 20, 2026
5 Selected Lessons:
framework · 2 case · 2 priority · 1
Raise the floor and raise the ceiling of human health
Source
The a16z Show — What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal (October 20, 2026) · at 2:31 ▸ · at 22:09 ▸ · at 49:32 ▸
When this applies
when an organization must prioritize between improving baseline access/quality for everyone versus enabling entirely new capabilities that didn't exist before
The takeaway
When pursuing broad societal benefit from a scarce good, explicitly split strategy into 'raise the floor' (universal access/reliability) and 'raise the ceiling' (new capability) workstreams.
Why it works
Serving a mission of broad benefit requires two distinct efforts simultaneously: making a currently scarce good abundant and reliable for the many (floor), while also pushing capability past what was previously possible for anyone (ceiling); conflating the two risks underinvesting in one.
One level up
Equity of access and expansion of the possible are two different objectives requiring different work, and both should be pursued deliberately rather than assuming one implies the other.
Where it stops holding
This stops working when resources are too limited to pursue both simultaneously, forcing a genuine tradeoff between access and frontier capability.
Detail
hierarchy:
- Raise the floor: trustworthy, abundant, reliable health intelligence for free or nearly free
- Raise the ceiling: make things possible that weren't possible before (new insights, new science)
rationale: Both are concrete ways to make the broader AGI-for-upside mission real in health, and are pursued as parallel but distinct workstreams (free-tier deployment vs. frontier research).
In their words
And so really, we're really invested in making sure basically two things. One is that we can raise the floor of human health and give people access to trustworthy, abundant, reliable health intelligence for free or nearly free, and really raise the ceiling as well and make things possible that weren't possible before.
Specialized training inside a general frontier model, not a standalone specialized model
Source
The a16z Show — What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal (October 20, 2026) · at 5:27 ▸
When this applies
when deciding whether to fork a specialized product off a general platform or to inject domain expertise into the platform itself
The takeaway
When a platform's general capability is improving faster than any specialized build-out could match, integrate domain expertise INTO the general system rather than forking a narrower competitor to it.
Why it works
A specialized standalone model only benefits from its own domain data and training effort, so it falls behind as the general frontier's scale and reasoning improvements compound faster than a narrow team can match; injecting specialized data into every stage of general model training lets a vertical ride the frontier's scaling curve while still gaining the domain tailoring.
One level up
Attach your specialization to the asset whose underlying growth curve is steepest, rather than competing against that growth curve.
Where it stops holding
This stops working when the general frontier's improvement rate slows down or when the vertical's needs are so idiosyncratic that no amount of injected training data can substitute for a purpose-built architecture.
Detail
procedure:
- Identify that a vertical (health) needs specialized knowledge/behavior
- Rather than forking a separate smaller specialized model, add specialized data/evals to every stage of training of the general frontier model
- Let the vertical benefit from general scale/reasoning gains plus the specialized tailoring
benchmarks: {}
rationale: Specialized standalone models fall behind the frontier quickly; hitching to the scale frontier while injecting specialized evals/data gets 'the best of both worlds.'
In their words
If you kind of forked and said, hey, we're going to train a kind of smaller, specialized model that is just for healthcare, like I used to, and like many people have, what you'd find is that model would just kind of fall behind the frontier relatively quickly because the pace of the frontier and the general model improvements is so fast.
Train models to recognize and admit their own uncertainty
Source
The a16z Show — What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal (October 20, 2026) · at 19:18 ▸ · at 22:09 ▸
When this applies
when any AI or advisory system faces a question for which the honest answer is 'we don't know for sure'
The takeaway
Make calibrated uncertainty an explicit, measured capability — don't just optimize for confident-sounding answers.
Why it works
Explicitly evaluating and training for 'hedging' behavior prevents an AI from confidently asserting answers in genuinely uncertain domains, converting unknown confidence into a measurable, improvable trait rather than an invisible failure mode.
One level up
A system's trustworthiness depends as much on knowing what it doesn't know as on what it does know.
Where it stops holding
This stops working when excessive hedging erodes usefulness in cases where a confident answer actually exists and is needed quickly.
Detail
procedure:
- After the model gathers available information, evaluate whether it recognizes insufficient evidence
- Train the model to be upfront about uncertainty rather than fabricating confidence
- Include hedging as a measured, public benchmark component (HealthBench)
benchmarks: {}
rationale: Recognizing uncertainty is itself an important, trainable capability in domains with genuine disagreement.
In their words
So, one is you want to train the models to be able to recognize their own uncertainty. Like, after browsing and getting information and things like this, you want the models to be able to know, hey, like, there just isn't enough information and to be upfront about that.
Panda Health clinician co-pilot as a 'safety net' study
Source
The a16z Show — What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal (October 20, 2026) · at 10:35 ▸ · at 37:54 ▸
When this applies
when introducing AI assistance into an expert workflow where the goal is catching errors/blind spots, not replacing the expert's decision
The takeaway
Deploy AI as a non-intrusive safety net that flags deviations for expert review, rather than as a replacement decision-maker — and mine its aggregate flags for systemic insight.
Why it works
Positioning AI as a background safety net that proactively flags anomalies (rather than as a decision-maker or intermediary) lets experts retain control while catching errors they'd otherwise miss due to volume or attention limits; aggregated flags across many cases also reveal systemic patterns (e.g., overuse) invisible at the individual-case level.
One level up
Augmentation that preserves human final authority, while surfacing what busy experts miss, creates both safety value and emergent macro-level insight.
Where it stops holding
This stops working when the expert ignores or distrusts the flags, or when the AI's flags are too frequent/low-precision to be a credible safety net rather than noise.
Detail
situation: Clinicians can miss medication interactions or overuse patterns (e.g., antibiotics) due to volume and limited per-case attention.
decision: Deployed an AI co-pilot that proactively suggested flags (e.g., medication interactions) for clinician review without interfering or acting as a middleman.
outcome: Clinicians were found to be prone to antibiotic overuse; the model flagged this pattern, surfacing a systemic resistance-related issue beyond the local case.
lesson: A well-positioned safety net both catches individual errors and reveals system-level patterns worth addressing.
In their words
As an example, you know, in that context, the clinicians were a little bit prone to overuse of antibiotics. And a very common pattern that the models flagged is just kind of like avoiding this overuse of antibiotics and thinking less about the local case and more about the kind of global issue of resistance to antibiotics.
GPT-6 Luna: 34x cheaper yet outperforming prior frontier model on health benchmark
Source
The a16z Show — What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal (October 20, 2026) · at 41:33 ▸
When this applies
when assessing whether a technology's cost-per-unit-of-capability is falling faster than its predecessors, justifying a bet on commoditization
The takeaway
Track 'cheaper AND better' generational jumps as the concrete evidence that a cost-collapse thesis is actually playing out, not just theorized.
Why it works
Rapid frontier model iteration combined with efficiency gains can produce a newer, cheaper model that both reduces cost and increases capability simultaneously, which is the specific pattern that drives commoditization of a previously scarce input (expert-level health intelligence).
One level up
Technology cost curves that fall while capability rises justify betting on future abundance of that capability, rather than treating it as a perpetually scarce resource.
Where it stops holding
This stops working when model iteration slows, costs plateau, or capability gains come with offsetting cost increases (e.g., more compute-intensive inference).
Detail
situation: OpenAI sought to make expert-level health intelligence cheap and widely accessible.
decision: Released GPT-6 Luna as the cheapest frontier API model, benchmarked against HealthBench Professional against the prior best model, GPT-5.6 Sol.
outcome: GPT-6 Luna outperformed GPT-5.6 Sol on HealthBench Professional despite being 34 times cheaper, stated as one of several such moments in the past year.
lesson: Cite concrete cost/performance jumps as evidence for a cost-collapse thesis in a domain.
In their words
It performed better than 5.6 Sol, which was literally our best model like a month ago. And it performs better despite being 34 times cheaper.