Large Universe Model/Applications/Large Universe Models for public health
ApplicationLarge Universe Models for public health
Last Tuesday's estimate is still moving. Only a model that never stops updating can represent that honestly.
What it ingests
Case reports, wastewater sampling, prescription and pharmacy volumes, school and workplace absenteeism, genomic surveillance, emergency-department presentations, and the reporting-delay distributions of each of those sources.
Nowcasting is belief revision
Public health data arrives late, incomplete and biased. A case counted today may have occurred nine days ago. Reporting delays vary by weekday, by jurisdiction, and by how overloaded the reporting system currently is.
The consequence is that the estimate of the past keeps changing. What you believed about last Tuesday is revised on Wednesday, Thursday and the following Monday, as the backfill arrives. A static answer to "how many cases were there last week" is not merely imprecise; it is a category error.
A Large Universe Model represents this natively. Beliefs about past intervals stay open and revisable, and the revision is reported rather than silently applied.
Multi-stream reconciliation
No single surveillance stream is trustworthy alone. Case counts depend on testing behaviour. Wastewater is unbiased by care-seeking but noisy and lagged differently. Absenteeism is broad and unspecific.
Weighting them into one posterior — each with its own precision, lag and known biases — is exactly what a Large Universe Model does with evidence. When they diverge, the divergence is itself the signal: wastewater rising while cases stay flat is a plausible early warning that testing behaviour has changed.
Why a Large Language Model cannot do this
The data postdates any cutoff, the estimate must be maintained rather than generated, and the provenance of each contribution matters for public communication. A Large World Model contributes nothing here — there is no scene to simulate.
Limits
Surveillance models inform decisions with serious consequences, and a well-calibrated interval is not the same as a correct one. The value of the Large Universe Model framing is honesty about uncertainty and its movement, not a claim to greater accuracy than the underlying data supports.