All news

The coronavirus response shows machine-readable data is the scarce input

Scattered blank sheets on a pale table beside one neat stack filed in a graphite-blue archive box with a cobalt divider

The White House asked AI researchers to mine 29,000 coronavirus papers and pooled 330 petaflops of supercomputers. Clean domain data now gates applied AI.

Five institutions were needed to turn more than 29,000 scholarly articles, over 13,000 of them in full text, into a corpus machines can read. That corpus, the COVID-19 Open Research Dataset or CORD-19, built by partners including the Allen Institute for AI, Microsoft and the National Library of Medicine, is what the White House science office asked the country's AI experts to mine on 16 March, with research questions set with the World Health Organization and tools and findings going to an open challenge on Kaggle.

The order of work carries the lesson for capital. Within a week of a public health emergency, the inputs applied AI needed were clear. Models and talent were available, and what had to be built was a clean, machine-readable corpus in a specialist field. In applied AI the defensible asset is domain data prepared for machines, together with the experts who can judge the output, more than any single model trained on it.

Compute arrived second and was pooled quickly. On 22 March the White House, the Department of Energy and IBM launched a COVID-19 High Performance Computing Consortium: 16 systems with more than 330 petaflops of capacity from national laboratories, universities and large cloud providers, matched to research proposals through one portal. Processing power came together in a single consortium, while the corpus needed five institutions. The pandemic is testing many business plans at once, and this part of the AI stack has a clear user and a clear job.

In health, science and other expert fields the opening sits between raw information and models: ingestion, curation, extraction of entities and relations, and evaluation by domain experts. Public datasets like CORD-19 lower the cost of a first product, but they are open to everyone, so a company needs its own data loop with users to stay ahead. Public programs ease access to compute for research-stage teams. The harder task is proving that results change what scientists or clinicians decide, and Fanaura asks exactly that proof of applied AI founders in science and health.

Access is not impact, and so far the crisis has proved only access. Tools built on CORD-19 still have to produce findings that researchers use, beyond challenge rankings. Publishers have to keep coronavirus literature open and machine-readable after the emergency, which would set a precedent for other fields. The consortium's own allocation will show the balance between old and new methods: how much of its capacity goes to machine learning, such as models for protein structure and drug screening, alongside classic simulation.

Applied AI is passing this stress test on data and compute, though not yet on results. Value will settle with teams that own curated domain data and the trust of expert users, which is why science-heavy AI needs patient capital. In a crisis the scarce input turned out to be a clean dataset, not a faster chip.

For founders

Building tools that turn scientific or clinical information into data machines can use? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.