1. A Revolutionary Blood Test Could Lead To Predictive And Prescriptive Analytics
Currently used in 500 million tests annually in the US, the Complete Blood Count (CBC) provides descriptive data relative to population averages, but it misses the individual insights necessary for precision medicine. In other words, the CBC is a low-value diagnostic tool that overlooks the transformative potential of longitudinal, personalized phenotyping. Accordingly, Medicare and Medicaid reimburse CBCs at only ~$10 per test.
Recently published, a groundbreaking study in Nature reveals1 that CBC metrics are much more predictive in the context of haematological setpoints—patient-specific baselines that typically are stable for decades. Those setpoints include hemoglobin levels, red cell distribution width (RDW), and white blood cell count (WBC) that are like biological fingerprints and predict outcomes more effectively than population-based thresholds. Deviations in hemoglobin setpoints are excellent at predicting mortality from all causes, while RDW is correlated to atrial fibrillation and WBC to Type 2 diabetes. This shift from averages to individual profiles heralds a paradigm shift in diagnostics.
To unlock the CBC's full potential, our research suggests that diagnosticians should shift from static, one-off tests to longitudinal phenotyping that can flag patterns associated with early health risks. Key to the analysis are single-cell genomics technologies—such as 10x Genomics' scRNA-seq and Standard BioTools' CyTOF solutions—that analyze blood samples at the cellular and molecular levels. Those methods could transform CBCs by revealing the detailed molecular changes behind setpoint deviations.
Cost is the main hurdle. Single-cell genomics’ costs would have to drop ten-fold to match the current CBC price point of ~$10 per test. The learning curve associated with unit growth will continue to drive the price point. During the past 15 years, as single cell genomic tests have scaled one million-fold, the cost has dropped by ~40,000x.2 Given the 500 million in CBC tests per year in the US, the market potential for advanced single-cell analysis is vast.
The combination of routine CBCs with single-cell genomics and AI-driven analytics could shift health care toward predictive and prescriptive medicine. A healthcare system in which deviations from personal baselines trigger timely interventions, targeted treatments, and better outcomes could reduce drug and other healthcare costs.
The CBC is ripe for reinvention, an opportunity that could transform not only diagnostics but also healthcare economics, bringing high-value, personalized medicine to millions of people. The transformation should begin now.
2. Can Off-Grid Solar Power AI?
Last week, researchers from Paces, Scale Microgrids, and Stripe published a new paper exploring the economics and feasibility of off-grid, solar-to-power AI datacenters. Why off-grid? Currently, interconnecting to the grid takes five years. The authors suggest that a solution behind-the-meter could take only two years, as shown below.
Source: Baranko et al. 2024.3 For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.
Importantly, the paper explores the real estate challenges associated with large-scale solar deployments. While private land suitable for large-scale solar development exists, ARK’s research suggests that orchestrating deals at scale will be a major hurdle. We encourage anyone interested in the energy landscape to read the full paper.
3. For The Finale Of OpenAI’s "12 Days of OpenAI," A Model Advanced Reasoning Capabilities
Last week, OpenAI concluded its "12 Days of OpenAI" by releasing two new frontier reasoning models, o3 and o3-mini.4 The most capable offerings from OpenAI to date, the models now dominate on several deeply challenging benchmarks designed to test the ability both to reason and to execute real-world tasks.
The largest sequential leap in performance to date on the agentic coding benchmark SWE-bench verified, o3 scored 71.7%, beating the previous record holder, Amazon's Q Developer Agent, by a staggering 16.7 points, as shown below.
.png)
Source: ARK Investment Management LLC 2024, based on data from OpenAI as of 2024 and SWE Bench as of 2024.5 For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.
The o3 model also outperformed o1's scores not only in research mathematics, but also on PhD-level science questions and the ARC-AGI benchmark, a unique puzzle-based benchmark. On ARC-AGI, o3 scored 87.5%, strikingly above o1's best score of 32%.
Clearly, advances in reasoning are enabling o3 to chain together actions and perform more complex tasks reliably. OpenAI demonstrated, for example, that o3 was able to create its own code generator and executor, ultimately using the tool that it built to create and run a script to evaluate its own performance on a PhD-level science benchmark. The unprecedented ability to execute a series of complex tasks autonomously represents a meaningful step toward a future in which AI agents perform real work for users that extends well beyond the limited capabilities of a chatbot. While not yet available broadly to the public, OpenAI has released both o3 and o3-mini to public safety researchers for further testing.