AI · Oct 28, 2026
RTX Spark laptops get open speech models and a dial for background agentsMicrosoft’s health AI roundtable agrees on one thing: medicine’s bottleneck is adoption, not algorithms
· Updated · Signal Desk newsroom
Analysis of reporting by Microsoft Research. How we work

On a Microsoft Research podcast, Eric Horvitz, Eric Topol and Ava Amini map AI’s promise in diagnosis, imaging and biology. Their sharpest point is that proven tools still aren’t reaching patients, and that better data is the next constraint.
Key takeaways
- Eric Topol argues proven imaging AI, backed by large randomized trials in mammography and colonoscopy, is largely not being used in patient care.
- Horvitz describes Microsoft’s sequential-diagnosis work in Science, which puts AI in the clinician’s position of choosing questions and tests.
- Topol calls it the first demonstration of higher diagnostic accuracy at lower cost, while Horvitz notes the physician comparisons have limitations.
- Amini and Topol say biology lacks the large activity-first and long-term patient datasets that made AlphaFold possible.
- Horvitz predicts that within ten years some fatal cancers will become chronic diseases thanks to AI-driven breakthroughs; that is a forecast, not evidence.
Three optimists who keep saying ‘implementation’
Microsoft Research’s podcast The Shape of Things to Come put host Doug Burger with Microsoft chief scientific officer Eric Horvitz, Scripps Research’s Eric Topol and Microsoft principal research manager Ava Amini, who co-leads Project Ex Vivo with the Broad Institute. It is a company podcast, and two of the three guests work for Microsoft, so expect enthusiasm. What makes it worth a listen is how often the enthusiasm turns into frustration with the same word: implementation.
Topol is the bluntest. He argues the field has fixated on generative AI while ignoring where deep learning is better proven: medical imaging, from retinas to chest X-rays to ECGs. He points to a mammography randomized trial of over 100,000 women and 44 randomized colonoscopy trials, and says most American women still don’t get state-of-the-art AI reading of their mammograms unless they pay for it in a limited network. His explanations are institutional, not technical: a slow-to-change profession, thin health-system margins and a lack of demand from doctors or patients.
The diagnosis research behind the optimism
Horvitz describes Microsoft’s sequential-diagnosis work, featured on the cover of Science. Instead of handing a model a full case and grading its answer, the system starts with a chief complaint and some history, then decides which questions to ask and which tests to order while weighing their cost, and when to commit to a diagnosis under uncertainty. Topol, the one non-Microsoft guest, calls it the first demonstration of an agentic system diagnosing more accurately than physicians and other frontier models at lower cost. Horvitz is more measured, saying the comparisons with generalist physicians come with limitations.
That caveat matters. Cost-aware diagnosis is the right framing, because medicine’s real problem is rarely whether a model can name a disease from a complete file. But a research benchmark is not a clinic, and Microsoft’s broader biology push, like its Quine research system, will ultimately be judged by outcomes in the real world.
Data is the next wall
Amini and Topol converge on data. Topol says whole-patient models combining health records, raw images, genomics and other layers could predict when disease will strike and test what-if interventions, but deep, long-term patient data barely exists at scale. Amini uses AlphaFold as the analogy: it worked because the Protein Data Bank existed, yet those structures are static. Biology needs activity-first datasets that map sequence to function and cell responses to perturbations, chosen intelligently because the full space is too vast to measure. Topol thinks mature whole-patient models will arrive before a true virtual cell.
That is a useful corrective to the idea that bigger models alone will crack biology. The panel’s implicit argument is that the next big health AI investments should be in data collection and clinical deployment, not just model training.
From the desk
The forecasts at the end are bold. Horvitz predicts that within a decade the public will credit AI with turning some fatal cancers into chronic diseases and slowing Alzheimer’s; Topol talks about preventing Alzheimer’s and cancers. We would file those as aspirations from people well placed to pursue them. The more actionable message is Topol’s complaint: AI tools validated in randomized trials are sitting unused, and patient demand may be what finally moves health systems. Ambient scribes, which he notes are taking off in the US, show adoption is possible when the benefit is immediate for clinicians.
Who it speaks to
- Patients
- Proven AI reads of mammograms and other images exist, but access is limited and sometimes requires out-of-pocket payment.
- Health systems
- Low margins and risk aversion are slowing adoption of tools with randomized-trial evidence.
- Biology researchers
- The panel sees activity-first and long-term patient datasets as the binding constraint.
- Clinicians
- Ambient scribes and diagnostic assistants could return time to patient care, if deployed well.
What to watch
- Prospective clinical trials of sequential-diagnosis agents outside research settings.
- Insurer and health-system decisions on paying for AI image interpretation.
- New large-scale biological datasets aimed at activity and perturbation rather than static structure.
- Results from Project Ex Vivo on defining and targeting cancer cell states.
FAQ
Who was on the podcast?
Host Doug Burger with Eric Horvitz (Microsoft), Eric Topol (Scripps Research) and Ava Amini (Microsoft Research).
What is sequential diagnosis?
An approach where an AI system starts with limited patient information and chooses questions and tests, weighing cost, before committing to a diagnosis.
Why isn’t imaging AI widely used?
Topol cites slow adoption by the profession, low health-system margins and a lack of demand from doctors and patients.
What’s holding back AI in biology?
The panel says datasets: biology lacks large activity-based data and long, deep patient records.
Companies: Microsoft
More on Microsoft
AI · Nov 12, 2026
Microsoft Research finds 8% of coding-agent sessions drive 44% of AI costs, plus a KLS proofAI · Oct 29, 2026
Microsoft research finds coding agents struggle most with reading unfamiliar codeAI · Oct 30, 2026
A single robots.txt line to block AI training: what the OpenAI-Google-Microsoft-Anthropic pledge does and doesn’t fixAI · Nov 19, 2026
At Ignite, Nvidia and Microsoft pitch one stack from Vera Rubin racks to RTX Spark laptops