Tempus AI Lifts Lung Cancer Testing by 24% at ASCO — May 29, 2026

0:00ASCO 2026 opens today, and AI just lifted lung cancer biomarker testing by 24%. Welcome to MedAI Times Podcast, your daily update on medical AI. Don't forget to like and subscribe.

0:12Tempest-Next lifts lung cancer biomarker testing by 24%. Mayo Clinic brings 30 oncology AI studies to ASCO. Sylvester AI reads bone marrow to personalize myeloma therapy. Penn Medicine flags AI gaps in patient cancer info.

0:28NEJM AI publishes MES open data standard for health AI. Anthropic ships Claude Opus 4.8 with dynamic workflows. Anthropic in talks to raise $65 billion at $965 billion.

0:41Mythos moves toward broader release after 10,000 bugs found. That is a massive list for today. It really is. So, imagine building a billion-dollar artificial intelligence model that can perfectly predict cancer on a pristine, you know, controlled test track.

0:56But the second you drop it into a real messy hospital, it just suddenly forgets how to drive. Yeah, it just completely stalls out. Right. And today we are looking at exactly why medical AI is simultaneously crashing into the friction of the real world while also, you know, achieving these massive life-saving breakthroughs.

1:11So, welcome to the deep dive. Today is May 29th, 2026. Which is a huge day, by the way. Huge. Because today is the opening day of the American Society of Clinical Oncology Conference, ASCO, in Chicago.

1:23So, our mission for you today is to explore how AI is finally moving past that sort of bench science hype and into actual clinical deployment. Right, getting to the patients. Plus, we're breaking down a sweeping new data standard, trying to save these models from failing, and an industry-shifting update in the general AI space from Anthropic that literally just dropped right on the eve of this conference.

1:47Yeah, the timing on that Anthropic drop was not an accident, I'm sure. Oh, definitely not. But let's jump right into the ground reality at ASCO, because we've spent years talking about retrospective AI studies, right? Where models are trained on clean, old data.

2:01But today, we're actually seeing deployment wins. We are, and it's a fundamental shift in the whole narrative. I mean, the presentation turning the most heads today is from Tempest Next. Oh, yeah. The lung cancer study. Yeah, because it's a real-world, multi-center, perspective study.

2:16They didn't just look backward at old records. They actually deployed this across six different community health systems, tracking 662 patients with early-stage, non-small-cell lung cancer in real time.

2:29Wait, I want to highlight the community health systems part of that, because that is where the vast majority of you listening would actually get treated. Like, we aren't just talking about pristine, academic ivory towers with massive research budgets.

2:41No, exactly. And the setting is crucial. Tempest deployed an AI clinical decision support system, an AICDSS, to analyze unstructured electronic health record data. And the outcomes within 90 days of diagnosis were just substantial.

2:56What were the actual numbers on that? They saw a 24% increase in testing for the ALK biomarker, an 18% increase for EGFR, and a 13% increase for PD-L1. Okay, let me stop you there and clarify what those biomarkers actually are, because it changes the stakes entirely.

3:12ALK and EGFR aren't just, you know, random letters on a chart. They are specific genetic mutations within the lung cancer. Right, they dictate the whole treatment path. If a doctor knows a patient has an EGFR mutation, they don't have to use broad-spectrum, highly toxic chemotherapy.

3:28They can prescribe a highly targeted pill that specifically inhibits that mutated protein. So, the AI isn't diagnosing the cancer itself. No, it's not. It's auditing the workflow to make sure these life-saving tests just aren't forgotten by the humans.

3:43That's spot on. It acts as an omnipresent safety net. I mean, picture a community oncologist seeing 30 patients a day. They get a 50-page PDF of pathology results, type notes, lab values. That's just too much data for one person.

3:56Way too much. The human brain gets fatigued, and a crucial indicator buried on, like, page 14 might just be missed. But the AI is constantly reading that unstructured, messy text in the background. It flags the chart and alerts the physician like, This patient meets the criteria for a specific biomarker test that hasn't been ordered yet.

4:14It's essentially a hyper-vigilant medical detective looking over the doctor's shoulder. And it's actually changing clinical behavior. I read Tempest is already expanding this platform to breast, colorectal, ovarian, prostate, and urothelial cancers.

4:28Yeah, they're scaling it up fast. And it seems like this push from theory into practice is everywhere at ASCO today. I mean, I saw the Mayo Clinic alone is presenting over 30 oncology AI studies. Oh, Mayo's presence is massive this year.

4:41Yeah. They are presenting data on biomarker-driven bladder therapies, multi-cancer early detection, and an update on their red mode pancreatic model. Wow, 30 studies is a lot. Is there a common thread there?

4:54Yeah. The overarching theme of their 30 studies is analyzing the tumor microenvironment. They're moving beyond just looking at the cancer cells themselves, and using AI to analyze the surrounding tissue, the blood vessels, and the immune cells that feed the tumor.

5:08Now, that concept of looking at the environment around the tumor leads perfectly into what I think is the most fascinating abstract at the conference. Sylvester one. Sylvester Comprehensive Cancer Center at the University of Miami, they have an AI that is literally reading routine bone marrow biopsy slides for multiple myeloma patients.

5:28And myeloma is a cancer of the plasma cells in the bone marrow, right? It is, yeah. Traditionally, a human pathologist looks at a stained bone marrow slide under a microscope to quantify the obvious cancer cells. But the Sylvester AI is extracting hidden immune cell signals from those routine slides.

5:44Hidden signals, like things humans can't even see. Exactly. It's seeing patterns that human eyes simply cannot quantify, specifically regarding spatial biology. Spatial biology. Okay, so instead of just genetic sequencing, which is basically just reading the ingredients list of the tumor, spatial biology is looking at the baked cake.

6:02That's a great way to put it. It's looking at how the immune cells and the cancer cells are physically interacting with each other on the slide. That analogy hits the core mechanism perfectly. Because knowing a specific immune cell is present somewhere in the sample isn't enough.

6:16You need to know if it's actively infiltrating the tumor or just, you know, sitting passively on the sidelines. Right. Location matters. By analyzing those physical spatial relationships, this AI predicts whether a newly diagnosed multiple myeloma patient will benefit more from advanced immunotherapies, like daratumumab, or if they should bypass that entirely and go straight to a stem cell transplant.

6:41Wait, so instead of the traditional trial and error approach, where we give everyone the population average protocol, cross our fingers and wait six months to see if they survive, this AI is literally matching the therapy to the patient's unique immune biology on day one.

6:56On day one. That is the holy grail of precision medicine. I mean, that's incredible. It really is. But as we look at these brilliant performances from Tempest and Sylvester, we kind of have to address the growing friction. The messy hospital reality.

7:10Right. Because the logical question is, if these tools are so effective, why aren't they running in every single hospital across the country right now? Yeah, where's the bottleneck? Because clearly the compute power is there and the models are smart enough to do it.

7:24Well, Penn Medicine is presenting extensive data at ASCO, arguing that the bottleneck is increasingly the patient. The patient? How so? We are seeing massive patient-side gaps in AI literacy, trust, and equity of access.

7:39For instance, they're highlighting the intersection of GLP-1 weight loss drugs like Ozempic and breast cancer screening. Wait, how do weight loss drugs connect to AI cancer screening? OK, so GLP-1 drugs rapidly reduce body fat, and breasts are composed of fat and glandular tissue.

7:54When the fat is reduced, the breast tissue becomes radiographically denser on a mammogram. Oh, I see where this is going. Right. So if an AI screening model was trained on data from 2019 before the GLP-1 craze, it doesn't understand the sudden widespread shift in breast density across the population.

8:12It might actually misread the mammogram. Wow. And when patients learn that an AI is analyzing their scans, but might be confused by their weight loss medication, trust just evaporates. They also highlighted the Basar Center Patrol Study, which is tracking over 500 patients to refine MRI and age-specific PSA thresholds for genetic prostate cancer screening. Again, trying to build that trust.

8:32I see. Because if patients feel like they are being handed off to a brutal algorithm that doesn't understand their specific context, they'll just refuse the AI-assisted care. Exactly. And Penn is showing that if patients don't trust the tool, the smartest algorithm in the world is essentially useless.

8:49Yeah, but it actually goes deeper than patient trust. There is a severe technological trust issue among the clinicians themselves, highlighted in the new state of clinical AI report 2026 from the Stanford-Harvard ARISE Task Force.

9:04Oh, I read through that section. The report focuses on a terrifying phenomenon called model drift. Yeah, model drift honestly sounds like a massive liability. For you listening, I like to think of an AI model in a hospital like a hyperlocal weather app.

9:17It gets incredibly good at predicting rain in Seattle because it learned the specific atmospheric rules of Seattle. But if you suddenly move that exact same app to Miami without telling it, it's going to fail completely because the underlying data, the humidity, the temperature patterns have entirely changed.

9:33The weather app is a much better way to conceptualize it than simply saying the AI breaks because the ARISE audit found that predictive models that pass internal validation with flying colors lose accuracy incredibly fast once they're actually deployed.

9:47Just from moving them to a new hospital. Or even within the same hospital. The mechanism behind that drift is just the shifting reality of healthcare. A hospital might buy a new MRI machine from a different manufacturer with a slightly different pixel resolution, or the local demographics might change, or a new medical billing code is introduced.

10:06And because the AI was trained on the old reality, it drifts. It starts throwing false positives or missing diagnoses because the data it expects no longer matches the data it receives. Exactly. Without continuous PCCP style oversight and recalibration, these cutting edge assistants become actual clinical liabilities.

10:25The ARISE report is pushing for this because they recognize that an AI model is not a static piece of software like Microsoft Word. Right. It's a living statistical model that has to adapt. But there was another part of that ARISE report that really alarmed me.

10:38They flagged the rapid widespread use of something called Shadow AI. Oh yeah, that is a huge issue. This is where doctors and nurses are bypassing the hospital's official IT systems entirely and just using consumer large language models on their personal iPhones to help with their clinical work.

10:57Like they're pasting real patient notes into random chatbots. Which is a severe hyperexposure waiting to happen. Yeah. To understand why it's happening, you really have to look at the immense administrative burden on these clinicians.

11:10The software they are forced to use is just terrible. It's notoriously clunky. Electronic health records often require like 40 clicks just to generate a standard discharge summary. So desperation wins out over compliance.

11:23I mean, if a doctor realizes they can synthesize a complex 40 page patient history into a clean summary in 10 seconds by just taking a photo of it with their phone, they're going to do it. They just want tools that actually work. Right. And the core issue connecting model drift and Shadow AI is the fragmentation of healthcare data.

11:40Every hospital system, Epic, Cerner, custom databases, they all speak a different digital language. The data is just a mess. Which brings us perfectly to the structural fix that just dropped yesterday, May 28th. The journal NEJM AI published a review on a new open source framework called MEDs.

11:59By Matthew McDermott from Columbia and his colleagues. Right. That stands for the Medical Event Data Standard. I kind of look at MEDs as the USBC of clinical AI. The universal adapter concept is highly relevant here because the goal of MEDs is to standardize the underlying data infrastructure so developers can finally write the model once, run it anywhere.

12:17Right. Because before USBC, you had a different charting cable for every single device you earned. That is healthcare data right now. So how does MEDs actually standardize this massive chaotic web of hospital information? It strips away all the proprietary complexity.

12:32MEDs represents every single observation in an electronic health record as a typed event, breaking it down into just four basic pillars. Timestamp, code, value, and patient. Wait, so instead of a massive unreadable spreadsheet with hundreds of customized columns that only one hospital understands, it's just, when did it happen?

12:51What was the medical code? What was the test result? And who was the patient? Fundamentally, yes, that's it. And crucially, it stores this data in a small set of Parquet files. Parquet files? Yeah. Parquet is a columnar storage format that is incredibly fast and efficient for querying massive data sets.

13:09Unlike traditional row-based spreadsheets, it prioritizes algorithm transportability. And it's already being used across 21 institutions. It formats major data sets like Emicke, EICU, MedSysDV, and the UK Biobank.

13:22So they can all benchmark identically. So if you fix the data layer with MEDS, the models become much easier to transport and recalibrate, which directly fights that model drift we talked about. But all of this medical AI, the Tempest triage, the Sylvester spatial biology, the MEDS data processing, all of it is ultimately downstream of the general foundation models, the core engines of AI.

13:44And Anthropic just released a massive update, literally on the eve of ASCO. Anthropic shipped Clod Opus 4.8 yesterday. And keep in mind, this is just 41 days after they released Opus 4.7. 41 days! The pace of iteration is staggering.

13:59It really is. And the specs are wild. Agentic coding jumped from 64.3% to 69.2%. Multidisciplinary reasoning with tools rose to 57.9%. And its knowledge work ELO hit 1890.

14:12Plus the fast mode is 2.5 times faster at a third of the cost. OK, let's talk about what Opus 4.8 actually does. I know the benchmarks and the ELO ratings are higher across the board, but the capabilities themselves have fundamentally shifted.

14:24They have. The most groundbreaking introduction here is dynamic workflows. Opus 4.8 is no longer just a conversational chatbot where you type a prompt and wait for a single response. It is now designed to orchestrate hundreds of parallel sub-agents in a single session. I want to break down what a dynamic workflow actually looks like in practice for you listening.

14:40Let's take an oncology researcher sitting at ASCO today. Under the old paradigm, they might ask an AI to summarize one specific research paper. What does Opus 4.8 do differently? OK, so with a dynamic workflow, that researcher can give Opus 4.8 a high level, really complex goal. The AI acts as a project manager. It spins up a swarm of specialized worker bee agents.

15:05A swarm. I love that word for this. Yeah. So one sub-agent is deployed to scrape and read every single post or presentation published at ASCO today. A second sub-agent pulls the researcher's internal patient data, formatted in the new MEDS standard, of course. Of course.

15:19A third sub-agent cross-references the new ASCO findings against the patient data to find matches. A fourth checks the statistical validity of the connections. And finally, the manager AI grades the work of all those sub-agents and synthesizes a comprehensive report.

15:33So it's like instantly spinning up an entire department of expert interns who all work in parallel. And it's doing this faster and cheaper than the previous version. The financial markets are clearly reacting to this level of capability because there is a massive funding story breaking alongside this release.

15:51There is. TechCrunch and others are reporting that Anthropic is in talks to raise approximately $65 billion. $65 billion in a single funding round. Yes. Led by folks like Altimeter, Dragoneer, GreenOaks, and Sonoya. And this is at a reported post-money valuation of roughly $965 billion. That is approaching a $1 trillion valuation.

16:16That would more than double their valuation from just last month and potentially put them ahead of OpenAI. We should caveat that these are reported figures and the company hasn't officially confirmed the close of the round. But still, the scale of capital is unprecedented. Where is that $65 billion actually going? It's earmarked for two massive priorities.

16:33First is compute scaling. Likely funding a deal with SpaceX and Colossus One to build infrastructure we really haven't seen before. And the second. The broader rollout of a project called Mythos. Let's unpack Mythos. The preview via Project Glasswing is already public and they've partnered with about 50 major organizations.

16:51We're talking about AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, NVIDIA. Basically, the load-bearing pillars of the global digital economy. Yeah. They have everyone on board. Mythos is an AI system focused purely on cybersecurity.

17:06And in its first month to port across these partners, it found over 10,000 critical and high-severity vulnerabilities. 10,000 zero days and critical bugs in a single month. Wait, a zero day, for those who might not know, is a software flaw that the software creator doesn't even know exists yet. Meaning there are zero days of protection against it. Finding 10,000 of them is mind-boggling.

17:27It highlights the dual-use nature of this technology perfectly. Because out of those 10,000, 530 have been formally disclosed. And 75 have already been patched with public advisories. Anthropic is actually signaling broader general availability in the coming weeks.

17:41Which is wild because an AI capable of finding that many vulnerabilities is a catastrophic weapon if it falls into the wrong hands. Exactly. But in the right hands, however, it acts as the ultimate proactive immune system for the internet.

17:55That concept of an AI immune system brings us right back to the paradox of today's deep dive. When you look at the sheer horsepower of Opus 4.8 managing swarms of agents, and Mythos finding 10,000 hidden flaws in the global digital infrastructure, and then you look back at the hospital floor...

18:10The contrast is jarring, isn't it? It really is. We have artificial intelligence advanced enough to personalize multiple myeloma treatment by reading the invisible spatial biology of bone marrow. Yet, simultaneously, doctors are so overwhelmed and frustrated by their clunky hospital IT systems that they're resorting to shadow AI on their personal iPhones, just to summarize patient notes.

18:31It is the ultimate collision of the future and the present. It leaves us with a critical question to chew on. Will the future of medical AI be defined by billion-dollar top-down enterprise deployments like Tempest-Next and the MEDS data standard? Or is it going to be driven by a grassroots rebellion of desperate clinicians demanding frictionless consumer-grade tech right there in the exam room? It's an unstoppable force meeting an incredibly immovable object.

18:58We are going to find out very soon which paradigm wins or how painfully they merge. Since ASCO just opened today, and there is so much ground to cover, we want to know from you. Do you want a daily ASCO 2026 AI-focused mini-briefing tomorrow through June 2nd to track all these abstracts and deployments? Or should we just roll the big headline news into the regular daily deep dive? Let us know what you want to hear.

19:21Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.