A Text-Only AI Outdiagnosed GPs. Google Still Won't Ship It.

Nathan Labenz welcomes Khaled Saab of Google DeepMind and Vivek Natarajan of Google Research for Natarajan's third appearance on *Cognitive Revolution*, the most of any researcher on the show. The plan is a tour of five papers: radiology report generation, diagnostic dialogue, Med-Gemini, a multimodal extension, and a new therapeutics model that mixes chemical notation with language. Labenz's standing pitch for AI, which he restates up top, is an AI doctor on a personal device, day or night, at a tiny fraction of first-world cost. Some of the human-competitive numbers they will discuss, he notes, still sit on Flamingo, a model Google published in April 2022.

The radiology paper fine-tuned that early multimodal model to write full natural-language reports, not just classify a pathology. Natarajan calls Flamingo one of the original few-shot vision-language systems, and says the work had been a couple of years in the making. Geoffrey Hinton's old line that AI would replace radiologists has become a bit infamous; Natarajan still thinks he was not far off. Labenz supplies the shortage in human terms: his wife's cousin, a physician, said her hospital can wait as long as 60 days for a scan to be read. Replacement debates, in the short term, sit on top of a staffing hole.

The result that matters, Natarajan says, is the clinically relevant task of report generation. They took a relatively modest vision-language model by today's sizes, fine-tuned it on high-quality radiological data from the open MIMIC repository, and watched performance approach what radiologists can do. The more interesting arm was collaboration: when AI drafts went to human radiologists for edits, the composite system beat AI alone and humans alone. That is the first empirical nod he cites to Stanford AIMI's Curt Langlotz: AI will not replace radiologists, but radiologists who use AI will replace those who do not.

Eighteen months of shelf life between Flamingo and Flamingo-CXR is long in foundation-model years. The data work underneath it is why. Naive reports refer back to history the model cannot see, so the team has to strip those references and teach the system to reason only from the image in front of it. On Med-PaLM they had used the same datasets without the cleaned reports curated by Pranav Rajpurkar and colleagues at Harvard, and they paid for it. Combine that cleanup with a relatively small language model, Natarajan says, and performance can match larger ones. Saab adds the evaluation tax: generative systems no longer output a single label, accuracy is the wrong metric, and the gold standard is a panel of clinical experts, which is slow and, in medicine, necessary.

The paper Labenz now puts in almost every slide deck is AMIE. The base model is PaLM 2, a generation behind Google's latest, which he notes would still have been revelatory if dropped into 2020 or even 2022. Through a text-only chat, the system produced more accurate differential diagnoses than human general practitioners. Other doctors gave it higher marks, and simulated patients scored it higher on empathy and on making people feel heard. Labenz says he cannot get that result out of his head.

The team was surprised too. After Med-PaLM, a group of them sat around and asked how long diagnostic medical conversation would take. The unanimous guess was at least two years. Six months later, AI was exceeding GPs both on New England Journal of Medicine case challenges and on diagnostic dialogues. Until AMIE, Natarajan says, models got a clean case summary and were asked for a diagnosis, a grand challenge that still is not everyday practice. A real visit is investigation under uncertainty: history, follow-up questions, labs, then a differential and a plan. He had not seen language models proven at that kind of work in any field.

ChatGPT and Gemini, he argues, still let the user drive. In clinic the model has to drive: pick the next question, acquire information efficiently, and walk the patient through a journey. That is a different skill, and it was not obvious that systems could be trained to do it. The reason they could, in his telling, is a simulation learning environment that lets self-play generate orders of magnitude more text encounters than any human doctor will see. He immediately cools the headline. The comparison is not entirely fair: doctors are unused to text, they convey empathy with face and voice, and AMIE never tires after a long clinic day. The progress, he says, is still undeniable.

Saab walks through how they trained a model to run a diagnostic conversation: gather what the patient is going through, ask follow-ups, then land an accurate diagnosis and perhaps a plan. The first instinct was obvious. Collect real doctor-patient visits, transcribe them, train. The first AMIE trained that way did not have the conversations they wanted. Transcripts were noisy: awkward pauses, pointing, utterances that lose their meaning on the page. So they built a multi-agent synthetic data framework instead, inspired by work showing capable LLMs can curate high-quality synthetic data.

Three agents sit in the loop. One plays the patient from a generated profile of symptoms, social history, medical history, and family history. One plays the doctor. A third, the critic, is instructed on what a high-quality dialogue looks like, then feeds comments back for another generation round. Coverage comes from a vignette generator: tens of thousands of patient profiles via web search and other tools. After skipping the sponsor break, Saab explains they start from the same PaLM 2 checkpoint with role prompts, then fine-tune AMIE on both patient turns and doctor turns so the same model gets better at both sides. That outer self-play loop, regenerate data as the model improves, is how the system compounds.

Tens of thousands of profiles sounds tiny next to a 15-trillion-token pretrain, Labenz notes. Natarajan's reply is that they are not training from scratch: PaLM 2 has already seen a similar order of tokens to Llama 3 or GPT-4. The right comparator is a human career, which back-of-the-envelope is tens of thousands of encounters. Cross those profiles with socioeconomic status, medications, travel, and personalities that are talkative, worried, or adversarial, and the combinatorial space hits millions or hundreds of millions. Conversations roll into the millions; they stopped a checkpoint for the paper after seven outer-loop rounds, which Labenz says is past the five where most critic-loop papers top out. The trick is that every round must add net-new information, or the model saturates.

Labenz asks the world-changes question: a new COVID variant, a new fact. Do you rerun everything? Natarajan says each self-play round already expanded the vignette pool, and a new variant can be added to the fine-tuning mix. The cleaner long-term answer is Med-Gemini's tool use, especially web search, so real-time retrieval sits on top of clinical reasoning. Saab wants both: keep training on new information, and use search for the fact that appeared yesterday. The medical-specific piece Natarajan names last is scalable oversight. Models need honest uncertainty, the ability to tell the patient they do not know, and the willingness to back off and ask an expert. That behavior, he says, is still being studied, and it is the missing piece for safe deployment.

Labenz expected the obvious next paper after AMIE: a multimodal visit, video chat, images in the thread. Med-Gemini is not that. It is a family of models, some on Gemini 1, some on 1.5, maybe still a Flamingo in the mix, each dialed into a narrow task across pathology, scans, genetics, 2D and 3D imaging. Why the Swiss Army knife? Natarajan treats Med-Gemini as the Med-PaLM analog on a much stronger base, with a new headline: native multimodal understanding over million-plus token windows. Med-PaLM had USMLE as the public hook. Med-Gemini does not have one trick. It is a platform of newly feasible tasks.

The generalist they proved last year, he says, was a research question. Robotics had a few examples; medicine did not. Once they started asking how to deploy, the trade-offs arrived: which data, which tasks, latency, throughput, model size. Cramming everything into the largest model is often not worth the cost for a given application. What the real world wants is a menu. Some jobs need almost no reasoning and very low latency, so they keep small models. Some need advanced clinical reasoning and million-token context, and those get a different option. The research challenge of a journalist, or generalist, model no longer felt unsolved. The product challenge did.

Saab frames the paper as a first exploration of how to specialize Gemini for medicine: textual reasoning with an agentic web-search framework, medical images because the base is multimodal-first, and then the 1-million and now 2-million context breakthrough. That unlocked stuffing multiple EHR logs into the window, talking to a surgical video, and asking questions across 12 genetics papers. Qualitative multimodal conversation, not just one-shot Q&A, is what he thinks will translate clinically. An AMIE-style paper on this stack is coming, he says, but it needs slower expert evaluation. Natarajan admits the field's benchmarks are falling behind the models. A 6% bump on MMLU is a tiny fraction of what the system can do, and everyone is stuck mixing static numbers with demos until better measures exist.

Labenz pushes back on cost anxiety. Gemini 1.5 Pro, he notes, is $7 per million input tokens, and he would pay $500 a month for something that works, still an order of magnitude under what it replaces. Natarajan agrees prices are falling and that Flash, then on-device Nano, is the capabilities-versus-cost sweet spot. Google's habit, which he had to learn, is not to aim at the top 1 percent. Think billion-user scale. In California $500 a year is not high; in large parts of India and Africa it is a no-go. If pricing becomes a barrier for 99 percent of people, they have amplified disparities in care instead of closing them. That is the non-goal.

Saab adds that text AMIE is already dramatically cheaper than a clinical expert or a night of self-research, and that video or audio would change the bill. Labenz shares the accessibility vision, then asks the sharper path questions. Why not Tesla's playbook: sell an expensive version first and subsidize the cheap one? Google has the balance sheet to skip that. And from the other end: many people in the world cannot do better than AMIE today. At what point does sitting on it become a moral failure, especially if the comparison class is Stanford rather than no doctor at all? He cannot imagine a 2027 in which the work continues and deployment is still optional.

Saab remembers raising his hand at an internal summit soon after joining Google: language-model services are free, why not an AI doctor? The answer taught him the subtleties of FDA regulation and medical evaluation. Natarajan says the dilemma is the single biggest question that keeps him up at night. They can build societally powerful tools; they have to communicate progress without pretending capability is reliability. Papers and products carry disclaimers for a reason. Viral tweets about ChatGPT beating a doctor are anecdotes. Regulators will not move on anecdotes.

The development path he wants looks like drugs. Start in simulation, which for AMIE meant patient actors, not real patients. Then put the system into real clinical workflows as an intervention, read out the data, and check safety before efficacy. Step capabilities up through phases. He hopes it does not take ten years. Dumping unproven systems on parts of the world that already lack the best resources, the places he comes from, would be the wrong kind of ethics. Testing safety itself needs oversight: a doctor ready to step in by phone if AMIE goes wrong. Those resources, today, live mostly in well-resourced Western organizations, which is why they are partnering there first, with announcements he says should come soon.

There are no shortcuts, he says. Take them and you set the field back decades. And everything in healthcare moves at the speed of trust among patients, doctors, and regulators. Doing it right, counterintuitively, gets you to the future sooner. Labenz admits AI has scrambled his own politics: he is afraid of pedal-to-the-metal scaling and still wants his AI doctor even if it is only on par with a human. Waiting for 10x better may be the only way to buy that trust. Societal acceptance is the other slow variable. Natarajan's mother mostly meets AI as the wall between her and an airline human. He would not be surprised if only a tenth of the population has really used GPT-4 or Gemini. Technology that has to work at a billion users still has decades of diffusion in front of it.

Labenz had teed up uncertainty-guided search and never closed it. Naive tool use, Saab says, can hurt. If Med-Gemini writes search queries for every question, fetches the web, and stuffs the results in, irrelevant facts give the model more ways to latch onto the wrong thing. Two fixes made it work. First, generate multiple answers, find the conflicts, and write queries aimed at those disagreements, so search targets what the model is actually confused about. Second, train with search results in context so the model learns to extract the useful span and not over-rely. It is not a token-level perplexity flag. It is a contrast of full responses, which is also a natural escalate-to-human signal.

Natarajan calls the current recipe unoptimal: they would rather not regenerate every time, and they want verbalized uncertainty from the model itself. Search is only the first tool in a coming agent stack that will talk to more specialized systems. For this paper, the instantiation was tuned to benchmarks. He also mentions, almost as color, that Saab's wife Kiren has video of him coding the work at gas stations over three months.

The Med-Gemini extension paper is where Labenz reads the radiology numbers aloud. Across two chest X-ray datasets, AI reports beat the original radiologists' reports by absolute margins of 1 and 12 percent. On normal cases, 57 and 96 percent of AI reports were judged equivalent or better; on abnormal cases, 43 and 65 percent. That, he says, is hard to read as anything but clinical utility for 2D chest films. They also report the first large multimodal-model generation of 3D CT reports, with 53 percent considered clinically acceptable. All of this, as he understands it, is post-training with distinct encoders, not CT scans dumped into pre-training.

Will natively multimodal soon mean every modality in the pretrain? Natarajan wants generalized encoders: one 2D encoder for natural and medical images, one 3D encoder for video and CT, not hundreds of specialists. Today you still get a boost from specialized encoders, so adapters and post-training hang on. If a 3D encoder can handle CT, MRI, and video equally well, more of that data can mix into pre-training. They are not there yet. Saab agrees general encoders would be nice and still wants medical specialization for behavior: how you follow up with a patient in a diagnostic conversation, even after the pixels are solved.

The last paper on Labenz's list is TX-LLM: chemical structure represented as the C-H-O strings every premed student remembers, woven into training text beside natural language. He wonders whether interpretability will later find higher-order concepts in those weights that humans did not name. Natarajan steps back to motive. Fleming and Jonas Salk were practicing clinicians who saw patients by day and ran lab work at night. The mission is medical superintelligence that democratizes care; the process feels like encoding the biomedical universe, from DNA, RNA, and protein up through imaging, EHRs, and population health. The composite system he wants is a hybrid AI physician-scientist: better access, better biology, better therapies, personalized care at scale.

TX-LLM is an older, Med-PaLM-era, text-only instruction-tuned model, not a conversational one. The result is that a single generalist can cover the therapeutics pipeline from target identification to clinical trials. Some tasks still lag specialist models; transfer across therapeutic modalities is the interesting part. They have not done enough work, he says, to unearth unique internal insights. A separate discovery line uses generative output rather than those representations. Hallucination is a problem; the flip side is creativity, and science needs creativity. Late last year, with Stanford collaborators, the models hypothesized a causative gene for a mouse hearing-loss phenotype; CRISPR knockouts validated it. They have since pointed the same method at human variants of unknown significance, the rare and undiagnosed diseases where doctors are scarcest.

Labenz sees an AI scientist on the Kurzweil-ish curves of the late 1990s: foundation models plus agentic recovery plus calls to AlphaFold-class tools. He is inspired and a little afraid; he would not bet on out-discovering it over the next couple of years. Natarajan answers with fire, steam, electricity, nuclear: dual use is old, society eventually figures out beneficial use, and he is hopeful. Saab names the blocker. Right now a hypothesis still goes to a wet lab for months. A biological simulator would make the loop much faster. They are deliberately not wiring the system to a robotic lab, Natarajan adds, because they do not yet know how to control it safely. The expert stays in the loop.

Labenz states his public position: adoption accelerationist, hyperscaling pauser. He wants the AI doctor and the AI coder quickly, and he is worried about a few more orders of magnitude producing something genuinely superhuman. Interpretability is going better than he expected eighteen months ago and still feels behind the scaling drumbeat. The concrete test for the guests: if Gemini 1.5 Pro, or a 1.5 Ultra, were the best model they would ever get, could post-training, validation, and elbow grease still deliver the AI-doctor vision?

Saab does not see why superhuman tools should scare us by default. Calculators already beat math PhDs at arithmetic; computers already beat us at recall. As long as the tools are controllable and beneficial, better-than-human on some axes is fine. He can wear both hats in the same week: AMIE is mind-blowing, and the systems still fail on basic things. Hallucination is what keeps him up, including in AI doctors. You can do a lot of good by building AMIE on Gemini 1.5 if progress paused there. Closing the dumb error modes may still need continued model improvement, and he does not know whether scale is the fix.

Natarajan answers directly. Yes, you can build a system that democratizes care with current LLMs. He takes inspiration from Waymo: no AGI, functioning self-driving, years of simulation, safety investment, and systems engineering. Medicine needs the same, plus process supervision to improve the reasoning path, plus data that covers the long tail. With today's base capabilities it will take more time; with stronger models, less. If progress stalls at Google, they keep pushing the vision on what they have and simply arrive later. Labenz hears that as confirmation that a couple of years of integrating known techniques may matter more than the next training run. He wants that fact in the room while people talk about a China arms race and trillion-dollar clusters. Flamingo is a spring 2022 model, and it already writes useful radiology reports.

If scaling plateaus, Natarajan says, what changes is moats, go-to-market, and commercialization, not the possibility. These systems are no longer science fiction. He asks for rational optimism rather than hype. Saab closes with a smaller origin story. Listening to Labenz interview Natarajan about Med-PaLM taught him the papers in enough depth that he could impress his future collaborator in the Google interview. He thanks the show for democratizing AI education, then both guests sign off.