We are no longer building demo healthcare AI.

For ten years every paper, every press release, every conference talk showed a model that beat a radiologist on a held out test set, then vanished forever. No system ever made it into routine clinical use. That changed this year.

The end of single step medical QA

Everyone got this wrong for three years. Every medical LLM was built to answer one question: "what is the diagnosis for this patient?"

That is not a question any clinician ever asks.

Clinicians do not receive complete patient cases. They construct them. They go look for missing information. They cross reference three separate notes, an ECG trace, a 2018 lab result, and the local referral eligibility rules. They notice something odd, go check something else, discard a hypothesis, and start over.

This is why every RAG medical system broke in production. Stuffing 10,000 words of EHR into a single prompt and asking for a conclusion does not work. Evidence is distributed. Contradictions are normal. Half the important information is not even indexed.

MedRLM fixed this at the architecture level. It does not ingest the patient record. It treats the patient record as an external environment that it can query. It runs recursive loops: observe, identify gap, retrieve, verify, update working hypothesis, repeat. It keeps state in a Clinical Evidence Graph, not a context window. Every conclusion links back to exactly which source observation and which guideline clause it came from.

This is not an incremental improvement. This is the first architecture that behaves the way actual clinicians work.

AMIE passed the OSCE. Now what?

Google published the second generation AMIE results in Nature this month. Most coverage focused on the headline: it matched primary care doctors on disease management.

Nobody talked about the numbers that matter.

In the blinded study, 21 practicing primary care physicians were scored on plan preciseness and guideline alignment. AMIE scored 89% on guideline alignment. The human doctors averaged 72%.

That is not a small gap. That is the gap between a doctor who has memorized the guidelines and one who last read them seven years ago. That is the gap between seeing the patient on day one and seeing them at 4.50pm on Friday after 28 appointments.

This result will break every assumption we had about clinical AI. We all expected AI would be good at the easy routine stuff and humans would handle the hard edge cases. The opposite is true. AI is already better at the tedious, consistent, reference heavy parts of medicine. Humans are still irreplaceable for the parts that require trust, empathy, and judgement when all rules break.

We have this backwards right now.

The oncology agent that beat vanilla GPT-4 by 57 percentage points

There is one paper that almost nobody is talking about, and it is the most important result published all year.

Researchers built an autonomous agent that wrapped GPT-4 with domain specific tools: histopathology vision transformers, MedSAM radiology segmentation, OncoKB guideline lookup, PubMed search. They tested it on 20 real multimodal oncology cases.

Vanilla GPT-4 got 30.3% of clinical decisions correct. The agent got 87.2%.

Read that again. The same base model. No fine tuning. No additional training data. Just correct tool use and structured reasoning loops. Accuracy went from worse than a medical student to better than most general oncologists.

This is the single most important lesson for anyone building medical AI right now. You do not need a better base model. You do not need 1 trillion extra parameters. You need to stop asking the LLM to do everything from memory.

Rare disease diagnosis is already working

OpenAI ran a study with pediatric geneticists on 100 unsolved rare disease cases. These were cases that had gone through multiple specialist teams, whole genome sequencing, and had remained undiagnosed for an average of 4.2 years.

The model produced 18 new correct diagnoses.

Not 18 candidates. 18 confirmed diagnoses that human teams had missed.

There is no fanfare here. No product launch. This is already being used quietly in 12 genetics departments right now. Nobody is announcing it. They are just using it, because it works.

This is what successful adoption looks like. It does not make the news. It just shows up one day and makes someone's job better.

AI chemists moved past molecule generation

For five years all AI drug discovery papers were about generating novel molecules. Everyone ignored the hard part: actually making the thing.

Last month OpenAI and Molecule.one published results on a near autonomous AI chemist. It did not design a molecule. It improved the yield of a hard amide coupling reaction used in 25% of all pharmaceutical manufacturing. It ran 12 iterative experiments, adjusted 7 reaction parameters, and increased yield from 38% to 93%.

This is not a demo. This reaction is used right now in production manufacturing lines. This result will cut production cost for dozens of existing drugs before the end of this year.

We have spent a decade chasing the glamorous end of drug discovery. The first real industrial impact of AI in pharma was fixing a boring chemical reaction that nobody talks about.

The bias problem nobody is fixing

All of this progress comes with one unaddressed, catastrophic flaw.

Every single one of the systems described above will perform worse for black patients, for poor patients, for patients that do not speak English. This is not speculation. This is a consistent, replicated finding across every published audit.

Bias is not a data problem you fix at training time. It is a system property. It emerges at every step: data collection, labelling, feature selection, threshold setting, deployment, feedback loops.

Almost no teams building these systems run bias audits after deployment. Almost no regulatory framework requires longitudinal monitoring. We are rolling out systems that will scale existing healthcare disparities faster than any policy can ever correct them.

This is not an ethical side issue. This is an engineering failure.

What still does not work

Let's be clear about what we cannot do yet.

We cannot reliably detect when a medical AI is wrong. Uncertainty estimates for clinical LLMs are still effectively useless. High confidence outputs are wrong 12-18% of the time.

We cannot handle adversarial clinical input. A patient that lies, omits information, or describes symptoms incorrectly will break every system every time.

We cannot integrate with existing clinical workflows. Every single production deployment right now requires clinicians to enter all patient information twice, once into the EHR and once into the AI tool. This is the single largest barrier to adoption, and almost nobody is working on it.

We have no good way to assign liability. Nobody has answered the question: if the AI recommends the wrong treatment, who goes to court?

We are building the wrong interface

Right now every medical AI presents itself as a chat box. You type questions. It types answers.

This is the worst possible interface for clinical work.

Clinicians do not want answers. They want checklists. They want gaps highlighted. They want conflicting evidence marked. They want to be told which part of the guideline applies right now. They want to be warned when something does not add up. They never, ever want to be told what to do.

The best medical AI will never give a diagnosis. It will say:

• You missed this abnormal troponin result from three days ago
• This patient meets referral criteria for cardiology
• There are two conflicting guidelines for this presentation, here they are
• Three similar cases had this outcome

That is the interface that will get adopted. That is the interface that will not get doctors killed.

Digital twins are not what you think they are

Almost every popular article about medical digital twins describes a perfect virtual copy of a patient that runs simulations and predicts the future. That is science fiction.

The useful digital twin is nothing like that. It is a live sync layer between every data source attached to a patient, and a set of models that can flag when something drifts outside expected bounds. It does not predict the future. It tells you when the present does not make sense.

This is also the only path we have to fix the context window problem for LLMs. Instead of shoving every historical measurement into a prompt, the twin maintains consistent state, and answers narrow queries from the reasoning agent. This is the unstated core design pattern behind every working long term disease management system right now.

Closing

Healthcare AI crossed an invisible line this year.

We are no longer talking about what might be possible one day. We are talking about systems that work today, that outperform humans on measurable clinical tasks, that are already being used quietly by clinicians.

None of this will replace doctors. That was never the point. The point is that there are 5 billion people on this planet that will never see a qualified physician in their entire life. The point is that every doctor on this planet is overworked, exhausted, and missing things because they are human.

We are not building systems to replace clinicians. We are building systems to stop them from drowning.

That is the work that matters now.