5 min read

ChatGPT plugs into your medical records

ChatGPT plugs into your medical records
Nº 01 · The Lede OpenAI Field report

ChatGPT ingests medical records

ChatGPT ingests medical records
Fig. IOpenAI · Filed 24 Jul 2026.

OpenAI launched Health in ChatGPT, letting eligible U.S. users connect medical records and Apple Health so the assistant can reason over their actual labs, vitals, and history rather than generic symptoms. Built with input from hundreds of physicians, it treats health context as a first-class connector — the same plumbing (permissioned data linking across conversations) that's been powering Gmail and Drive integrations. Consumer LLMs now sit directly on top of clinical data streams that used to require a portal login and PDF export.

Read the source

Also discussed on X.

Foundation models predict CAR-T response
Fig. IIbioRxiv · Filed 24 Jul 2026.
Nº 02 bioRxiv Field report

Foundation models predict CAR-T response

Single-cell foundation models predict durable CAR-T response in a new bioRxiv preprint, holding up even when input cell-type annotations are noisy or wrong. The result moves scFMs (single-cell foundation models — large models pretrained on millions of transcriptomes) from generic embedding tools toward clinically-actionable predictors, and narrows the gap between benchmark performance and real-world messy labels that has kept most of these models off the trial-design table.

Read more
Protein LMs take a germline shortcut
Fig. IIIbioRxiv · Filed 24 Jul 2026.
Nº 03 bioRxiv Structural biology · Protein design

Protein LMs take a germline shortcut

Protein language models retrieve adaptive immune receptors partly by memorizing germline V-gene identity rather than learning true antigen-specificity signal, a new bioRxiv preprint shows. Anchors a debate reference point for the whole PLM-for-immunology field: benchmark wins on TCR/BCR retrieval need to be discounted by how much of the score comes from the germline shortcut — a pattern of inflated PLM scores we've tracked — before anyone claims the model has learned recognition.

Read more
Also Filed · Three Briefs from the queue
Nº 04 arXiv Agents · Infrastructure

Interpretable agent for surgical margins

Agent-guided concept discovery surfaces interpretable relational features for surgical margin assessment, aiming for pathologist-inspectable reasoning rather than a black-box classifier. Moves intraoperative-margin AI toward the auditability floor regulators are starting to expect.

Read
Nº 05 arXiv Clinical AI · Evaluation

Gene expression from clinical images

M3-Gen generates gene-expression profiles from paired clinical and imaging data, with an interpretability layer over the multimodal fusion. Pushes image-to-transcriptome inference — a fast-moving capability frontier — toward outputs a molecular pathologist can actually check.

Read
Nº 06 X Field report

Physician-informed health UX

The Health rollout was shaped by early testers and physicians, echoing the medical-records launch above: OpenAI is signaling that clinician-in-the-loop design is now the product story for consumer health AI, not a compliance afterthought.

Read

Reply with your discoveries. A human reads them. Forward freely.

Agentic Discovery  ·  Nº 63  ·  24 Jul 2026

Editor's Note

OpenAI wires ChatGPT into personal health data today — plus foundation models keep chipping away at CAR-T and immune-receptor prediction.

 

Nº 01 · The Lede  —  OpenAI  —  Field report

ChatGPT ingests medical records

ChatGPT ingests medical records

Fig. I  OpenAI · Filed 24 Jul 2026.

OpenAI launched Health in ChatGPT, letting eligible U.S. users connect medical records and Apple Health so the assistant can reason over their actual labs, vitals, and history rather than generic symptoms. Built with input from hundreds of physicians, it treats health context as a first-class connector — the same plumbing (permissioned data linking across conversations) that's been powering Gmail and Drive integrations. Consumer LLMs now sit directly on top of clinical data streams that used to require a portal login and PDF export.

Read the source →

Why it matters

The 300M-people-per-week health-question funnel now has a persistent, record-aware backend — which resets the reference point for what a general-purpose chatbot is allowed to know about a patient, and forces every clinical-AI vendor to answer why their siloed tool is better than a consumer product that just read the whole chart.

 

Nº 02  —  bioRxiv  —  Field report

Foundation models predict CAR-T response

Fig. II  bioRxiv · Filed 24 Jul 2026.

Foundation models predict CAR-T response

Single-cell foundation models predict durable CAR-T response in a new bioRxiv preprint, holding up even when input cell-type annotations are noisy or wrong. The result moves scFMs (single-cell foundation models — large models pretrained on millions of transcriptomes) from generic embedding tools toward clinically-actionable predictors, and narrows the gap between benchmark performance and real-world messy labels that has kept most of these models off the trial-design table.

Read more →

 

Nº 03  —  bioRxiv  —  Structural biology · Protein design

Protein LMs take a germline shortcut

Fig. III  bioRxiv · Filed 24 Jul 2026.

Protein LMs take a germline shortcut

Protein language models retrieve adaptive immune receptors partly by memorizing germline V-gene identity rather than learning true antigen-specificity signal, a new bioRxiv preprint shows. Anchors a debate reference point for the whole PLM-for-immunology field: benchmark wins on TCR/BCR retrieval need to be discounted by how much of the score comes from the germline shortcut — a pattern of inflated PLM scores we've tracked — before anyone claims the model has learned recognition.

Read more →

 

Also Filed  ·  Three Briefs from the queue

Nº 04  —  arXiv  —  Agents · Infrastructure

Interpretable agent for surgical margins

Agent-guided concept discovery surfaces interpretable relational features for surgical margin assessment, aiming for pathologist-inspectable reasoning rather than a black-box classifier. Moves intraoperative-margin AI toward the auditability floor regulators are starting to expect.

Read →

Nº 05  —  arXiv  —  Clinical AI · Evaluation

Gene expression from clinical images

M3-Gen generates gene-expression profiles from paired clinical and imaging data, with an interpretability layer over the multimodal fusion. Pushes image-to-transcriptome inference — a fast-moving capability frontier — toward outputs a molecular pathologist can actually check.

Read →

Nº 06  —  X  —  Field report

Physician-informed health UX

The Health rollout was shaped by early testers and physicians, echoing the medical-records launch above: OpenAI is signaling that clinician-in-the-loop design is now the product story for consumer health AI, not a compliance afterthought.

Read →

 

· · ·

Reply with your discoveries. A human reads them. Forward freely.