The headline release was EvolutionaryScale's ESM3, trained on 2.78 billion protein sequences with 98 billion parameters. Its most striking result was esmGFP, a novel artificial green fluorescent protein that the company estimates would have taken nature roughly 500 million years of evolution to produce — generated through human-led chain-of-thought prompting. Larger ESM3 models solved twice as many atomic-coordination tasks as smaller ones, and the model was open-sourced.

Models kept getting bigger

The trend line is unambiguous. ProGen (2020) had 1.2 billion parameters and ProtBert 0.42 billion; ProGen2 and ProtT5 arrived in 2022 at 6.4 billion and 1.2 billion; ESM2 reached 15 billion in 2023; ESM3 hit 98 billion in 2024. Each jump in scale has come with better protein prediction accuracy — ESM C, released in 2024, posted improved structure-prediction results in the CASP15 challenge.

Four results worth remembering

• AlphaProteo (Google DeepMind) designed the first protein binders for several targets including VEGF-A, a protein linked to cancer and diabetes. Its binders hold together roughly 10 times more strongly than existing state-of-the-art designs, with some estimated to be up to 300 times more effective; for the viral protein BHRF1, 88% of its designed binders bound successfully in the wet lab.
• AlphaFold 3 extended the series beyond protein structures to interactions with DNA, RNA, ligands, and antibodies. On protein-ligand docking it reached 80.5% of predictions under 2 Å RMSD — and 93.2% when the binding pocket was specified in advance, against 59.7% for the Vina baseline.
• A Stanford virtual AI lab — a principal-investigator model, a scientific critic, and three specialist agents in immunology, computational biology, and machine learning — autonomously designed 92 nanobodies targeting SARS-CoV-2, with over 90% successfully binding in validation.
• Google's Connectomics team reconstructed one cubic millimeter of human brain at synaptic resolution: over 5,000 slices at 30 nanometers each, capturing around 57,000 cells and 150 million synapses.

The databases underneath

None of this works without public data. Since 2021 the major protein science databases have grown sharply: UniProt by 31%, the Protein Data Bank by 23%, and the AlphaFold Database by 585%. The infrastructure is old — PDB dates to 1971 and was the first open-access digital resource in the biological sciences, Pfam to 1995, STRING to 2000, UniProt to 2002 — but AI-generated entries are now what drives the growth. Within biological-sciences publishing in 2024, function prediction was the most researched AI-driven protein topic at 8.4% of papers, followed by structure prediction (7.6%), protein-drug interactions (3.0%), and synthetic protein design (0.7%).

Microscopy is following the same curve. Light-based microscopy foundation models doubled from four to eight in 2024, and while no electron or fluorescence microscopy foundation models were released in 2023, four of each appeared in 2024.

MedQA, introduced in 2020, draws over 60,000 clinical questions from professional medical board exams. OpenAI's o1 set a new state of the art at 96.0%, a 5.8 percentage point gain over the best score posted in 2023 and 28.4 points above where the benchmark stood in late 2022. Like other general-knowledge benchmarks, MedQA now looks close to saturation — the report treats that as a signal to build harder evaluations, not as proof of clinical competence.

Two randomized trials, one awkward finding

• In a 2024 single-blind randomized trial with 50 US-licensed physicians working through complex clinical vignettes, doctors given GPT-4 alongside conventional resources scored 76% — barely above the 74% of doctors using conventional resources alone. GPT-4 by itself scored 92%, a 16-percentage-point margin over unassisted physicians, and there was no time saving in either direction.
• A 2024–25 randomized controlled trial with 92 physicians looked at management reasoning rather than diagnosis: treatment decisions, risk-benefit trade-offs, patient preferences. Here GPT-4-assisted physicians did beat the control group, by about 6.5 percentage points — but GPT-4 alone still performed on par with the assisted physicians, and the assisted doctors spent slightly longer per case.
• The pattern across both is the same: excellent standalone model performance does not automatically transfer to human-AI teams. Closing that gap is a workflow, training, and interface problem, not a model-capability one.

Where AI is actually deployed: paperwork

The clearest real-world win of the year was ambient AI scribes — systems that listen to the physician-patient encounter and draft the note. Kaiser Permanente Northern California launched one in late 2023 and thousands of clinicians adopted it before the pilot ended. A Stanford study of a fully integrated, automated scribe found 55% average uptake among physicians, roughly 30 seconds saved per note and about 20 minutes less EHR time per day, with self-reported burden down 35% and burnout down 26%. Investment in ambient scribe technology reportedly reached almost $300 million in 2024.

Stanford Health Care also fully implemented an AI peripheral arterial disease screening model under its FURM (Fair, Useful, Reliable, Measurable) framework. It is expected to reach roughly 1,400 patients a year and operates without external funding — one of only two of six evaluated use cases to make it all the way through.

The data bottleneck nobody solved

More than 80% of FDA-cleared machine learning software targets medical image analysis, yet the training sets are small. The Cancer Genome Atlas — one of the most comprehensive public histopathology collections — holds 11,125 patient samples across 32 cancer types, and histopathology models are often trained on fewer than 1,000 samples when genomic or proteomic labels are required. The geography is narrow too: most US cohorts used to train clinical machine learning algorithms came from California (22), Massachusetts (15), and New York (14), with most states contributing none.

The scale gap against general-purpose AI is stark. GatorTron, a large clinical LLM built to extract patient information from unstructured records, was trained on 82 billion tokens; Llama 3 was trained on 15 trillion — nearly 182 times more. On the imaging side, RadImageNet contains 16 million image-equivalent tokens against roughly 6 billion for DALL-E, about 375 times more. MIMIC-CXR (377,000 images) and CheXpert Plus (around 226,000 radiographs) are important resources but still small beside ImageNet's roughly 14 million images.

5.5 — And then the ethics field got funded

The AI Index meta-reviewed thousands of medical ethics studies for this chapter. NIH grants for medical AI ethics projects went from 25 in fiscal 2023 to 337 in fiscal 2024 — after just 2, 3, and 7 grants in the three preceding years. Total funding followed the same shape, jumping from $16.3 million in 2023 to $276 million in 2024, an almost 17-fold increase in a single year. Whatever else 2024 was, it was the year the ethics of medical AI stopped being an unfunded side conversation.

What the literature is actually about

• Bias and privacy are the two dominant concerns in 2024, followed by equity, transparency, and trust. The ordering has flipped since 2020, when privacy was the more prominent topic and bias trailed it.
• Tool coverage is lopsided. OpenAI's GPT series was discussed in 86 medical AI ethics publications in 2024, up from 42 in 2023 — an order of magnitude more attention than Google, Meta, Anthropic, Mistral, Cohere, or xAI models combined received.
• Evaluation interest exploded alongside it. A PubMed search for 'large language model' returns 1,566 papers since 2019, of which 1,210 were published in 2024 alone against 353 in 2023. An early-2024 systematic review of NLP performance on healthcare tasks found over 500 papers, concentrated on medical knowledge (419) and diagnosis (178).

Synthetic data as the privacy workaround

Because the real data is scarce and sensitive, 2024 research leaned hard on synthetic data. Studies validated it for privacy-preserving clinical risk prediction — modeling lung cancer risk in ever-smokers from the UK Biobank with ADSGAN, PATEGAN, and DPGAN, where the synthetic distributions closely matched the real ones. A Nature study used a generative image model to optimize drug formulation in silico, correctly predicting the percolation threshold of microcrystalline cellulose in oral tablets. And fine-tuned classifiers augmented with synthetic data outperformed ChatGPT-family models at extracting social determinants of health from clinical notes, while showing less sensitivity to race, ethnicity, and gender descriptors.

Deployment still runs through a handful of vendors. EHR adoption reached roughly 90% for any system and 80% for certified systems as of 2021, and a 2023 American Hospital Association survey found predictive-model use concentrated in Epic, Cerner, and Meditech networks — with Epic, Cerner, and CPSI hospitals mostly running vendor-developed models. Whether AI-enabled health IT reaches rural and underserved settings, which face broadband and infrastructure limits, remains an open question.