Algorithms & Adipose - Clinical AI, Explained

In this episode we talk about agents - what happens when AI stops answering and starts doing, why errors compound across steps, and why the threshold for letting it act is far higher than for letting it advise.
Paper: Jiang Y, et al. MedAgentBench: A Realistic Virtual EHR Environment to Benchmark Medical LLM Agents. arXiv:2501.14654. 2025.
Substack: https://algoadipose.substack.com
Music credit: NCS Dream – Di Young

In this episode we talk about multimodal AI - how a model "sees" an image, why that's the wrong word for what it does, and why it can give you the right answer for entirely the wrong reason.
Paper: Jin Q, et al. Hidden flaws behind expert-level accuracy of multimodal GPT-4 vision in medicine. npj Digit Med. 2024;7:190.
Substack: https://algoadipose.substack.com
Music credit: NCS Dream – Di Young

In this episode we talk about benchmarks and why "it passed the boards" means less than you think. Contamination, obsolescence, and how to read an AI paper's methods like a clinician.
Paper: LiveMedBench: A Contamination-Free Medical Benchmark for LLMs with Automated Rubric Evaluation. arXiv:2602.10367. 2026.
Substack: https://algoadipose.substack.com
Music credit: NCS Dream – Di Young

In this episode we talk about the three ways to make AI "medical", fine-tuning, retrieval, and prompting, what each one actually does, and how to tell them apart when a vendor is selling to you. Plus why a general model, prompted well, beat the specialized medical one.
Paper: Nori H, et al. Can Generalist Foundation Models Outcompete Special-Purpose Tuning? Case Study in Medicine. arXiv:2311.16452. 2023.
Substack: https://algoadipose.substack.com
Music credit: NCS Dream – Di Young

In this episode we talk about training vs. using - how these models are built, why they have a knowledge cutoff, and why they never actually learn from your corrections no matter how many times you correct them.
Paper: Singhal K, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-180.
Substack: https://algoadipose.substack.com
Music credit: NCS Dream – Di Young

This episode: Why a prompt isn't a request but a set of conditions, why "be accurate" accomplishes nothing, and why letting a model say "I don't know" is the highest-leverage move you can make.Paper: Pandit S, et al. MedHallu: A Comprehensive Benchmark for Detecting Medical Hallucinations in LLMs. EMNLP 2025.

Substack: https://algoadipose.substack.com

Music credit: NCS Dream – Di Young

This episode: retrieval-augmented generation - giving the model a library instead of a memory.

Paper: Zakka C, et al. Almanac — Retrieval-Augmented Language Models for Clinical Medicine. NEJM AI. 2024;1(2).

Substack: https://algoadipose.substack.com

Music credit: NCS Dream – Di Young

This episode: hallucination. The model produces a correct answer and a completely fabricated one by exactly the same process.

Paper: Pal A, et al. Med-HALT: Medical Domain Hallucination Test for Large Language Models. CoNLL 2023.

Substack: https://algoadipose.substack.com

Music credit: NCS Dream – Di Young

In this episode we talk about tokens and context - the model's working memory. What's in it, what falls out, why long chats degrade, and why a bigger context window is not the same as a bigger attention span.

Paper: Liu NF, et al. Lost in the Middle: How Language Models Use Long Contexts. TACL. 2024;12:157-173.

Substack: https://algoadipose.substack.com

Music credit: NCS Dream – Di Young

what actually happens in the seconds after you hit enter after typing your prompt.

Paper: Goh E, et al. Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial. JAMA Netw Open. 2024;7(10):e2440969.

Substack: https://algoadipose.substack.com

Music credit: NCS Dream – Di Young