Why is the cheese sandwich the downfall of AI carb counting, even with “3D scanning”?
A slightly rhetorical question, I’ll admit, but a pertinent one. Once again, I have been trying out the latest (read most heavily pushed on Facebook […]
AI
Five AI models, three users, one finding: the settings came from the textbook, not the data
We took Nightscout data from three people and gave it to 5 Large Language Models with a detailed prompt to see how they would respond.
The results weren’t what I was expecting.
Explore a Decade of Diabetes Insights with Diabettech AI
After 10+ years of documenting the diabetes landscape, Diabettech is launching diabettech.ai—a custom research tool designed to help you navigate a decade’s worth of articles and reports. Powered by an LLM interface, this tool allows you to “interrogate” the site’s data to find specific answers on everything from cures to n=1 observations.
Key highlights of the beta launch:
Targeted Accuracy: Built to stick strictly to Diabettech articles and comments to minimize “hallucinations.”
Research-Focused: Optimized as a precision tool rather than a standard chatbot to provide relevant article links.
Sustainable Design: Privately funded and optimized for efficiency to manage resource costs and query speeds.
Ready to dive into the archives? Visit diabettech.ai to start your research and share your feedback on the beta experience.
ChatGPT Health: Promise, Practicality, and the Reality for People with Diabetes
Artificial intelligence has been hovering around diabetes care for a while now, usually wrapped in glossy demos and ambitious claims. Carb-counting apps that “just work”. […]
Beyond Snaq-qing. Digesting your glucose patterns.
I spent a long time discussing Snaq’s use of AI for food recognition in the previous article. In this one, I’m looking at what it […]
Snaq. A diabetes AI to help you carb count. Badly.
Here we review Snaq, a diabetes management app that aims to simplify carb counting through food photo recognition. Despite its functionality, the app struggles with food identification and portion sizes, often leading to inaccurate outputs. Users must contribute data to improve its learning, raising concerns about its subscription cost versus effectiveness.
In conversation with… Unpacking the “Why” Behind the Errors
Here we discuss the LLMs design and purpose in the context of interacting with and advising on health data, and demonstrate why the two may be at odds with one another:
https://wp.me/p7O2EL-2BM
In conversation with… File uploads and interpretation. Where the wheels fall off!
In this final section of the “In conversation with” experiment, we provided “access” to complete Nightscout data for the LLM models.
Unfortunately, this is also where the wheels rather fell off, and whilst the results gave an initial appearance of credibility, upon further investigation it became clear that what was presented was rather less then effective and could have been very unsafe.
Read on to learn more: https://wp.me/p7O2EL-2BE
In Conversation with…. Reviewing user data and system settings using AI
In this article, we look at the abilities of the LLMs to review data charts in image form from Nightscout and see what they suggest in terms of changing treatments, and how the contextualise those changes.
As a key way that users are interacting with these models, this is a critical insight into what you might get.
It’s a big read with lots to go through, but we’d highly recommend taking the time. It revealed a lot of unexpected items that are potential gotchas…
In Conversation with…. Carb Counting with AI: A Deep Dive into How Different Models Stack Up
In the dynamic world of diabetes management, anything that can help us accurately estimate carbohydrates in our meals is a game-changer. Imagine a future where […]