“And if we want AI to improve care rather than just reward the best-resourced organizations, then governance has to be designed for the hardest settings, not just the easiest ones.”
really well said here. enjoyed this read a lot. this now one of the fields i’m most eager to see properly adopt ai.
Thanks so much for reading! I really appreciate the feedback.
I agree with your phrasing of “proper” adoption - the unfortunate fact is proper governance is not convenient. It is also not quickly profitable, though I think it can be extremely profitable if you design a robust system that clinicians really trust.
The main gap right now is that the people who are technically capable of building such systems often lack the clinical expertise, and vice versa. The combination of “clinician” and “exceptionally strong Python coder” and “project manager” and “machine learning expert” in one person is practically zero.
There are people like me who are taking a crack at it with the help of frontier AI models, but the lack of determinism when it comes to codebase execution is a real slowdown point. I can create plans that say exactly what functions need to be replaced and how (literally line-by-line, which is a level of precision most people like me don’t have) and the coding agent will STILL find a way to just straight up not do the task correctly.
How do you think someone like me, who has a very clear vision but very little coding experience, should tackle a huge project spanning multiple domains?
honestly this is the frontier. i think vibe coding is fantastic enabler. it really is the larger projects that span domains that are the toughest still. the large context leads to context rot.
newer 1m token models and better memory management techniques help, but for something like this you need to create a master planning document for your llm to refere to as you build. a planning mode really helps. separate agents for each project help as well.
“And if we want AI to improve care rather than just reward the best-resourced organizations, then governance has to be designed for the hardest settings, not just the easiest ones.”
really well said here. enjoyed this read a lot. this now one of the fields i’m most eager to see properly adopt ai.
Thanks so much for reading! I really appreciate the feedback.
I agree with your phrasing of “proper” adoption - the unfortunate fact is proper governance is not convenient. It is also not quickly profitable, though I think it can be extremely profitable if you design a robust system that clinicians really trust.
The main gap right now is that the people who are technically capable of building such systems often lack the clinical expertise, and vice versa. The combination of “clinician” and “exceptionally strong Python coder” and “project manager” and “machine learning expert” in one person is practically zero.
There are people like me who are taking a crack at it with the help of frontier AI models, but the lack of determinism when it comes to codebase execution is a real slowdown point. I can create plans that say exactly what functions need to be replaced and how (literally line-by-line, which is a level of precision most people like me don’t have) and the coding agent will STILL find a way to just straight up not do the task correctly.
How do you think someone like me, who has a very clear vision but very little coding experience, should tackle a huge project spanning multiple domains?
Thanks again for your readership and support!
right back at you. it’s been a great.
honestly this is the frontier. i think vibe coding is fantastic enabler. it really is the larger projects that span domains that are the toughest still. the large context leads to context rot.
newer 1m token models and better memory management techniques help, but for something like this you need to create a master planning document for your llm to refere to as you build. a planning mode really helps. separate agents for each project help as well.
I’m working (overstated…lol) on an app for clinics and govt. in Honduras…feel free to have a look. It’s really far from completion though…
https://github.com/lostlilbot/Medpak
https://medpak-5559.d.kiloapps.io/