Hacker News new | ask | show | jobs
by ungreased0675 10 days ago
I’m skeptical that general purpose LLMs are a good fit for a very specialized medical analysis task. Something trained specifically for the task would be the path I’d explore.
4 comments

Does it say anywhere that he’s using general purpose models for the analysis? Fine tuning open weight models is generally available for pretty minimal cost, I’d say his reference to vibe coding is how he is building the software not how the software functions.
General models are very cheap to get started with though. So even if they are less than ideal, you can use them to get a company going and then make something more efficient.
> I’m skeptical that general purpose LLMs are a good fit for a very specialized medical analysis task

If they can save more lives without harming other lives I would gladly take it during the analysis. Even saving 1% more lives is an amazing achievement!

How will this lead to saving lives? The only possible consequence is abortion?
If a couple gives birth to a baby with a debilitating condition and dies after 6 weeks, that’s an additional baby death. If instead of having that baby, a couple gives birth to a healthy baby, it will likely survive into adulthood, meaning there’s one less baby death.

It sounds like you just take issue with abortion. That’s more of a you problem than a problem with quantifying baby deaths.

This is an anecdotal example, isn’t it? What do people today abort for, after genetic diagnosis? Down-Syndrome and other markers with highly random and non-fatal outcomes.

I have no problem with abortion itself.

in some cases early and correct diagnosis saves lives. may not be applicable for the OPs situation, but imagine running full diagnosis is so cheap at some point in the future that every kid when they born get that diagnosis cheaply and automatically. And doctors do not need to "guess" based on symptoms.
I think author meant he vibe coded some deterministic system. Maybe comparing the whole genome to a healthy one, and pointing out all diffs?