This project is very cool! One readme nit: "Pack a stateful virtual machine into a single file (.smolmachine) to rehydrate on any supported platform." For awhile I thought this meant that you could rehydrate a machine's memory like you can with a firecracker vm, but as far as I can tell you can't? It's stateful == disk?
It’s impossible to disengage the deployment of technology with the way the technology was deployed. I’m not saying anything that I’m not saying. I believe that there could have been a world where deploying technology in all these classrooms had a positive effect
The increased review burden is also happening inside companies. It’s genuinely hard to keep up with the volume. I was a little surprised to see my comment downvoted. I’m not saying we shouldn’t be able to delete slop PRs. Of course we should! I’m saying that the pr should change at least _somewhat_ in response to how much programming is changing.
Also worth stating that I have been ranting about contract first reviews for 10 years and it’s not just in response to llm written code.
They need to talk about how the pr itself should change. The text diff just is not the right thing to center. We should be using ai to chunk changes into reviewable bytes and to align on semantics and contracts.
Not the same but this gives me an idea… what if there was a map reduce for doms as a web primitive. Like imagine if I could make a dom (or feed) that was some selection and transformation of another dom
I feel like itd be cool to try prompts based on an adversarial justice system… attorney agents arguing both sides, a judge ruling on “the law”—adherence to instructions etc
I’m surprised the hospital said they were confident it was safe. I wonder what gave them confidence? I’m struggling to think of what data I could have on hand that would convince me it was really safe. Also, to folks saying that randomness is clumpy… did you read the article? I think a bunch of nurses that notice they are all getting sick (and then FIVE of them getting brain tumors) should be taken quite seriously. I’d start with the assumption that there IS an environmental problem and then figure out what it is.
Idk I used Claude Code recently and revised all my estimates. Even if the models stop getting better today I think every product has years of runway before they incorporate these things effectively.
RAG + analysis on health data has huge potential. We need to tread carefully, obviously, but I have also used these models + RAG on personal health data at times when the docs were scratching my head. It was very helpful, although I approached it a bit differently. In my case it was my infant in the hospital for breathing / eating issues. I took data from the network tab in mychart and added context additional context, then asked Claude questions. My main goal was to be prepared for rounds in the morning. I wanted to understand what types of information each test was going to give me.
It was pretty good! Unsurprisingly, it did hallucinate sometimes or miss some nuance. But, I knew I was talking to something that did that. It was helpful for me, even with errors. I didn't, however try to just say "hey given all this what's going on with my baby"
On another topic:
If I were you, I'd make this thing speak FHIR or some other common format. That'd make using this thing in an ecosystem of other tools a lot easier. There are even FHIR graphql apis you could just polp on top maybe.