

Yeah, totally agreed.
It’s a tough balancing act to filter the snake oil while taking viable moonshots, but I think that’s a niche that the US regulatory regime has generally done well on, all things considered.


Yeah, totally agreed.
It’s a tough balancing act to filter the snake oil while taking viable moonshots, but I think that’s a niche that the US regulatory regime has generally done well on, all things considered.


Based on the linked paper, it reads like they took an existing open weight LLM, tacked on a scheme to parse and output genomics data, and finetuned it a bit. It’s the same approach chatgpt and other big-name models use for reading and producing images, but applied to a different data format.


I don’t disagree with the general statement… but this was specifically done with an LLM, so it seems like a moot point?


They’re skipping that whole “human trial” thing by going for a compassionate use exemption
https://www.fda.gov/news-events/public-health-focus/expanded-access


With ya all the way on that front 🤙


There’s absolutely nothing whatsoever forcing these companies to use GNIS names, even for users within the US; they’re doing so purely out of their own discretion.
Outside of federal contract work, you can use whatever names you want in a mapping product.
BGN decisions are not required to be adopted by nonfederal domestic entities, such as private company applications (e.g., Google Maps and Apple Maps); however, some nonfederal entities have chosen to do so or generally have a policy to defer to GNIS nomenclature.
They hate us because of our freedom scp-to-prod pipeline

Agreed. You can even take it a layer deeper and argue the physical waveforms reflect the same trend:

Bouba is a cleaner, smoother pattern with a few dominant frequencies while kiki is a chaotic signal all over the spectrographic map.


This whole account is obvious “agentic” spam and y’all just upvoting without reading a damn thing.
Click the link if you don’t believe me.


Un día de ventana abierta
Mi vecino vino a verme
Estaba llena de desilusión
Me miró en los ojos, y me dijó
Ya dijimos no, pero el sí estaba en todo
Todo lo que hay
On a day of open windows
My neighbor came to see me, filled with disillusionment
He looked me in the eyes and told me:
We already said no
Yet the “yes” lingers on
In everything that is
From Chilean-American artist Nicolas Jaar
(The Yes/No refers to the referendum vote that nominally ended the Pinochet dictatorship.)


Np!


I work in this space. I agree with the gist of what you’re getting at; the catch is that it’s mostly just feeding the beast rather than breaking the dynamic.
With LLMs, the problem is that they’re neither objective nor consistent, so the insurance companies are just going to increase their headcounts and hire more people to contest submitted claims. Deterministic output would change the picture in a big way, but I’m not seeing much appetite for that in the market for some reason.
There’s a lot more potential for disruption in the form of advanced monitoring that can automatically flag and document complications in an objective way (which is my professional wheelhouse, for the sake of disclosure), but even then I expect payors to mostly pass the increased reimbursement rates on to their enrollees.
A single-payor model is the biggest real cure for what ails our system, and I think I speak for a lot of us in the industry when I say that we’d happily lose our jobs to see that happen for the greater good.


I’m not a C guy, but afaik the closest thing to a useful manual in print is the K&R book (written by Kernighan & Ritchie, authors of the language).
I have a copy on my shelf and it’s… alright. It’s much more focused on basic language constructs than pragmatic usage and opinionated style. Good for what it is, but you should pick up something more recent and opinionated to go alongside it IMO, based on whatever C dialect you want to focus on (C89/C23).


Yeah, O’Reilly books or any other printed stuff that strikes your fancy are great options. DDIA and Learning the Bash Shell are personal faves.


She died three years later, and I can’t help but think that was also related. The WH press pool was her life.
Shocking that a 92 year old woman would just up and die like that.
You wanna keep Mitch McConnell around too? He’s been in the senate so long; it would be a shame to make him quit.
Absolutely based. Dan Gackle is an insincere scumbag of the highest order.


now that AI does everything 20000 times better than I ever will
Now I’m no cognitive behavioral therapist, but, uh… this is what I think is catching you.
I use AI plenty. I write most of my code with it. The garbage it spews out if you don’t hand-edit or get ultra-specific with design instructions is just atrocious, even if it looks sleek in the gestalt. You can give it small tasks and one-shots that it handles perfectly, sure, but it will rot a large codebase to the core if you start letting it make any real structural decisions.
There’s simply no replacing architecture and focused thought, and writing code by hand isn’t going anywhere, even if the majority of the industry works with AI. I mentor junior engineers, I’ve seen the worst spaghetti they can make, and I can nonetheless confidently assure you that you’re better than “frontier” AI across multiple domains.


For the common-case risk model in the post body (SaaS user), the easy option is to set up a secondary git remote over SSH on a cheap cloud VPS. Then you can simply run git push backup-server or the like. Bonus points for setting up a cronjob to push to your backup box periodically, or to take periodic tarball snapshots of the repos on the server.
Personally though: I self-host forgejo with registration disabled, and take automated snapshots with ZFS occasionally in case I truly bork things.
Given simple primitives and runtime permissions, a skilled hacker can even bootstrap one of these “shells” from scratch, underlining the continued importance of banning math and computer science education in our schools.