> this means that autoresearch will find the most optimal model for your platform in that time budget
I'm looking forward to finding out what model is optimal on my rtx3090
One thing I'm concerned with is that the model with best bpb after 5 minutes in smaller setups are only about ~10M Parameters in size which is too small for some emergent effects.
I’ve been working on Thymis, an open‑source device management platform, and just published a short demo showing how to take a fresh Raspberry Pi and boot it straight into a fullscreen YouTube kiosk in under 5 minutes.
The interesting bit (to me, at least) is zero‑touch provisioning:
You only flash one generic image per device type.
On first boot, the Pi connects back to a controller (cloud or self‑hosted).
It automatically pulls Wi‑Fi + kiosk config, provisions itself, and starts Chromium in fullscreen.
Optional VNC lets you see the screen remotely, even behind NAT.
The blog post is a simple demo (digital signage is a common request), but the broader aim is scaling fleet management for Pis and other NixOS‑based edge devices without manual setup.
Would be curious how others here have handled digital signage, kiosks, or multi‑device deployments — we tried to minimize the usual “burn SD cards + SSH into everything” pain.
In my opinion it's a tragedy there are so few resources in using "Propositions as Types"/"Curry–Howard correspondence"[0] in didactics in tandem with teaching functional programming to teach structured proof-writing.
Many students do not feel comfortable with proof-writing and cannot dispatch/discharge implications or quantifications correctly when writing proofs and I believe that a structured approach using the Curry-Howard correspondence could help.
Horrifying half-knowledge in the explanation of annealing section:
> Annealing starts with a weak piece of metal. In such a piece, the metal’s atoms are spread unevenly: Some atoms are close enough to share magnetic bonds, but others are too far apart to bond. The gaps that are left lead to microscopic deformities and cracks, weakening the metal.
But the result is great! Beautiful algorithmic design work!
The company I work for right now uses NixOS for all* Bare-Metal Hosts and VMs that run atop them. Personally I run my Home-Lab/Personal-Computing-Setup all on NixOS.
You have to consider that Nix is a language (turing complete, for describing build-processes) and treat it that way too. It has a similiar, if not harder, learning curve to other languages. Especially since most people are usually not exposed to concepts such as lazy evaluation, functional programming, etc.