Note that Garnix is more than a CI system. It also allows for hosting apps with a quite brilliant and AFAIK unique solution to interdependencies. https://garnix.io/blog/call-by-hash/
I can relate to that. If I'm not careful catching up my internal model and understand of the problem space and the specification of my project while the agent keeps spitting out code I feel very uncomfortable. Perhaps I have very high standards for my code quality for various reasons? But its extremely frustrating and tiring to me to check all the implicit assumptions the agent so boldly makes at high speed.
Usually, I start with a solid idea that needs hands on experience to fully developin, so I need to stay in the loop. Than I have to force myself to carefully review the code and point out architectural issues (usually there sneak in a lot of them). If I don't force myself, I end up in a limbo where the agent produces code faster than my mental model following.
Even if I have a well developed domain knowledge and a good idea what I want, I then definitely feel some kind of, what I guess compares to, slot machine effect. "Maybe now I get what I want. Argh, not quite. Okay, feels nearly there. Perhaps only this prompt. Ah, there is something missing...".
I'm really curious how you handle and approach this.
I think the pelican on a bike is known widely enough that of seizes to be useful as a benchmark. There is even a pelican briefly appearing in the promo video of GPT-5, if I'm not mistaken https://openai.com/gpt-5/. So the companies are apparently aware of it.
Sadly, only the mounting rim is normed. The form of the body is not. Meaning, containers from different manufacturers don't stack necessarily and lids are not interchangeable.
NixOS makes this rather easy too. There exist modules for synapse, livekit, etc (MAS module is still a Pull Request, though) and the setup is quite doable: https://wiki.nixos.org/wiki/Matrix
However, you still need to know what you are doing (the manual helps) and connect the pieces (in theory there could be a nixpkgs module that does this for you but apparently nobody did bother). Once its done you can lean back.
I've been running my homeserver happily for > 5 years and it was fairly straightforward.
That is only partly correct. Devenv.sh misses outputs defining packages compared to flake.nix, i.e. there is no way to deterministically build the project as a nix package. It's the developer environment devenv competes with. Both can even work in tandem. Devenv for the devenv, flakes for the rest.
The compiler is FOSS under a permissive license Apache 2.0). Only the online editor, similar to Overleaf, is not Open Source. Please check the facts before hitting reply.
Scientific/Numeric/Data-Python is essentially a DSL around C-API which creates friction (try for example to map a custom function over a Pandas column). Whereas in Julia, it's just Julia. It's liberating to just extend and use a library written in the same language. It leads to surprising synergies.
> First, we evaluate, for each voxel, subject and narrative independently, whether the fMRI responses can be predicted from a linear combination of GPT-2’s activations (Fig. 1A). We summarize the precision of this mapping with a brain score M: i.e. the correlation between the true fMRI responses and the fMRI responses linearly predicted, with cross-validation, from GPT-2’s responses to the same narratives (cf. Methods).
Was this cross checked against arbitrary Inputs to GPT-2? I gather, with 1.5 Billion parameters, you can find a representative linear combination for everything.
Live View (a WYSIWYM Editor) - you type markdown, but see it rendered, comes to mind. Also some interesting extensions/add-ons centered around knowledge management are arising.
Unfortunately, Obsidian itself is not open-source though; but worth checking out.