"meta" and "personal". one of you is lying.

systemd is all you need

And someone else's machine

systemd and ansible

You're absolutely right

any particular reason not using terra or sol, given the generous limit quotas and frequent codex resets?

I run out of my 20x account with Sol medium and some Sol high in about 2 days

Because Luna feels right and just enough for me.

might as well rethink about their pricing on gpt-realtime. for a quick test, spoke to that model for ~5 minutes, ended up spending near $1.

They haven’t launched the new realtime in api yet

the moment debate becomes philosophical it's over

Bonsai vs Qwen (quick) Benchmark: https://github.com/ArmanJR/PrismML-Bonsai-vs-Qwen3.5-Benchma...

Bonsai is qwen3.6 based, not 3.5

Likely apples / oranges

Bonsai 8B and 1.7B were on Qwen3.5 the benchmark is from a few months ago. However I'll add Qwen3.6 to the benchmark too.

qwen3.6 starts at 27B

can someone tldr me why choose apple container (and its ui) over docker (and orbstack)

I can’t speak about orbstack, but I’ve worked with docker desktop and podman desktop for years on macOS. Those programs start up a virtual machine that consistently eats ram regardless of whether or not you are running containers in it. Apple container looks lighter weight. In the age of ridiculous ram costs, you gotta save resources.

In addition to memory saver that another person replied about, Docker Desktop also has an MCP server functionality and marketplace (almost all free) and huge AI focus. You can hardly compare it to the others at this point.

I was doing the following at the same time on my MBP this week:

* running a bunch of containers + MCP servers for Claude and Codex on Docker Desktop

* heavily using Claude Code with Fable and packer to build cloud marketplace images

* having Codex write some tests and git flows and reviewing the work in vscode

* automating a character in a Wine-based 1st party RPG in the background running at full resolution

* watching anime on Plex in between Claude Code prompts

It's all about your machine. Docker Desktop is not my worry and if you're a Dev you should have a nice laptop with 32-64GB or more, Apple Silicon Max CPU, etc. This goes for Fusion or UTM also if you want to run a Linux Desktop.

I use docker CE with all container/tui interfaces on all of my Linux systems, but Docker Desktop is nice for macOS or Windows. I almost forgot about Docker Desktop's Gordon, and the AI assistant will do things like analyze your Dockerfile or compose.yml. Super handy.

> if you're a Dev you should have a nice laptop with 32-64GB or more, Apple Silicon Max CPU, etc.

Really depends on what you're building, to be honest.

Docker Desktop's memory saver shuts down VM when containers are not running.

Additionally, Docker/Podman/Orbstack start a single VM, where memory is shared between containers.

On the other hand, Apple Containers create a separate VM for each container, which results in higher memory usage due to Linux kernel overhead, as well as the fact that kernel will try to use most of the available memory for file caching.

one main marketing leverage of 23andMe, AncestryDNA, etc are fulfilling the curiosity of people who want to know which part of the world their genes are from. I guess that dataset should be preparatory.

Problem with those providers - they only check 700K positions out of 3 billion and there is no mapping quality or allelic depth data in those dataset and this is critical for assessing whether the detected variant is a false positive or real.

It's not suitable for health investigations since most of DNA is not sequenced and genotyping technology is known to produce high rate of false positive for rare mutations.

(I'm the solo-founder of Gene Inspector Pro, mentioned in the blog post). AMA. :)

It's a bit ironic, since FTDNA and MyHeritage (which uses FTDNA's lab) have switched to NGS now, so they presumably could deliver those notorious "health insights", at least better than 23andMe. But they aren't in that market, and 23andMe shows no inclination to switch. They're probably licking their wounds after the user hack and buyout fiasco.

Need to check if they do 30x read depth or much less - ancestry doesn't need 30x, so companies may just do 2-3x reading, which is not enough for anything health-related due to high chance of errors.

MyHeritage uses 2x. FTDNA uses a custom targeted enrichment panel, so high accuracy in selected regions important for genealogy.

funny how we may have to wait even longer for llms to pick up this update in their pre-training

It's actually fantastic for job interviews. Use that trick with my clients when asked to review candidates.

Is there seriously no solution to this? Perhaps something we fan do post training? For example add the new features to SKILLS.md? But the trade-off here is of course tokens.

My working methodology has been , LLM output is a jumping point and to use my experience, knowledge and basic understanding to N+1 it.

So for bleeding edge stuff it works out well or in places where documentation is not great like Apache Flink.