astrostl

MacBook Pro M5 128GB RAM

614 GB/s of memory bandwidth

MacMini M4 with 64GB of RAM

273 GB/s of memory bandwidth (also only currently available with 48GB)

When it comes to inference speed, you want your model to fit in memory, and then to have as much memory bandwidth as possible. In this case a hypothetical Mini with 1TB of memory would still be over 2x slower with 27-35B models.

And FWIW I have an M4 Max MBP 128GB that I keep on a Roost laptop stand, with a separate keyboard/mouse/video. It does fire up the cooling jets when running local LLMs, but stays within tolerance for me on noise. I haven't heat-tested it on longer runs, but I imagine the risen airflow helps a ton.

Updated https://github.com/astrostl/surplies to scan for it too

FWIW I vibe coded https://github.com/astrostl/surplies to detect evidence of the Axios and LiteLLM malware, using StepSecurity's writeups as a data source.

FWIW I vibe coded https://github.com/astrostl/surplies to detect evidence of the Axios and LiteLLM malware, using StepSecurity's writeups as a data source.

Disagree. We all are — or should be — Linux kernel developers. What's more, we should align to a specific and singular VCS worldview informed by BitKeeper, which no longer exists, whether or not we used it. Therefore Git. Thank you for your attention to this matter!

Huge enabler for the mini keyboards for me: Fn + L/R for Home/End, Fn + U/D for PgUp/PgDn.

I have mixed feelings about it, but I'm going to fire somebody tomorrow for using a struct just to prove a point to the author.

Browser automation: chrome-devtools-mcp, playwright-mcp, etc.

The quality of the ChatGPT Mac app is a major driver for me to keep a subscription. Hotkeys work, app feels slick and native. The Claude Mac app I found so poor that I'd never reach for it, and ended up uninstalling it — despite using the heck out of Claude Code on a Max plan — because it started blocking system restarts for updates.

Yep. Consumer Reports' "Find a Car" page has sorting options for Overall Score, Road Test Score, Predicted Reliability, and Predicted Owner Satisfaction. I think they're a tremendous pro-consumer non-profit, and that a $39/year membership more than pays for itself by way of better major purchases.

Rivian, by the way, is the lowest-ranked of 26 covered auto manufacturers in terms of predicted reliability, below Ram and Jeep. The top 3 are Toyota, Subaru, and Lexus.

That doesn't seem like enough to entirely shape worldwide discourse around nutrition and sugar.

IDK, see the "BLOTS ON A FIELD?" by Science ("A neuroscience image sleuth finds signs of fabrication in scores of Alzheimer’s articles, threatening a reigning theory of the disease") or "The 60-Year-Old Scientific Screwup That Helped COVID Kill" by Wired (regarding the anti-scientific refusal to acknowledge it as airborne) for a couple of recent examples. Once underlying assumptions stop getting questioned, I think anything is at least possible.

Same. CMD-F, 'sqlite', no hits, skip and go straight to comments.

It's such an amazing business model:

- create a platform to host content others create

- get employees to ask for company-provided access

- almost none of these employees really use it

- collect subscription revenue indefinitely

This is why I use Go as much as reasonably possible with vibe coding: types, plus great quality-checking ecosystem, plus adequate training data, plus great distribution story. Even when something has stuff like JS and Python SDKs, I tend to skip them and go straight to the API with Go.

The information to properly land the plane is in the manual

Look, I already liked the nerdy blog post! I don't need even more reasons to like it.

Most recently (yesterday), vibe coding a better interface for Roblox screen time: https://github.com/astrostl/blockblox . Claude Code crushes, and I'm preferring Go for everything I can to take advantage of typing, quality ecosystem, and distribution. Still need to implement the QE side on this as I have on other things.

The OP confirmed that it isn't:

"is this using RDMA?" "No. It will be faster with that in the next release" [1]

1: https://x.com/awnihannun/status/2000243131779023329

Having been through cycles of manual writing with '#' and having it do it itself, it seems to have been a push on efficacy while spending less effort and getting less frustrated. Hard to quantify except to say that I've had great results with it. I appreciate the spirit of OP's, "CLAUDE.md is the highest leverage point of the harness, so avoid auto-generating it" but you can always ask Claude to tighten it up itself too.

I have Claude itself write CLAUDE.md. Once it is informed of its context (e.g., "README.md is for users, CLAUDE.md is for you") you can say things like, "update readme and claudemd" and it will do it. I find this especially useful for prompts like, "update claudemd to make absolutely certain that you check the API docs every single time before making assumptions about its behavior" — I don't need to know what magick spell will make that happen, just that it does happen.

Instead of software thermal control, I could add an actively cooled heatsink with PWM fan control. This might achieve similar temperature stability while using less power overall.

Love this (honestly). Interesting article!

I haven't seen convincing evidence that vitamin D supplementation is materially useful for anything but rickets. I get the impression that naturally high serum levels are an effect rather than a cause of other positive things, and that supplementation mostly increases serum levels without effecting positive things. It doesn't seem harmful either, so can't hurt might help?

In Germany, there’s a popular nonprofit which tests consumer goods for safety and publishes the results.

In the USA too: https://www.consumerreports.org/about-us/what-we-do/

To this day the single most-impressive tech demo experience I've ever had.

With aggressive Claude Code use I didn't find Sonnet better than Opus but I did find it faster while consuming far fewer tokens. Once I switched to the $100 Max plan and configured CC to exclusively use Sonnet I haven't run into a plan token limit even once. When I saw this announcement my first thing was to CMD-F and see when Sonnet 4.1 was coming out, because I don't really care about Opus outside of interactive deep research usage.

I got so dang sick of SD and enclosure fan failures on rPis. Adore my $159-in-2024 Beelink S12 (N100/16GB/512GB).

Time Machine can do that. Whether they're automatic or manual backups, you can boot your system in recovery mode and restore from a backup to fully revert to a previous snapshot. It will require a reboot and the speed of execution will depend on the size of the changeset from the current state.

Since Catalina the OS itself is in a read-only volume and everything else is separate - by default, respectively "Macintosh HD" and "Data" as exposed in Disk Utility. Strictly speaking, I think programs cannot change the OS or boot sector. The only thing I know of that can is a macOS upgrade itself, and it does automatically take a pre-upgrade snapshot for the ability to restore.

GoBack was designed by Wild File, Inc., a company located in Plymouth, Minnesota. The software was shown at COMDEX in November 1998 and released in December 1998.

GHOST (an acronym for general hardware-oriented system transfer), now called Symantec™ GHOST Solution Suite (GSS) for enterprise, is a disk cloning and backup tool originally developed by Murray Haszard in 1995 for Binary Research.

^^ GHOST was not descended from GoBack

As for something like GoBack on Mac, if you're using a recent macOS and have an APFS filesystem you can take/restore whole-disk snapshots with Time Machine, tmutil from the CLI, or a third-party tool like Carbon Copy Cloner.

Another practice I’m not hugely fond of is “leveling”

That said, I do believe there are three meaningful levels in management

Cool!

I went decades thinking that all prescriptions were imperfect until I started getting age-related presbyopia (reduced close focus abilities) and sought perfect correction or a clear reason that it couldn't happen.

The diagnosis: keratoconus (thinning and ultimately misshapen cornea). The prescription: highly-corrected rigid scleral contact lenses. The result: radical. I would describe my uncorrected vision as 480p, glasses as 1080p, and scleral contacts as 4k.

The scleral prescription makes my medium and far vision great while further limiting my ability to focus closely. I experimented with "monovision" to have one near-focused eye and one far-focused eye with my brain sorting out the difference, but I could tell that it resulted in lower resolution. I optimize by sticking with the medium-far prescription and wearing 1.5x reading glasses for anything within a few feet. It works great.

If you're in a similar boat, I can't recommend it enough to get checked for keratoconus. To get that you'll want to find a place that can do a corneal tomographic scan with a device like a Pentacam. Oh, and try to not rub your eyes and tell your kids the same because it's believed to be a major cause. I used to do it for fun as a kid to make colors while I was bored.