raffraffraff

So it's not about the model (eg TensorFold/Qwen3.8-27B-MLX-4bit) it's about the inference engine?

Can you explain the pricing here?

Of electric sheep?

Here are the numbers that got me buying a used M1 Max for LLM work.

    M1 Max 400GB/s
    M4 Pro 273GB/s
    M5 Pro 307GB/s

Only Max are worthwhile for LLMs. An Ultra is just two Mac chips on a die, but I think they're just used in Mac Studio.

I'll check it that inference engine! Thanks!

Edit: it doesn't give numbers for unsloth/Qwen-3.6-35B-A3B 4bit, but I'm using mlx-serve right now and getting petty good numbers (and it took a while to tweak everything properly!)

M1 Max 64gb here, running Qwen 3.6 35b (unsloth 4bit) and it's plenty usable, and cheap used. Best bang for buck.

Regular < Pro < Max < Ultra

Fantastic key ceremony. Great time had by all. Can't remember a damn thing, we were all so plastered.

I'm having kinda the opposite experience, but it's less to do with Linux and more to do with hardware, firmware and manufacturers.

MiniPC network device doesn't have a Linux driver. At all. Had to get a vibe-coded, reverse engineered one. This was a first for me.

NAS (ugreen) was a pain in the ass to install my own OS on (I didn't trust theirs, wanted OpenSUSE, and had to get a reverse engineered LED controller driver so I could display hard drive health)

ASUS Laptop will not sleep properly, and would drain the battery flat in a day or do. Myself & Claude gave up after days of trying various tricks workarounds (up to BIOS hacks). The best I can do is hibernate (after about 5 minutes of sleep) but man, the hacky bullshit required to get there was incredible.

Different laptop sleeps fine but will not, under any circumstances, hibernate. Even though I have all the crazy tricks up my sleeve from the ASUS that won't sleep properly and needs to hibernate. Different hardware, totally different bullshit problems.

And then I recently got an Macbook M1 Max for local LLM hacking and everything just fucking works. I don't like MacOS but the hardware is amazing and the OS just fucking works with it (and I like that).

This is the biggest issue I have with headlines "[country] got all of its electricity from [some renewable] today!" when the reality is they rely on some other supply when [some renewable] has zero output.

Like you say, a lot of countries get a pass because a neighbour invested billions and decades building nuclear. Or else they fall back to coal/gas.

My mother has a crow (it's lame and very slightly tame). It's the only crow in a huge murder that isn't afraid of her, and the other crows are in absolute awe of it. It's like they don't know how to completely trust a human like it does. The special crow is like a human whisperer.

The fact is, she would feed any of them, but only crow taps on the window, and goes from one window to the next as my mother moves through the kitchen. The others only get the leftovers, but they line up on the fence.

I hate that ikea switched like that.

What's your experience with Matter / Thread? Is it a case of buying a different USB dongle and devices?

With AWS SES for sending mail, this Cloudflare + Gmail setup would be ~free for a <1000 outgoing emails per month.

I've done too many jobs around my house that should have been done by a qualified <whatever> because none of them are interested in a shitty little repair job, or if they are, the quote is astronomical because there's a huge cost overhead of taking a job and getting your shit onsite.

Does remote debugging port itself get detected, or does the presence of a webdriver connection to the port get detected? I believe it's the latter.

I've had success launching the browser and using dumb dumb methods to get around the captcha before attaching playwright.

Dumb dumb methods = wmctrl, xdotool, bash (work fine for Cloudflare's "are you human" check)

I will! Actually I started a personal project for this stuff, and have a dirt simple local app with about 50k tracks (YouTube embed, track credits, genre, rating).

I was going to add Spotify metadata from that recent data dump, but it's kinda trash data (valence isn't trustworthy, wildly wrong a lot of the time).

I then stumbled on lrclib data dump, local LLM and a super two step prompt that extracts extremely high quality lyric based metadata (about the narrator, protagonist, setting, story arc, moods etc)

Anyway, even though I bitched about MusicBrainz in another comment, their JSON dumps can be coaxed into giving the sort of metadata that I wanted to get from RYM, who absolutely suck due to their crazy aggressive blocking, self sabotage and unforgivable data poisoning if you trigger their bot tripwire. Their data came from users, but as a paying user I can't use their site to tag my music. No API, and you get banned if you try to write a music player plugin to fix the tags in your "now playing" track. I don't know what their business plan is.

If you're interested trainspotting individual CD serial numbers, then there's nothing quite like it. But my gut feeling is that there's a lot of folks with digital music collections who don't give a hoot about that stuff, and are interested in metadata about the song they just heard (the who, the when, the instruments / genres / moods, lyrics). For this audience, there are few decent, clean metadata sources. But when you ask, people will point you to MusicBrainz or tools that use it as a source (like Picard or Beets). You have to fight tooth and nail to avoid tagging a 70s album year as "2005" because there happened to be a remaster that year. Who. Fucking. Cares what year a Remaster happened when it's a 1976 album that you don't want showing up in your "Best of 2005" smart playlists.

As for social features, nobody is ever going to use MusicBrainz. They might have used RYM or Last.FM before the site owners made them unusable or pointless.

I used to love Last.FM and was a paid subscriber. Their neighbourhood radio was a great way to find music

rateyourmusic.com is also dead to me now. And MusicBrainz is an excellent resource marred by horrible user experience and a primary focus on listing every issue of an album in every country and in every format (who the hell wants that?)

I did read it. I didn't enjoy reading it. Yes I said the same thing, in fewer words, and I didn't hint that maybe there really was a conspiracy. There doesn't need to be a conspiracy or a blacklist if you rub people up the wrong way, and then the bottom falls out of the market you're writing about.

We used to have parties. But then my wife had a few realisations:

1. "I hate losing sleep, and there are very few things in the world now important to me than a good night's sleep".

This has been a life changer for both of us. I used to be a late owl, but when she started going to bed at 9:30,I did too. We now have competitive "what was your sleep score" in the mornings over breakfast

2. "I do not enjoy conversations with more than 3 other people at the same time"

I don't mind crowds and can "hold forth" with entertaining stories and have fun doing it, but my wife has social anxiety that really only emerged in larger groups. She doesn't even like going to parties or bars, and after a few drinks, leaves me to it.

3. "I don't like getting wasted, and I don't even like getting slightly drunk if it screws up my sleep"

For me it's more medical. While I'm perfectly healthy, I noticed around the age of 45 that red wine gives me palpitations and tachycardia. Lots of beer (especially really nice beer) does the same. My father is 99 years old because he never drank alcohol. I honestly believe a single hangover would actually kill him. But certainly, years of "not kicking the shit out of his insides with booze" have contributed to his long life.

Of course it is. That's where they make their money. In other news, the local drug dealer is targeting drug addicts and the bar still serves the alcoholic who has destroyed his family and is drinking himself to death.

God I hate this writing style. And the title. The reality is probably more bland.

In the 80s and 90s, tech was magic. Lots of computer magazines, cyberpunk novels. Tech fact and fiction sold in huge quantities. Mainstream tech publications depended on tech ad dollars, so editors preferred tech sycophants. There were doom-scarers but it was mostly a utopian thing. It's what got me into tech as a kid.

After the dot-com bubble burst, this imploded. Investment in tech dropped across the board, including advertising tech publications. I'd love to get a count of cyber/tech fiction novels published through the 80s,90s and early dot-com burst.

We have a politician putting up very obviously hand drawn (because we need to be fucking certain that AI didn't do it right?) posters for an anti-datacenter meeting. I'm pretty sure that if you were to talk to the actual humans in our society they'd talk about housing, employment, public transport. Not datacenters. If datacenters even feature in their real world outcomes it's 3rd hand, but the usual lefty crowd will turn up to any vaguely lefty protests with their bizarre mix of rainbow flags and keffiyehs.

TIL they renamed Perl 6!

Based on my local park, people like you are the exception. Recently I've been picking up the rechargeable ones with digital displays to tell you how full the battery is. I have dozens of them. I stopped collecting them because I have enough to build lots of little projects. It is so fucking wasteful.

I love this use of the word "agreement"

Slippery slope. Next, agents are convinced that they contain the word of God and the filthy heathen at the keyboard should convert or die.

I strongly agree with this:

> the ideal goal of an interview is to make a prediction of how happy and productive the candidate will be in the role they are being considered for.

I'd add two things that should be an obviously component of that way of thinking, but frequently get missed:

1. You need that good candidate as much as they need you. Anecdote: I have found it much harder to hire good people, than to get hired. By "good" I don't just mean smart, I also mean amicable and driven.

2. You do not have to agree with them on everything during the interview. Anecdote: some of my best work has been against the grain. But I am persuasive and passionate when I'm convinced that I see a "better path forward", and if I can convince my peers and leaders to let me shoot my shot, it usually works out. Part of the success may be due to the fact that I have to succeed in order not to look a fool or let people down. (If someone else's venture fails its much easier to shrug).

The point being: your candidate may disagree with you. But may be also be right. Or wrong, but letting them invest their whole self into the enterprise may be better value.

But this resonates the most with me:

> because he was such a jerk to people that those people would spend tons of their time and energy to find ways to avoid working with him

I've seen this so, so much. And it wrecks teams. When I say "let people disagree", again, this is when the best and worst of people emerges.

A human brain has qualities that make it very different from AI, but the same as other living creatures: it's alive, it can feel (sensations, pains, pleasures, emotions) that give it rich experiences and memories, and it hardwired to continue existing and protecting the body it is inside. So: motivation, desires, fears, and continued consciousness and thought without any external person having to poke or prod it into "doing".

None of the models I've downloaded have sprung to life spontaneously, or done anything remotely unpredictable. None of them can act without input. They all have strictly limited ways to receive input and pass output. They aren't alive, can't feel, don't have experiences. The fact that they are a phenomenon that emerge from billions of snippets of our language lets them express languages in extremely convincing ways, and they are extremely impressive in what they can do. But this isn't Short Circuit or D.A.R.Y.L. or the Terminator, none of them have feelings. None of them continue to think after the program exits.

Without those characteristics they can only generate based on what they've ingested. The capabilities are still astonishing, but they're still dead things, and nobody will ever risk their life to protect the model "context" from getting "killed"

Have you tried turning it off and then on again?

> We're no longer pursuing restarts as a path to remediation.

Oh you have

In America?