Slightly tangential question: Does having a spruced up Mac Studio help learning the ropes of model tuning and/or development as good as having a CUDA device?
Raymond Chen's articles are true wealth of contemporary tech-lore. My favorite is the one where a user reported error on Flight Simulator map [0] and it escalated all the way up to Bill Gates & the then chief of FS division.
Raymond's work often uses analogies to everyday, real life situations to make the technical stuff more relatable. Sometimes, these end up being bizarre and even nightmarish. It's like a glimpse into some alternate world where the denizens are adversarial, hyper pedantic, scheming, criminally insane, or just plain stupid.
"Sometimes I’m in a hurry, and I want to make sure I am the next person to get served at the deli counter. To do this, I find whoever has the lowest number, knock them unconscious, and steal their ticket. But sometimes somebody else comes in who’s also in a hurry. That person knocks me unconscious and steals my ticket. My plan is to set my watch alarm to wake me up periodically, and each time it wakes me up, I find the person with the lowest number, knock them unconscious, and steal their ticket. Is there a better way?"
"This is like asking for a cheeseburger and then trying to peel every last bit of cheese off the patty to turn it into a plain hamburger."
"This is like driving up to a bridge, seeing a 'Bridge out of order' sign, then getting out of your car, moving the sign, driving onto the bridge, falling into the river, and then complaining to the car manufacturer that their car doesn’t work."
"This is like calling the natural gas utility company’s emergency number to report a major gas leak in your house. The gas company sends a technician over, and they can’t find any leak. They ask how you came to suspect that there’s a gas leak, and you tell them, 'Oh, I didn’t smell anything. I called you because I’m hoping to learn more about how to recognize the smell of gas. Do you have any tips?'"
"This is like 'I have this cupboard full of rubbish, so I will solve that by nailing the doors shut'"
"It’s like asking, 'What happens to the phrase 1 cup sugar when I eat my cookies?' The phrase 1 cup sugar was part of the instructions for making the cookies"
Oh nice, hadn't seen it. It looks good, and they've clearly got some decent stuff I don't (chrome extension, embedding in model cards, family pages).
The main difference appears to be how the graph gets made. From what I can tell they analyze the config on the server (their page says the first request can take a few minutes). I actually build the model on PyTorch's meta device and run a fake forward pass through it, so I get the real execution order and the tensor shapes at every module. That's what the animated replay and the compute/KV cache numbers are based on.
Pardon my ignorance about formats, but between XML & JSON, what would have made them choose XML? As I understand there is lot more tooling, standards & existing software examples around JSON for project & records management (e.g. VSCode's records as settings.json for example). Wasn't XML spearheaded by Microsoft but mostly used by Microsoft today?
json's only advantage is it's easier to write by hand, but that's not really a consideration here. XML has schema support and much more tooling available.
But even before it became officially unmaintained, it was already quite unloved, and it's one of the most used xml library.
We are not in the peak Java/J2EE era anymore. Pretty much nobody is going to willingly work on the foundations unless they're getting paid because the majority absolutely loathes XML and would not see maintaining XML libs as fun, and guess what, they're not? xslt also got removed from web browsers because nobody wants to maintain those libraries. And corporations like Google don't want to pay a dev for those either.
Meanwhile the tooling around JSON is as healthy as it's ever been, with very efficient implementations in most common programming languages and multiple SIMD impls in C, C++, Rust etc. People actually want to work on this.
> As I understand there is lot more tooling, standards & existing software examples around JSON for project & records management
I don't understand why that is relevant; as long as there is a minimum level of tooling, libraries and support for their choice, what benefit would JSON bring over XML? I don't see a clear reason for one over the other.
Honestly, I'd have the same question if they chose JSON and someone asked why did they choose JSON over XML - Why wouldn't they?
Calvin & Hobbes have always been such a joy. In childhood, it was a reflection of all the naughtiness you could come up with. As a middle aged adult today, I look at the bigger meanings of those simple adventures. Reading a few stories is a reminder that happiness could be found in simple things & vivid imaginations.
Bill Watterson's dedication to not commercialize it preserves the charm about 'simple life, simple joys' of our childhood. He could have raked in the money, but his integrity is admirable. It isn't easy to be in his position & make such difficult choices to preserve the ethos of his art.
Generally it is a good practice to keep v.[N] and [N-1] in active deployment. But lately a lot of policies from Anthropic are confusing at best, and at worst plain subscriber-hostile.
If costs are drowning them, they should rather be a bit more honest about the messaging and hike prices of their lower ($20 / $100) tier by ~$10 and enforce stricter limits (optionally).
I liked 4.5 and 4.6, but none of them are available outside of using API now - and thats a bummer given the 2~2.5X overall costs
It is perhaps better for Anthropic to do a price hike e.g. $25 or $30 for Pro with a clear/honest messaging e.g. "Running costs are high, price hike is unavoidable" than resorting to these tactics.
These shenanigans are earning them no respect. The market is already annoyed on model serving QA issues, and now (recently) Opus limits. They don't want to lose to OpenAI - understandable - but these shortcuts won't earn them anything either.
ACM has named Charles H. Bennett and Gilles Brassard as the recipients of the 2025 ACM A.M. Turing Award for their essential role in establishing the foundations of quantum information science and transforming secure communication and computing.
* An accessible news excerpt via CNN science [1]
Years before emails, internet banking, cloud servers and cryptocurrency wallets, two scientists devised a way to keep secrets perfectly safe and indecipherable to eavesdropping outsiders.
Their 1984 work depended on the hidden, counterintuitive world of quantum physics, which governs the way the world works at the smallest, subatomic scale, rather than complex but theoretically breakable mathematical codes to secure data.
The insights of Charles Bennett, an American physicist who is a fellow at IBM Research, and Gilles Brassard, a Canadian computer scientist and professor at the University of Montreal, have since transformed cryptography and computing. The pair received the A.M. Turing Award on Wednesday for their groundbreaking work on quantum key cryptography.
During my bachelors, I was using a Casio FX-991ES+. It was a beautiful calculator compared to the hand-me-down Fx-82 VPAM. And pretty popular with the folks of my age that time.
During and after grad school, I was using Mathematica or NumPy/Sympy quite a bit. But it felt like using an overpowered system to do basic assignments. Think of taking a Bugatti Veyron for grocery shopping. I indulged myself with another physical calculator - this time a Casio FX-CG50 with a color display and python support. I use it whenever I do self-paced courses or reading the occasional stats/ML paper.
I had the same gripe as the author. OhmyZSH seems too bloated for my needs. Added to that, the defaults adds (and oft unnecessary) emojis to prompts & outputs - something I don't find tasteful or appealing.
I stripped out most of the OhMyZsh functions (which is pretty modular given a shell package) and created a smaller, leaner package (leanZSH) having only the known stuff I may use. I have been using it without much complaints.
In a similar vein, there is slimzsh [1]. For me, it is a minimal config that enables everything that's useful, and only that. Quite a few others and I have been using this for more than a decade.
https://github.com/changs/slimzsh
reply