Autocomplete on iOS is a shit show. It regularly autocompletes non-grammatical text, which is unforgivable in 2025 when we have AI that can write coherent sonnets and code. Dictation is still at the same level it was at 10 years ago - complete shit. Carplay sometimes randomly starts playing music when I get in the car, other times it doesn't. My Apple Watch regularly can't find my heart rate, for long time periods. The HomePod app and Watch display incorrect information about what's playing on HomePods about 50% of the time. There's no way to filter text spam. The Messages app on MacOS doesn't let you filter by known senders. If you delete a text thread on iOS, it doesn't delete it on MacOS, so my desktop messages are cluttered with fucking donation requests from PACs. Try to do anything with Siri, even simple things like playing a song. It still makes bizarre mistakes. It can't answer basic questions about my calendar.
The list is endless, really. Everything looks "delightful" as fuck. Mac and iOS fonts, colors and text padding are immaculate, so it gives the impression of solidity and competence that isn't really there. A lot of things "mostly work" but aren't reliable, so I can't rely on them. They can list them as "features" but if I can't rely on them, I can't use them, because I don't want to deal with constant frustration. They act like all their systems are this one integrated whole that works well together, but it doesn't.
I don't think that everything will "just work" on Linux, but at least I won't be paying a premium for the privilege of having my needs as a disabled person ignored. I'll be able to customize my experience to meet my basic accessibility needs without fighting against a company that seems to hate me.
Siri has always been a shitshow for sure. I basically use it for timers and even that it fucks up a non-negligible number of times. For the few experiences I've had with it, at least Google assistant is a lot more reliable.
I also noticed worsened reliability in the heart rate tracking of my Apple Watch in recent workouts. It must have happened in one of the recent updates because it was fine previously. I could say it's programmed obsolescence but I'm sure I would be accused of conspiracy theory. But it is hard to interpret the failing reliability otherwise when it suspiciously happens after updates and around new hardware release.
In any case, I don't think the Apple Watch is a very good product for the price, so whatever, the next watch will be focused on sports and the competition has made great alternatives.
I completely share the sentiment that everything looks good but doesn't work that well in practice. There are so many random issues that make the hardware prices very unpalatable.
Ah well, everything changes, not always for the better. The pain is in transitioning to something else, but that's something that is very true for most tech related things since we can't ever agree on proper standardisation.
I believe SchemeFlow [0] is working on solving some of these problem, particularly with the insane reporting requirements. But of course, that still leaves the unions...
The unions will use their government connections to force the job to exist even if it doesn't need to be. They'll figure out a way to get a union driver sitting in an fully autonomous truck for some invented safety checkbox
I remember Louis CK said he had a hell of a time trying to run his own comedy shows so he could offer lower ticket prices for his fans, but because every bit of the theaters were unionized it got really expensive fast and failed. They couldn't even touch the curtains, they had to pay a union guy to stand around all day and his only job was pulling a curtain cord at the right time. Which was some NYC rule.
> The unions will use their government connections to force the job to exist even if it doesn't need to be.
And that's OK. At the end of the day, labor unions exist to help people and not robots.
People were worried robots were going to take their jobs. Then people were seeing their jobs get shipped overseas. Today, robots are finally taking over what jobs are left.
> At one of our dinners, Milton recalled traveling to an Asian country in the 1960s and visiting a worksite where a new canal was being built. He was shocked to see that, instead of modern tractors and earth movers, the workers had shovels. He asked why there were so few machines. The government bureaucrat explained: “You don’t understand. This is a jobs program.” To which Milton replied: “Oh, I thought you were trying to build a canal. If it’s jobs you want, then you should give these workers spoons, not shovels.”
My FIL spent 30 something years as a union carpenter and for about 15 of that was teaching underprivileged kids carpentry and finding them job placements while rebuilding national parks.
That union did just fine for America, it's better off for it having existed.
The 5 days work week, the paid holidays (in the country where they exist), and rights on the place of work wouldn't have been obtained without all the struggles of the working class in the past century, by striking and unionizing. They literally made the world better for everyone (except the people owning companies)
For argument's sake, suppose we live in a world where many high-quality models can be run on-device. Is there any concern from companies/model developers about exposing their proprietary weights to the end user? It's generally not difficult to intercept traffic (weights) sent to and app, or just reverse the app itself.
So far, our focus is on supporting models with fully open-sourced weights. Providers who are sensitive about their weights typically lock those weights up in their cloud and don't run their models locally on consumer devices anyway.
I believe there are some frameworks pioneering model encryption, but i think we're a few steps away from wide adoption.
Simple answer is they won't send the model to the end user if they don't want it used outside their app.
This isn't really anything novel to LLMs of AI models. Part of the reason for many previously desktop applications being cloud or requiring cloud access is keeping their sensitive IP off the end users' device.
To echo the other replies, the tokenizer is definitely not the bottleneck. It just happens to be the first step in inference, so it's what I did first.
Modal's GPU glossary is a good overview about how GPUs work [0]. Karpathy's LLM overview is a good high level overview on LLMs [1]. 3b1b's video (and subsequent videos) on transformers was excellent at helping me understand the math at a high level [2]. This matrix multiplication optimization worklog helped me understand writing better CUDA (not for beginner intro though) [3].
During this process I also asked ChatGPT a lot of questions.
I'm definitely open to suggestions about "how to learn" with all the new tools we have. I felt this has not been straightforward to figure out.
reply