Okay that’s even more pathetic. Is it using only CPU bound software rendering? I do 3d programming on MacOS. I can assure you that it’s not slow because it’s just super-duper complex. No, it just sucks.
>i didn't know that the ANE (and the data pipeline around it) was designed for CNN rather than transformers.
Multiple stories have reported that ANE came from Apple's self-driving car project that got canceled. (Makes sense since CNN is used for vision-related machine learning and enables cars to analyze their surroundings.) They spent 10 years and ~10 billion on research & development on a product that never got released so Apple is probably happy they're able to salvage some of that ai technology and put it in iPhones and Macs.
This is pure sunk cost fallacy. CNNs were from the deep learning ImageNet heydays, but now everything in that domain is equally done better by transformers. All you are doing is wasting area and saddling software with outdated hardware, and myopic PMs insisting on its use will create inferior products. Now that sounds a lot like the Apple AI efforts..
roads = ip packet network routing, DNS, telecommunications cabling
racist conventions/meetups = social media platforms like Facebook
Therefore, "I hate the internet" is analogous to "I hate the roads" because their perspective is dominated by seeing all the racists meetups that use the roads.
(Yes, people do "hate traffic" but that not's the same as "hating roads". Anyone who has travelled on unpaved dirt roads or gravel trails that make their teeth rattle will appreciate having paved roads. And that appreciation doesn't change even if a lot of KKK racist members also use roads.)
>the obvious (and rather malicious, I would say) PR play that parent comment points out.
The author Lawrence Ulrich references research from the Insurance Institute for Highway Safety. Why would it be "obvious" that they are paid PR shills for Tesla/Waymo?
I'm jaded and cynical so I can accept that they may be secret PR mouthpieces but isn't it reasonable to expect a little evidence rather than just throwing out a random accusation?
IIHS works for the insurance industry who will happily throw you under the bus if it makes things more predictable on their side. Not better, not lower claims, more predictable.
The alignment between what is good for you and good for the interests they answer two is not all that much.
But of course, they work very hard to market themselves as peddlers purely of safety, after all only a barbarian would be against safety <mockingly clutches pearls>
This isn't to say that some people they don't care, but giving a crap isn't gonna change the incentives.
How do using autonomous vehicles benefit their claims-settling? Wouldn't it make it unpredictable- in an accident between two self-driving vehicles, who is liable? What if both vehicles are from the same company? Doesn't this lead to extreme variability?
>The board game room looks like somewhere you'd have a conference call.
The author literally labeled that room the "Upstairs Office".
The rest of the house (kitchen, family room, etc) that's not dedicated to LAN computers looks like a normal Euro/Scandinavian type of clean minimalism aesthetic which a lot of people like.
My home currently has a lot of "traditional" style furniture and decorations hanging on the walls which some might call "warm & cozy". But others would disagree and think of it as "grandma's house" and as I've aged, I tend to agree. I'd like to eventually dial it back and redo the interior to be more minimalistic but I have other higher priorities than remodeling the house.
>, and you can learn anything from words an images.
A lot of info on Youtube is not available in alternative forms such as books or blogs with just words and images.
E.g. A random person uploads a video on how he repaired my particular washing machine by isolating the relay that broke. It's not in blog form because the guy is not an HTML expert or WordPress user. Instead, he simply used his smartphone to record a 10 minute video of what he did and uploaded it to Youtube.
Was I "fine" before Youtube?!? In the past with that type of situation, I would have spent $1000 on a new washing machine instead of spending $2 for a new relay part and soldering it myself. There was no equivalent resource available to show me how to DIY repair it. Trying to avoid $1000 by calling an appliance repairman wouldn't help because these days, they charge a $150 trip & diagnostic fee just to show up ... and then in the end, they say the electronics board that's fried is discontinued so they can't get the part which means the only the only remedy is to buy a new washing machine. So I would end up paying a guy $150 to tell me to pay another $1000 for a new washing machine... for a total of $1150. Youtube helps in avoiding all that. As a knowledge base for DIY learning how to repair random stuff around the house, there is no alternative to Youtube.
How did that person who uploaded the YouTube video figure it out? They didn't have a video to watch. So why couldn't you?
I'd say it's more of specific hypothetical, and I don't understand how one couldn't read about a fuse on a forum or in a manual and diagnose with a DMM. Sure, YouTube is convenient because we are all out of time, all the time.
It's also a further reflection of how helpless we have become with regards to so much of what we rely on.
>How did that person who uploaded the YouTube video figure it out? They didn't have a video to watch.
He spent a month trying different repairs (replacing pump, etc) that didn't work and then finally stumbled onto the electrical relay as the root cause. This is the deep link to the video I used: https://www.youtube.com/watch?v=3Gm_2qISYvk&t=2m41s
>So why couldn't you?
Maybe I eventually could have fixed it if I spent weeks on it and replicated all his failures but I didn't have to because ... he uploaded a video to help me. His video was the breakthrough I needed because it identified the particular relay. I actually had to watch more Youtube videos to learn how to actually replace the relay because the board is encased on epoxy resin and I had no idea how to deal with it. Various Youtube videos showed ideas of using an exacto knife and Dremel rotary tool to cut through the plastic case on the other side.
>, and I don't understand how one couldn't read about a fuse on a forum or in a manual and diagnose with a DMM.
One could hypothetically read the text of fixing a single relay of that particular washer model on a forum ... if... that guy in the vid hypothetically wrote a post on a forum... but he didn't. Sometimes, the mode of sharing is Youtube instead of a forum post. I can't control what other people do.
I did eventually find the official Samsung service manual PDF file and had to download it from some sketchy warez website because end-consumers are not supposed to have it ... but it only talks about board-level repairs ($250) instead of isolated repairs of replacing a single faulty relay.
Without hypotheticals or exaggeration, Youtube has literally saved me thousands of dollars. Showing how to fix cars, plumbing, ice makers, sewing machines, coffee machine, sprinkler systems, etc.
I'm with you on this, the knowledge sharing on YT is massive and in general not replicated in written form anywhere else. Anything that needs fixing involves a YT search for me these days, since that has the highest chance of turning up something useful and relevant. Only downside is that it's time consuming to find the right spot in the video and pause/rewind to understand how to do the thing properly. But having someone actually demonstrate how to do it, rather than just describing it with words, is so helpful and makes replicating it a lot easier.
> They didn't have a video to watch. So why couldn't you?
You can of course say the same thing about reading books, instruction manuals, asking somebody for advice, and so on.
Why should there exist any instruction manuals, when you could figure out everything in them for yourself. You aren't a stupid idiot, are you? Then why would you need a manual? And why would you ever ask other people for advice? That's only for stupid idiots who can't figure out things for themselves - not for hackers.
>So I would end up paying a guy $150 to tell me to pay another $1000 for a new washing machine...
I agree that there's a huge difference between $2 and $1150, but you're absolutely positive that only this Youtuber knew how to do it and that no repair man would've been able to?
Well, I'm reasonably certain there are dozens of YouTubers who figured it out. I wouldn't be surprised if I could figure it out myself, although it's unclear. But yes, repairmen, they would not be replacing a single relay. They would replace the whole board if it's available. Otherwise, the whole thing is scrapped.
Repairing a relay is something that's easy for someone to do by themselves. However, it is not worth it for repairmen. By the time they get the relay, which means a second trip for another $150, plus spend the time to unsolder it and re-solder it, so a bit more of their time, so now we're up to $400. Worse, they replaced the relay and sometimes it turns out the relay failed because something else is broke and another thing is just going to break the new relay right away and now they have to spend another $150 on a repair trip to figure out what was wrong that caused the relay to burn out, replace that part and replace the relay. We are very close to the cost of a new washing machine right now and that is assuming whatever is wrong is easy to repair and not the result of something else that is also broken. Thus, the right answer for a repairman is you make the simple, cheap, quick fixes you can or you just sell them a new one before you end up wasting more of their money than it
Modern factories are very efficient at making things, however, that only works when the assembly line is there. A repairman working alone doesn't have all the advantage of a factory. I was recently talking to one of our factory engineers at work and they said this part, whatever it was, took two 'man hours' to build in our factory. But a repairman just taking it apart to replace one cheap seal and putting it back together would be at 10 man hours. This is why more and more things are going to modules that can be replaced. If a repair person can come out and look and say, this module is broken, grab a new one, shove it in, and it works, they can be in and out quickly and it's cost-effective. If they have to tear down to the relay level and diagnose things if you're at why it really broke, that quickly becomes more labor and thus more cost than it's worth.
Your devices in the 1960s did work and they were worth repairing because they were a lot simpler. Those devices also didn't have nearly as much automation in the assembly line and so there were many more man hours to build in the first place. They were also, although this is less true depending on what we're talking about, but very often had unreelable parts that because they were broke often, they specifically designed those for your parable. You fixed a 1960s TV by replacing all the tubes, something that was not very hard to do because they put more expensive sockets under each tube to make it easy
>I realize I’m somewhat limited (16GB RTX 9070), but still, it seems really far off from the kind of experience even a basic $20/month subscription gets me.
I just ordered a new Mac Studio M5 Max 128GB $5899 ($6400 with tax) to be able to run the bigger "consumer size" models in the 70B parameter range (~96 GB). That said, I have no illusions that this expensive setup with a Qwen Flash coding LLM will be comparable to a $20/month subscription. Even upgrading to an even more expensive Mac Ultra 256GB for $10000 to hold a bigger model still won't be comparable. Apple hasn't shipped my Mac yet and I'm still considering cancelling it and downgrading to a smaller 64GB RAM config ($4299) to save $1600.
Why did I initially spend the extra $1600 if I knew ahead of time that it wasn't as good as cloud AI? Because I thought I could use some local LLM for the easy tasks or when I hit cloud rate limits. No issues with privacy so that wasn't part of the motivation at all. I just wanted some local AI capability to augment a subscription. I've not totally convinced myself of the cost/benefit of this.
Based on today's consumer hardware landscape, you're paying very high prices for crippled capability compared to the cloud AI subscriptions. We're also in a transition period where the next iteration of hardware improvements have some compelling features for local AI. Apple's upcoming M7 (2027 or 2028) is anticipated to have better GPU and neural engine to help with prefill TTFT. AMD Strix Halo is about to release 192GB system which is a big upgrade to their current 128GB ai pc. Maybe apply my $1600 savings towards those newer products. Those future products will still be very expensive but maybe the cost/benefit will be better.
> Why did I initially spend the extra $1600 if I knew ahead of time that it wasn't as good as cloud AI? Because I thought I could use some local LLM for the easy tasks or when I hit cloud rate limits.
The maths don't check. With Deepseek Flash one goes a very long way with 1600$ - even 10$/month, for easy jobs, are more than 13 years, and at a higher quality.
Oh no doubt. But one does have the guarantee that no bits left their home and that's a big one for some.
The low hanging fruit stuff for me is more something I use it for because I have the local LLM setup running anyway. It wasn't the reason I bought it, but now that it's there I might just as well use it as much as I can.
> Oh no doubt. But one does have the guarantee that no bits left their home and that's a big one for some.
No, this is a big misconception, and part of the cargo cult.
Use cases like the parent's are essentially about having a local LLM handle the leftover tasks. By that point, a lot of bits (main/big tasks) have already left home anyway.
It not only about it being an expensive setup (or not), and also other considerations:
- There's no guarantee of the $20/month service, and it likely has some limits compared to dedicated hardware token wise.
- Model are becoming more and more efficient, in many cases an M1 Max Mac Studio is still capable with 32 GB. 128 GB ram may not be the necessary baseline.
- Folks may think they want to only have a general model running locally (it's the comparable after all from the cloud providers), but we have to remember if the tasks we're trying to do ultimately are more specific than general and if there's space for the smaller models to do that.
I think the M5U Ultra 96gb is the sweetspot in that price range. It has more compute and bandwidth so you get to run models better sized to its hardware. I believe the Max would be too slow; personally I'm getting this SKU because I think it'd suck to get the 128gb Max and then discover it's too slow, and end up just using cloud providers anyway.
I agree about getting the Ultra if you're interested in AI (LLM) inference speed. I'm a little perplexed as to why there isn't a RAM option in between 96GB and 256GB, though. For instance, I believe Deepseek v4 flash runs a lot faster on (https://github.com/antirez/ds4) with 128GB than 96. I assume the higher SKUs have low enough sales that Apple doesn't want to be manufacturing too many different RAM configs as that would eat into their margins. (I say this as someone interested in AI hardware in general, even the 96GB is out of my price range).
It's the upsell ladder. This is how they get you to buy a more expensive tier or two. At first I was only going to look at 64gb or 128gb Max, but ended up here.
I also think it’s possible that it may be best to continue waiting on the Mac side of the house despite the increasing prices.
I think that some of the hardware design folks have been blindsided by AI demand and we haven’t really gotten that next generation AI hardware yet, to the point where buying M5 isn’t going to make sense in a couple of years.
Rumors seem to be that the M7 is the generation that Apple is looking to push AI performance much further.
I’m not sure that Apple anticipated this specific route that computer hardware has gone and I don’t think M5 and previous iterations were really specifically architected for local AI performance, more like they happened to be pretty good at it.
> Rumors seem to be that the M7 is the generation that Apple is looking to push AI performance much further.
What specifically does that mean? The M5 series has 10 cores per 128 bits of memory bus; are the cores unable to keep up with the RAM? I thought they did and memory bandwidth was usually the bottleneck. But the memory bus is already very highly clocked and goes up to 1024 bits wide so it's hard to picture memory bandwidth having a huge leap.
LLMs token generation is memory bandwidth constrained. If the M7 has double the memory bandwidth as some speculate [1], then it will help with LLM performance. There is also expectation (possibly unfounded) that the GPU will have improved matrix performance to help prompt processing as well.
14GT/s is a goal but it might be a while, I think 11-12 is more likely in that time frame.
50% more pins... we'll see. If they can reasonably make that fit then even more shame upon the traditional desktop CPU makers for sticking with 128 bits for so long.
Keep the memory. You’ll be glad you did when you realize that you’re better off with a solid coding model plus a good voice model and also a lightweight all-rounder all running at once isn’t of loading dynamically (slowly). It also helps if you want to be able to run a browser, IDE, and container environment.
> AMD Strix Halo is about to release 192GB system which is a big upgrade to their current 128GB ai pc
Big upgrade to memory capacity but memory speed is only going up by a few percent, so its still going to be slow with more than a few B active params (I have one)
you can get 4xGB10 for <20K so that gets you about the same tg and pp will be probably better. Power consumption though will be something like 200W idle so that's a bummer. And you get VLLM and SGLANG unlike them mac where one has to use oMLX (nice but not the same concurrency or cache performance)
I have 128GB M5 Max laptop and sill prefer to run things on other boxes in basement because its no fun to have the primary device being overloaded.
>What changed, why suddenly they adopt C++ features they explicitly excluded?
Project Valhalla, which includes the effort to add value types was announced in 2014. They've been working on it for a while.
As for "what changed" ...
Back in 1990s when Java was conceived, there was an idea that CPUs in desktops had plenty of extra cycles that were being wasted and could therefore be used to reduce mental load on developers. It was the same "cpus are cheaper than developers" idea that is repeated today with "tokens/cpu are cheaper than developers". With that philosophy, James Gosling talked about Java's "everything-is-an-object" as a mental simplification for developers. All the extra indirections of pointer-chasing to box unbox primitives and/or iterate through arrays of objects wasn't seen as a penalty (again, "CPUs are cheap; devs are expensive").
Now having value types that are contiguous in RAM is a big deal for performance. Avoid a bunch of pointer chasing. Even C++ best practices were affected. E.g. the traditional tradeoffs you learned from from classroom textbooks of linked-lists being faster than arrays for middle-of-list insertions was no longer always correct in the new world where CPUs are caching adjacent RAM areas to try to reduce the memory wall issue. So O(n) could be faster than O(log n) depending on the size of the data structure and interactions with RAM pre-fetch, etc.
Depends. Are you searching for the complicated thing by identity, or in a system that caches identities and/or interns objects of the type you’re handling? All
Of those can result in searches being word-based and thus vectorizable/cache-sympathetic more often.
So now the result of `new HashMap<>()` should be backed by an array for the first 200 elements or so? Potentially the size depending on the L1 size etc.
The 1990s were the peak of "you will be able to buy a better computer in a year and a half" and "the cost of computing power is going down rapidly" and, for me, the 2010s were the decade where you couldn't sell specialist VCs on any non-columnar query engine because they were all impressed by mechanical sympathy, more so than the mainstream programmer.
Today we're in the age where we can't count on your next computer being faster than your current computer or being more affordable, so the trade-offs look quite different -- it is feeling more like the 1980s where the Apple ][ line lasted almost a decade longer than Apple expected with (mainly) minor improvements in performance.
> Project Valhalla got its start in 2014, with the goal of bringing more flexible flattened data types to JVM-based languages, in order to restore alignment between the programming model and the performance characteristics of modern hardware. (In some ways, it got started much earlier; the designers of Java wanted to include value types in the initial version of the language.)
>I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked.
Having owned 3 MacBook Pros since 2008, the decision to make my next computer be a Mac Studio came down to (1) MacBook thermal throttling that slows down CPUs when it starts to overheat and (2) easier upgrade of Mac Studio SSD with after-market storage module whereas the MacBook requires more complicated disassembly and hot air gun to dislodge the surface mounted SSDs.
I have a brand new M5 Pro MacBook Pro I don't like it when the fans turn on. The Mac Studio will be faster and quieter for the same workloads.
Sorry for not being clearer. I don't like the MacBook's noise when the fans turn on.
The Mac Studio has bigger heat sinks to delay the need for thermal management -- and if its fan does need to turn on, the bigger size means it's still silent instead of the high-pitched whooshing noise the tiny fans make in the MacBook.
You can get large under-laptop cooling pads with giant fans that can help this somewhat. It's not part of the laptop, but help if it's docked at home. It's not totally ideal since the bottom of your laptop doesn't have radiator fins, but does help.
>European restaurants typically use a team service model,
American chain restaurants also use the team model. The server that writes down the order often isn't the same person bringing out the hot plates of food. Maybe less of team than Europe because the waitress taking the order will typically be the same one refilling the drinks.
It's the white tablecloth high-end restaurants where the one waiter is the only person handling all interactions with the table.
What you are describing is not team service because the server "owns" the customer and are the one who will collect the tip from that table.
What you are noticing is that many chains have a "full hands" rule: if any server with free hands walks past a table with dirty dishes on the way to the kitchen, they are expected to pick them up no mater who's section it is. Also, if they are heading out from the kitchen with free hands they are expected to run out some dishes no matter what table they are for.
>the server "owns" the customer and are the one who will collect the tip from that table.
In many USA restaurants, the waitstaff tips are also shared into a pool with the runners, busboys, and kitchen staff. The % may not be equal however.
>What you are noticing is that many chains have a "full hands" rule:
But that contradicts the idea that "USA restaurants almost always over staffed there." The chain restaurant could just "overstaff" with dedicated table runners with no need for a "full hands" rule.
It doesn't seem like USA restaurants are overstaffed nor do owners have the incentive to overstaff. The very common scenario is the hostess seats you at the table and the waiter doesn't even show up for 10 or 20 minutes to take the order. The patrons are frustrated and wondering why it's taking so long. That's a symptom of understaffing where the waiter is handling too many simultaneous tables.
> In many USA restaurants, the waitstaff tips are also shared into a pool
This doesn't change anything because the incentive is still to drive up the bill to drive up the tips. This is caused by the staff being primarily paid in tips.
> The chain restaurant could just "overstaff" with dedicated table runners with no need for a "full hands" rule.
Unlike servers, bussers and runners are generally paid minimum wage or close. Only over staffing servers helps make more income. Over staffing support staff runs up costs.
USA restaurant staffing is widely studied and your personal anecdotes don't change the conclusion: compared to most anywhere else, restaurants in the USA are over staffed with servers.
>Only over staffing servers helps make more income.
There are several gaps in your assertion about USA restaurants being overstaffed with servers.
#1 the mathematical relationship between servers and income is questionable: Servers acting as quasi commissioned salespeople by upselling patrons on "another beer?" or "anyone here want to order desert?" ... can be time-staggered interactions across multiple customers as the server rotates among their assigned tables.
Let's say, a restaurant has 20 tables and initially has 5 servers. Each server rotates among their assigned 4 tables. For a single turnover of 1 set of 20 tables, the diners have been asked(upsell) "Do you want desert?" 20 times. Let's say the owner wants to "overstaff the # servers" to be 1-for-1 and have 20 servers to 20 tables. The # of upsell events is still the same at 20. In what scenarios do overstaffing servers increase restaurant's income? (On the other hand, understaffing servers causes diners to twiddle their thumbs for half an hour after they finished eating because the overbusy server hasn't gotten them the final bill yet which causes less turnover of the tables which leads to less restaurant income.)
It's also not clear why you need tipped servers to upsell. Fast food places like McDonalds have trained minimum wage workers to constantly upsell... "Do you want fries with that? Do you want to upsize that drink?"
#2 the economics equilibrium of what tipped servers want to be paid and whatever staffing levels the restaurant owners want. Trying to overstaff the restaurant means waitstaff notice they don't work as many tables which means their tips (thus total income) is lower. This causes them to quit, which then causes the restaurant to keep trying to hire new waitstaff. Many restaurants chronically run "shorthanded" because of this negative feedback loop. (Which is one of the causes of extra delays between the hostess seating the diners and the time the waiter finally shows up.) Therefore, adding more servers to increase restaurant income isn't going to work.
>USA restaurant staffing is widely studied
The only studies I found analyzed restaurants not being able to hire staff including the servers. No study talked about restaurants having overstaffed servers in the USA as a proven driver of extra income.
EDIT to reply: >In Europe a typical place with 20 tables will have 2-3 servers and no runners. We go from 8-10 floor staff down to 2-3.
Your premise doesn't work for USA because with the hypothetical American restaurant that only has 3 servers with no runners and those 3 all sharing the workload "as a team" for all tables ... still means those 3 servers want to work for tips. That negates your parent post saying the higher # of staff is the barrier to eliminating tipping.
Instead, the barrier to end tipping in the USA is the inertia of the ingrained culture where servers expect tips and patrons keep paying them.
There are several high-profile restaurants in New York City, etc that tried to eliminate tipping by raising prices and they eventually reverted back to traditional tips.
>Hopefully it is clear how 5 servers dedicated to 4 tables each and backed up by runners can do a lot more upselling than 2-3 staff doing team service.
It isn't clear because those same lowered server count of 2-3 can still ask the same upselling questions... "Do you want desert?" Asking that question isn't mathematically constrained by the # of servers. Instead, it's constrained by the # of tables.
> Let's say, a restaurant has 20 tables and initially has 5 servers.
This is already overstaffed and you are doing the math backwards. The place with 5 servers in the USA almost always has runners as well making total floor staff 8-10.
The comparison is not a mythical restaurant with 20 servers for 20 tables. In Europe a typical place with 20 tables will have 2-3 servers and no runners. We go from 8-10 floor staff down to 2-3.
Hopefully it is clear how 5 servers dedicated to 4 tables each and backed up by runners can do a lot more upselling than 2-3 staff doing team service. And they don't need to because their salary isn't dependent on it.
Locking down 1 server to a strict set of 4 tables is very inefficient as well especially when one section is slammed and another is half empty.
You will find plenty of study of tipped wages and restaurant economics but they generally don't use words like "overstaffed" instead they say things like analysis found that owners systematically offload slow day risk to tipped employees.
reply