> "The speed of AI development right now is just crazy," Brooks said. "I can't imagine where we're going to be a year from now, three months from now, or even a month from now," he added.
I don't think I'm taking this out of context when I say this is unintentionally correct. Apple still doesn't know what to do about AI.
Luckily, it doesn't matter because it's a solution in search of a problem. Most consumers aren't using AI apart from google search.
Everyone else is using it as a content scraper and praying nobody will step in to end the piracy/fraud.
The new Siri isnt that exciting at least on my iPhone 15 Pro Max. I know it's beta but it's sluggish and often says try again. I watched many videos on Youtube saying its amazing but maybe not so much on older phones? Also, I need a Siri I can talk in an unfettered manner from my lock screen while Im driving without having to unlock the screen. Probably a big ask for Siri to know my voice via voice fingerprint allowing unfettered access from my lock screen.
Others running the beta now on newer iPhones and enjoying it more so?
> The new Siri isnt that exciting at least on my iPhone 15 Pro Max
On my Pro 16 it has its ups and downs - I still can't get it to "play my running playlist on shuffle" whilst running (this is the only thing I used Siri for before the beta and it would improve my life immeasurably if it worked). But it responds to things like "how long will it take to drive to the AirBnb booking in my inbox", and "when is X playing a concert in Y - add a calendar entry with details" perfectly.
This is a beta and I have hopes, but I can imagine it will run better on a 17 and later
> Luckily, it doesn't matter because it's a solution in search of a problem. Most consumers aren't using AI apart from google search.
This is... a view.
Maybe I live in a strange sphere of strange ("normie"-ish) people, but the people around me are for sure using AI. Mostly chatgpt to be fair. They use it to compare products that they intend to buy, identify plants in nature, create travel plans, find interesting places to visit nearby, give movie suggestions based on what they have previously enjoyed and so on and so forth. AI is becoming a very integrated part of their reality. To "google" something and digging through the search results manually is very rapidly being replaced by asking chatgpt, for better or worse.
Chatgpt is so damn good for cooking it is unbelievable. It will learn your family's tastes over time, you can tell it what pantry staples you keep is stock, and you can take any recipe you find online and ask for stuff like "find a way to make this preparable in 45 minutes instead of 2 hours, what trade offs will I be making?"
I agree! I think it's because there's so much cooking material in its training data. I wonder what proportion of the internet is food blogs and recipes... Probably a lot!
This is true but I can do the same on free ChatGPT without even logging in. I wouldn't pay $5 a month for that functionality, much less $20 (or >$1000 to be able to run it at home).
SOTA AI for "serious" work is in a different position, used by fewer people but with big pockets and sometimes a pathological dependence on it.
Nobody does what to do, they are just throwing trillions at it betting they’ll figure out.
If they don’t they’ll be screwed, if they do Apple will quickly catch up with a better and more refined product, as they’ve always done.
Running models on-device on a Mac is immensely annoying though. Figuring out what will work out of BF16, FP8, BF16+FP8, NVFP4, INT8, GGUF ... the list goes on ... is 'non-obvious' at best. Apple do little to support with tooling. There's MLX, but unless you're happy to transform a model to that format yourself you'll be lagging a long way behind.
Apps like LMStudio, Ollama, Draw Things, etc do a great job of simplifying it but it's still a pain.
How is it a pain exactly? It’s just learning and only takes a day or two to get up to speed. We seem to have forgotten that for the past fifty years doing all kinds of tasks on computers has been tedious and involved and time consuming to even get working. My first computer had 48kb of RAM and to play a game you had to load it off cassette for five minutes. That was annoying. Having LM Studio download a model and load so you can chat or attach an agent it is effortless and easy in comparison.
In general I just don't like not using a cross platform architecture of some kind.
MLX is fine, but the cross platform alternatives (ONNX) are terrible, GGUF often lacks and doesn't have the "easy convert" that some of the commenters below say MLX has.
Ok, I didn’t want to take the bait but this one’s just too much.
> “He also described a shift toward running AI locally rather than in the cloud – a move motivated by privacy, security, and the rising cost of inference as agents consume more tokens.”
Classic Apple. No more just beating the “security and privacy” drum, now its “tokens are expensive!”
<neanderthal voice/> Cloud scary. Cloud expensive. Mac good. Buy Mac!
> “He also singled out what he calls ‘transparent AI’ on iPhone and iPad, referring to features scattered throughout the operating system and third-party apps that work quietly without announcing themselves as AI.”
<neanderthal voice/> Apple use AI, Apple just not say it. Apple smart, not lagging behind industry! Buy iPhone!
How about you invest in developing your own models, correctly? And provide a secure and private inference cloud service on your fancy Apple silicon? And integrate that into your platform so Siri gets smarter without you farming queries out to Google Gemini? Bill me for it in iCloud+ I’ll probably pay for those tokens.
But why should apple invest in developing their own models? Why would it be correct?
Or phrase it in a very similar ask, why don't they invest in power plants? The model space is truly crowded, what do they gain or recover suppose they are SOTA? Across the Pacific they are pumping out free models that are only 6-12 months behind. What business sense does it make for Apple to develop their own models?
They don’t believe in this model and never had. You sound like the people who screamed that Apple was dying because they were not making a netbook style Mac in 2009.
Apple is the only big tech company with a non existent financial exposure to the current capex bubble.
Let big dogs bark at the moon. They are the loud ones, at least until the moon implodes.
I’m not seeing how it’s bad that a company is pushing in the direction of user hardware ownership. Of course it’s self-serving, that’s what companies do, but with most of the rest of the industry increasingly leaning in the direction of eliminating powerful general purpose computers in favor of thin dumb clients with useful compute being gated by subscriptions, it’s nice to see some dissent.
AI features not being constantly shoved in my face and just selectively silently integrated where it’s most useful is preferred to what the rest of the industry has been doing, too. I think most of us are pretty sick of AI getting tacked onto things that don’t need it and then given prominent promotion and UI positioning, potentially at the cost of features we actually use.
They could be doing more, sure, but directionally this all seems fine?
It’s not for the AI inference, it’s for the tool calls and desktop GUI app workloads and browser. There aren’t any on-device models capable enough of real work that can run on lower end Mac Minis. But for running a few browsers and GUI apps, you’re much better off buying a Mac Mini than paying for a more expensive and worse-performing container in the cloud. Browsers were not designed to run in Linux containers but they run optimally on baremetal desktop OSes. An M4 Mac Mini beats the single core performance of any VM you might rent in the cloud, in terms of raw compute per dollar (Geekbench scores).
Apple is a premium brand with high brand loyalty. Do you not think even 1 billionaire would want something like that? Even to just say that they bought it. Apple could sell things at a price point much more than $10k.
There's only so many billionaires and at Apple's scale you would not offer such a product for public sale even if you do custom builds for the rich and famous.
Apple makes product lines with assembly lines, its not a hand fab or custom build type of place.
At the end of the day, it’s just an Apple computer, not a Ferrari or an Aston Martin. I hardly think an Apple computer can be considered as a luxury item, unless they release it as a limited edition
I just went to Apple.com and specced out the top of the line 14" Macbook Pro, and the price is $9,849.00. Well over 10K with California sales tax.
I could absolutely justify having that machine for the actual work I do, and I'm not even doing any of the really hard AI things. If I were, I might prefer a Thelio Mega workstation from System76, which is $90,383.00 fully loaded.
Alas I can not afford a 10K personal computer right now, which is not the same thing.
> I cannot imagine a personal computing usage which can justify a 10k machine.
If Apple released a machine that would let me run, say, DeepSeek V4 Pro or GLM 5.2 locally at 100 tok/s for $10k, I might hurt myself running to get my credit card.
But then, I'm also posting this sitting in my truck while my family attends an event, with a Vision Pro on my face so I can monitor 7 Claude Code sessions without constantly switching screens on my laptop.
$10,000 divided by five divided by 12 is only $166.66 a month without interest over 5 years a Mac M2 Ultra or Mac M3 Ultra are easily usable computers over five years, so are most of the medium to high-end laptops that Apple sells for one person or locally in a small company.
It’s not necessarily out of reach or unusable over the course of time, the thing I keep hearing over and over is how usable many people find Mac’s very useable over the course of time, obviously software support plays a big part of that.
On top of that, AI models are getting more useful despite their smaller size. The future is a personal computer future, not a mainframe computer one. Yes, larger computers are useful, but most people will not be using that larger computers, which will be relegated to universities and larger companies.
Apples on stage use cases for their hardware and software makes me wonder if they actually use computers over there, or what a "job" at apple entails.
I am unsure that apple themselves understand why their hardware (top end & bottom end) has been so successful, without this understanding leaning into these use cases isn't really going to be possible.
Obviously they are playing 12d chess. They stopped selling high memory machines, they stopped selling pro machines. They are the king of local Ai compute.
With their apple finger right there on the pulse, they are going hard on the VR/AR glasses, cars and folding phones. By the end of the year (tm) we 100% will have all the features that were showcased and demonstrated 2 releases ago.
Sadly, that is an outdated PoV. It has probably not been valid, since last century.
It's just that Apple isn't really focused on software development professionals, and it's still fashionable to throw shade on them, so we hear a lot of kvetching about it, in communities like this.
I feel that it does, but I’m also a dev. I used to run a multi platform shop for years, and have a pretty good idea of what kind of support various companies give.
That’s just excuses for example how does a relatively small company like Black Magic Design With one of the best programs in its class manage to support three operating systems MacOS, Windows OS and Linux OS simultaneously?
There are many larger software companies that can support all three operating systems, they just make up excuses/reasons as to why they can’t perform over the years.
Ages ago, back when the Macs would come out, my co-workers and I would take a bit of time to configure the most expensive possible configuration --- time was, it was pretty easy to hit six figures, but over time, that has gradually come down.
I asked an Apple (via a sales rep who visited our company to showcase in internal iPad healthcare app) to please do this for iCloud when iCloud Drive was in-development. We would have easily paid $50,000 for a rack-able Mac Pro you could point "managed" devices at.
How do you design XServe being tied to Intel the way you want to make it, and why would you, when you’re having problems with Intel, why? would you go any further with Intel hardware wise? Makes no sense. The next four years will be interesting with an engineer CEO in charge.
Apple knows the market demand for this type of device.
You may have paid $50,000 for it, but you’re only one customer. At Apple scale they need to focus their finite resources on the products that serve the largest market demand.
$50,000 rack mount servers are not a large demand.
From a historic standpoint though Apple came back from near-death because they differentiated by focusing on the consumer first.
While other companies were recycling the same beige boxes meant to be tucked under desks for home use, Apple came out with products in candy colors.
While Microsoft was rolling out new business process tool SKUs, Apple came out with GarageBand and bundled it in for free.
Apple is not the company prioritizing going after a Fortune 500 company to replace their fleet with Macs. So their focus isn't going to be to design products and features to try to get that deal closed.
If I had the capital I’d make an household inference appliance.
No peripherals except Ethernet, integrated compute (cpu+gpu+mem) and secondary storage (+mobo, psu). No accoutrements, just the minimum amount of hardware to run a model as a utility.
Even the appliance faceplate would be a display showing stats like an old HiFi stereo.
Fairly sure most iGPUs these days are zero-copy and can dynamically allocate memory so what does "unified memory" mean to you exactly? A wider bus would be nice but it's not exactly a groundbreaking new invention.
> Unified memory in Linux creates a single address space accessible to both the CPU and GPU, eliminating the need to manually copy data between system RAM and video memory. It is enabled via NVIDIA's CUDA, AMD's ROCm/HIP, or generic kernel-level Heterogeneous Memory Management (HMM).
So it does exist and is available for platforms that matter.
Unfortunately their chatbot, while amazingly fast, doesn't know anything about the company running it.
Anyway I wouldn't mind an ASIC running a diffusion language model locally. Even if eventually it would become dated. Beats outsourcing all that to a company that's running on VC money which in the future might either perish or worse - dominate the market and charge whatever they wish.
Yes. Solar thermal heaters on the roof are common in Florida and other parts of the south. Some people also use heat recovery devices attached to the AC condenser. Further north I've only seen natural gas heating (e.g. in very rich NYC exurbs). The amount of shade over the pool has a big effect.
I'm keeping an eye on Tenstorrent for this. Pricing seems like its going to end up being in between a super memory dense unified memory platform, and a purpose built GPU.
Definitely on the edge of what would make sense at home, but its interesting.
Oh please the neural engine is mostly useless for LLMs. Siri in iOS 27 is laughably pathetic and slow compared to GPT Live DESPITE sending personal context to their (attested) cloud to execute anything but the most basic queries. Still years behind.
I think the relevant comparison is the developer beta, which has access to the Gemini-powered Siri that will roll out publicly later this year. From the reviews I've seen, Apple won't be "years behind" (which surely they were) for long.
I did say iOS 27. I am using the new Siri. It’s better, but extremely unreliable (half of all requests fail) and slow. It should be like a Codex running on my phone with the ability to chain skills (intents) to execute a task, but it’s too crappy for that.
And the voice is still a poor text to speech model, very far behind GPT live.
My guess is the OS. People who want a server often enough want to choose the OS, Apple wants to supply the OS and the hardware together so they're not blamed every time the two turn out to be incompatible, as happened the other way round in the 90s.
Target audience - B2B. There are multiple videos of Steve Jobs saying that he hates B2B, because the people using the devices are not the ones making the purchasing decision. It is pretty much against Apples DNA, and all their B2B they have today is a means for them to sell more B2C.
Because they killed the market, no one would now buy a macOS server, when Linux distributions, and to a lesser extent FreeBSD, own the server room.
They would even sell less than Windows Server licenses.
By the way, they are down the same path with the workstation market, now that they only top level answer is the Mac Studio.
Workstation market wants flexible towers that they can customise to their own liking and special use cases.
The main reason Swift exists for Linux, is that app developers need to have servers somewhere, and if they want to share Swift code with the backend, well it isn't going to be on macOS Server.
whos going to buy one? You cant trust them not to kill it within a few years and cease all software updates AND make it impossible to install a different OS to keep it going. Until they stop being dicks about what you can do with the hardware you own it's a non starter.
Apple itself was a major user of Xserve, Apples needs for cloud compute are massive and growing, and Apple could probably rent Xserve as a cloud to justify the cost and sell it to privacy focused consumers and businesses.
Being a Mac switcher since 2003 I am as much of a fanboy as anybody else but this quote from the article caught my attention, and smells like PR.
> Many AI tools are also Mac-first or Mac-only
I fail to recall AI tools Mac-only general purpose AI or agentic tools. Most of the claws, harnesses, studios and inference engines seem to be multiplatform. You can say you can run then in a Mac with a nicer UI wrapper or whatever, but "Mac-first" or "Mac-only"?
oMLX[1] is the only one that comes to mind but it's not exactly unique to mac, it just runs MLX models and provides a nice gui. It does have the whole paged SSD KV cache thing, not sure if thats working on other platforms.
> Apple's Mac mini and Mac Studio have become the machines of choice for running AI agents, according to Doug Brooks, Apple's senior product manager of Apple silicon.
This is mostly an US phenomenon, no Mac mini nor Mac Studio around here.
Only Thinkpads and Macbooks laptops talking to hyperscalers.
Around where? They're pretty popular for it in the UK right now given our obscene energy pricing as they end up being one of the best low power options for local llm. If you're not in the local llm space you obviously wouldn't see it. It's like saying Tennis isnt popular around here then admitting you dont frequent a tennis court so wouldn't even know.
What I’m not sure to understand is that if you want to just run Claude code or openclaw type software with llm apis or subscriptions (and not run local models) to benefit from a local file system and always-on capability for ‘second brain’ type of workflows, I guess you don’t need a Mac mini but can run it on a raspberry pi or an old laptop ? Does anyone have experience with that ?
Yup, for openclaw and APIs you dont need a big PC. I run something lightweight on the RPI4 8gb. Many people run local LLMs which is where a mac is useful. Frankly I dont think you can beat the value of an openrouter subscription and API calls.
> Many people run local LLMs which is where a mac is useful
Unless you go for the very expensive options, most of the Mac Minis really aren't suitable for running local LLMs, they're painfully slow with prefill/processing input, and the models you are able to run don't handle long context very well, which these sort of long-running agents perform very differently with when you can.
I'll agree with your latter point, hard to beat the value of using something like OpenRouter or similar remote inference.
Even with local models, you can run the agent software and the inference workload on different hosts, which is what I'm doing at home. Beefy server responsible for inference, tiny VM on other server is running the actual agent software + RPC + bridges and what not.
Why not go direct to the source instead of paying an extra 5.5%? Seems like it'd be trivial to have AI wire up connections to your preferred inference providers and save yourself some money over time.
If you're referring to the markup charged by OpenRouter, you can use harnesses like OpenClaw/Hermes without it and go direct like you're saying. If you're talking about actually "routing", then you don't get that out of the box. However, the popular use of those harnesses doesn't often use the smart routing approach with a single agent. Instead, the approach is to create multiple agents, each with a role and a model tailored to that role by cost and functionality.
If you or anyone else don’t mind, I have a have a question or 2.
I use Claude Pro ($20/m) as a glorified search engine (no ads/SEO) plus simple hobbyist dev things (shell scripts, managing my Mac, apps etc.
I also use it for tasks like - “search the web for top ten selling EVs, put them in a table” and then iterate - pivot tables, charts, additional research”. It could be cars, it could be broccoli. Code Work has facilities to streamline this type of work, but I usually drop into the CLI.
How much if any functionality would I need to recreate if I switch to OpenRouter and would be match my costs with the API approach. I don’t want any cost overruns. With Codex or Claude, if I run of tokens, no big deal, I can wait.
Pretty much. Running a local harness calling an llm via APIs doesn't necessarily take a lot of resources. But whatever tasks you want that agent to do via tool calling will be limited by the resources of the machine it runs on if you run those locally, so that's what should inform your choice of specs in this case
...or any cheap VPS? I now do most of "second brain" things via pi harness with Opencode Go subscription, and it costs me like 20 bucks a year, with added benefit of "you can have tmux and open session realtime on whatever device".
With how apple seemed to be caught by surprise when it came to Macbook Neo demand, I'm not sure they have the quantities of SoC's around to handle the demand a Mini Neo could drive. Especially if they could do it for $299.
I will go out on a limb and say that's not going to be an Apple product, period. It doesn't fit anywhere in the value envelope.
The relevant questions here are: will the person using this machine also conceivably be wearing a pair of $549 AirPod Max? Or a $399 base Apple Watch? Does that person expect to pay more or less for their largest-screen computing device than their headphones?
Framing that way points toward a $350 price point being a laptop for young children (younger than Apple Watch age, so lower elementary). That's a whole different software experience beyond just the hardware.
It's anecdotal but the kind of people I know that bought Mac Minis for this purpose are what I'd call "light techies." They definitely know how to use an iPhone or a Mac but would struggle on the CLI of a Linux box.
Anyone who wanted the OpenClaw use case that is comfortable with Linux probably already has several Linux machines (including a few Raspberry Pis) on-hand.
> Is it a lack of knowledge from the users or do they really value iMessage integration that much?
My understanding is that the barrier to entry to using iMessage makes iMessage a LOT more secure from spam. If you want to do mass iMessages you have to register as a business with Apple, go through all sorts of checks and attestations, etc.
At any rate, iMessages are a lot more trustworthy than SMS. So being able to spam people via iMessage is very desirable. I recall a few months ago a guy posting his little spam-iMessage-as-a-Service product here on HN. You could build your little iMessage spam army using a bunch of Mac Minis...
I do exactly this. You can run the whole thing on a PI. I have actually installed asahi Linux on my Mac and I connect to it remotely so you can be sure I will never upgrade my Mac again because it’s already overbuilt.
Correct. You do benefit from some headroom for things like launching browsers etc but refurbs or mini PCs (with at least 16gb; ideally 32gb of memory) from the likes of Minisforum or GMKTec work well enough if you're wanting to spend a little bit of money.
Yeah I find the Mac mini trend is kind of baffling.
It seems like it's driven either by 1) people hearing Macs are good for AI, buying one, and using Claude for inference, not realizing that you interact with the anthropic API from an internet connected hair dryer. Or 2) people want their agents to have blue bubbles.
I find it hard to believe that enough normal people are doing on device inference is driving Mac Mini's out of stock. And even if they were the Mac mini is not actually a very good platform for it.
One aspect you're missing is that people running a claw type agent thing need to run it on a Mac to automate software in the Apple ecosystem.
Neo-Siri in iOS 27 removes the need for a lot of this, but before then, if you want to ask a robot about information that is stored in Apple notes, or to send an iMessage, a Mac mini is your only practical option.
The demand is not coming from 'normal Apple customers' it's coming from people who want a machine that can run local AI.
It has nothing to do with Macs being especially good at AI. It has everything to do with being one of the last 'cheap' devices being sold with that much unified RAM.
There are two angles to this. One is that if you want to integrate your agent setup into the Apple ecosystem you need a network connected Mac running 24/7.
The second is that the puck is heading towards local models. The people running their own 'Claws are usually experimenting running their own services either to save money or to explore the future where 95% of requests are handled on device.
> "people often want a system that's under their control, isolated from their primary machine, and capable of running 24 hours a day, seven days a week," said Brooks. "A Mac mini is an amazing system for that," he added.
These execs are so out of touch they believe Apple hardware to be "a system that's under their control", how does it come to this? Besides, a VM without bi-directional sharing of data gives you pretty much the exact same thing.
Did hundreds/thousands of developers really go out there and bought Mac Minis just because one prominent technology semi-celebrity happens to have used a Mac Mini for the development of their thing? Seems bananas people would spend hundreds on monies on something they barely grasp how it works.
The remote access story for macOS is absolute sadness, without Jump Desktop there would be zero performant ways to access that “system under my control”.
And all of that because Tim Apple fears any feature that could mean people could have less than one iDevice per person.
I'm curious what you mean. I have been accessing Macs remotely over SSH and VNC for like 20 years and it's always been easy and as performant as the network would allow.
The built in high rez screen sharing between Macs works well too. Through Tailscale I’ve accessed my main Mac from both the opposite coast of the US and from the other side of the Pacific and it works great.
Host resolution automatically matches that of client, image quality is great, framerate is decent, latency is minimal. The host creates virtual screens for the connection so connected screens don’t light up and the machine remains locked to anybody accessing it physically too, which is a nice privacy assurance.
The issue with the Apple is that they didn’t really develop any competitive local AI machine. Their strategy/marketing falls flat when you ask them how exactly they implement AI: they buy it from Google cloud. In the future local AI may become a thing but that’s 4-5 years away. I count the 2q-4q and atrocious performance as “local ai” only for the enthusiasts crowd not for people doing competitive work.
Apple could dominate this niche if they decided, for a while until prices fall, to eat some margin and bump up RAM in high end models. Couple that with a new M series chip with even faster AI performance.
It’s not a huge niche but it’s an influential one. They’d get the engineers and CXOs of AI ventures and a lot of academics and hobbyists.
For the platform it would keep them cemented as the high end vendor. In the long term it would position them to take advantage of any software or training breakthroughs that deliver frontier model performance at that scale.
Now just imagine they'd kept making the Mac Pro and enabled compute offload to GPUs. Or even just passthrough to Linux VMs. Would've been quite the AI machine.
Except that Mac ultra M3 they talk about is now only being sold in the 96 GB configuration. It’s no longer being sold in larger Ram configurations by Apple Apple because of the global RAM shortage. And you can not add memory after purchase because it’s integrated/soldered on.
> And you can not add memory after purchase because it’s integrated/soldered on.
For as much as I dont like this aspect of modern computing, I understand why it is done from a technical perspective. Power, heat, and performance are all "better" when ram is on the motherboard vs in a "stick".
Apple has totally failed to deliver interesting AI experiences so far ... and I still think they're going to be the dominant provider of AI in 5 years. We're just one or two advances in chips / models / both away from being able to run very good local models for free on mid-tier Apple devices. The privacy, cost, and latency story there will be too much for OpenAI/Anthropic/Google to beat.
Just writing this down so I can be praised/mocked in 5 years.
I'm also of this opinion, but also that it doesn't have to be Apple (but they are well positioned). What I've seen with running local models on my 48GB M4 MBP is really impressive - it's not the same level as hosted stuff, but it's better than what I was using a year or two ago.
I suspect we'll see a hybrid before an all or nothing. Local models for computer control or delegating, online models for things that need strong reasoning, planning, and knowledge access. Again, I'd be more than happy to be wrong. I just see models growing faster than the hardware can.
Apple has failed to live up to the Steve Jobs era and initial iPhone hype. It's not just "AI experiences", it's computing in general. Maybe the consumer sector is just dead/dying. Maybe consumers are just running out of cash because filthy VCs are destroying communities and forcing the 99% into poverty.
The one thing that is marginally exciting: the Apple SoC or M series chips.
It's unfortunate they are locked behind crappy macOS and other proprietary apple crap.
Unsurprising. Apple seriously thought the iPad would replace computers and usher in a "post-PC" word during their "What is a computer?" ad campaign era. Now they are sticking phone chips in laptop chassis.
For the general consumer, they were basically right though. Most people don't use laptops except for work. The primary computing device is the phone, and phones have basically become become mini-ipads in form factor since that ad aired.
> Apple seriously thought the iPad would replace computers
...for some users. See their "Mac is a truck" analogy.
And it has. My parents haven't owned a Windows or Mac machine in six years, since they got rid of the one I gifted to them a decade ago. Its all iPad and iPhone.
You should not buy a fully loaded Mac Studio for AI unless you absolutely NEED macOS. You will be wasting so much electricity idling on prefill while your GPU pulls 150-250w from the wall.
Buy an Nvidia Spark, then whatever cheap Mac you want to use as a thin client. There's no reason to force Apple Silicon's round peg into a square hole like AI inference.
I don’t really think the electricity cost of prefill is a significant differentiator, is it?
I would imagine the time spent waiting is more of the issue.
I also imagine for a lot of people that macOS being easy and software-rich is half of the appeal. Waiting longer for prefill is potentially better than fiddling in the Linux command line for hours for a certain buyer.
A real dummy can walk into an Apple Store, buy a Mac, download LM Studio, drag it into the applications folder, and they’re up and running.
Outperforms doing what? Inference is not a homogeneous workload, memory bandwidth correlates to decode speed and layer swapping but not necessarily inference speed overall.
The other half of that equation is latency, predicated on prefill performance which needs a powerful GPU and ideally ALU-level optimization to build larger KV caches quickly. Even the M5 gets smoked in this department, the M5 Max has a 50% longer TTFT on Qwen's 27b dense model at only 16k of context, which is a pretty typical starting context to use for agentic editing in normal apps like OpenCode/Claude Code: https://raw.githubusercontent.com/Osmantic/MMBT-Messy-Model-...
For agentic, 50-256k token on-device coding sessions, the Spark will be faster and consume less power running larger models. Without an external GPU (which Apple doesn't support), Apple Silicon will always be bottlenecked during prefill. Apple's failure to address this with their GPU architecture is a big reason why Apple Silicon viewed as a waste of time and money for professional datacenter deployment.
For agentic work, you just cache the prefill kv cache of the relevant system prompts. TTFT is a little slow (10s of seconds, oh no!) the first time you boot up a new harness.
I keep hearing people make this claim that TTFT is a problem, and… it just isn’t, if you’re running oMLX.
Folks in my camp keep saying this, and folks in your camp keep beating a drum we tell you isn’t resonating. Not sure why I keep bothering to argue; you can’t buy a high RAM Mac Studio like mine anymore.
They're not usable for deployment. They're perfectly fine for "enthusiast" low-end usage with 10-30B models, but the same goes for almost every dGPU made in the last 10 years.
I'm perfectly happy with Apple not becoming an "everything we do is AI-centric" business.
I'm fatigued by it all at this point. It's streamlining the interesting and fun parts out of my job (by practical necessity of use there), and if I used it half as much outside of work I'm sure it'd do the same there too.
> It's streamlining the interesting and fun parts out of my job
Interesting. For me it's streamlining the tedious and attentionally taxing parts of my work tasks. I love solving problems, I don't particularly love shaving yaks.
I don't mind the subtle ML integrations that they have put in the photos app: plant ID, recognizing faces, removing background, OCR text search (even for handwriting!), etc.
You’re listing every feature I’d like to remove; Meanwhile Android users have great holiday pictures at the bottom of the Pyramids because they can remove the people from their pictures.
Apple is doing something very different. Their AI experience for end users definitely has been a little behind.
Apple Silicon, however, has been quite unique for the last 4-6 years and it's increasing overlap with LLMS.
The model/chip optimizations are definitely improvements, the thing that is really standing out the past 2 years is how much the open source model community has been making possible, especially when you know a group of use cases.
Here’s the two main reasons why local inference won’t compete any time soon with the cloud:
1. Most useful LLM work is done in parallel. A Mac Mini can run one LLM inference thread at a time. The cloud can spool up dozens and spread that inference across efficiently batched operations over a fleet of hardware.
2. Faster inference hardware such as the chips from Cerebras and Groq cannot be run locally. But the advantages of running >5x the token throughput per thread can’t be overstated. Add in the multi-threading advantage and it’s a knock-out punch for local LLMs.
Local inference has a role: if you’re working with extremely private matters or you want an uncapped model that will talk dirty or generate NSFW photos, local is the only option. I think Apple and others will continue to also run a lot of useful workloads locally such as text editing suggestions, speech to text, text to speech, and image manipulation. As local hardware improves, these capabilities will get better too.
But, for most LLM work, the cloud will continue to dominate for a long time to come, if not forever.
You can always buy multiple Macs. I think Apple's great differentiator could be making frontier class models local. I you want more threads, buy more Macs.
I don't want to run any workflows on someone else's computers.
As far as I know, there isn’t an interface like nvlink that allows these macs to work in tandem; they would just send their data over ethernet, maybe thunderbolt/usb-c?
> A Mac Mini can run one LLM inference thread at a time.
That’s not accurate. With MLX, at least, parallel inference is both possible and useful. Model serving tools like LM Studio and oMLX support parallel generation with continuous batching, and the total throughput increases with it.
You are not wrong, but the practical reality of local hardware is to be batch-constrained in comparison with a multi-user inference cloud. You will never be able to compete cost effectively in your home lab with a cloud that has >100M end users streaming millions of inference requests per second across a gigantic fleet of machines.
Can I run a few inferences in parallel on my Mac Mini? Yes. But put 1,000 Mac Minis in a datacenter serving 1,000 copies of myself? That's going to be more efficient.
Oh certainly, a local machine can't parallelize like the cloud can. Does anyone think that, though? What it can do is provide good-enough inference to support common business needs, without creating unnecessary data risks.
Why on earth would you need to do that for most things? Not everyone wants to be the next Coreweave. It sounds like you're killing a fly with a cannon to me.
There is so much happening in that scene, where tokens/sec double or 10x
So I could see the same hardware doing 20 tokens/sec on a large model suddenly doing 200 tokens/sec in the future, a better device in the future doing 500 tokens/sec, while having vision models baked in, audio models etc
Users wont consciously switch to local, they will just have it and use it
and a toyota camry can’t compete with a lamborghini but the camry vastly outsells them.
you’re far far far more likely to see a camry or equivalent in an americans driveway than you are a super car.
you’re also more likely to see an enthusiast with a corvette or equivalent muscle car spending way more than it’s worth to tinker on that car in their garage than you are a super car.
I guess what I'm doing is not considered that useful then? I usually only have zero, one, or occasionally two things actively doing inference at a time, be it claude code sessions or one of the chatgpt/claude web interfaces, and i bet that's true for like 95% of people using llms. And anyway i bet even the hardcore people using a bunch of parallel agents would appreciate having access to local, private inference for some things.
You're obviously right though that cloud inference isn't going away anytime soon
What about when claude code starts spawning sub-agents? it's pretty good about doing that now even if a user doesn't request it. Does this could as parallel inference?
I haven't seen it do it by itself that much so far. Even then, a lot of the time is spent waiting for some shell command to complete, which isn't inference. I guess being able to do parallel inference is like having multiple cores in a cpu, you can get by without it by just sharing that single core, and having it just gives you a performance boost under certain circumstances
Interesting. For me it seems to be much more common that even basic requests start spinning up sub-agents. But I also routinely ask it to do this so maybe it's stored this as a preference somewhere deep in its bowels.
And yes, I agree. I find the experience better less for speed and more for context management. But it's far from necessary.
100%. Orchestrating multiple sub agents is a key requirement for many workflows. If only because it minimizes context sprawl (amongst multiple other reasons).
There is probably a middle ground though. Something like a main thread agent (say a local model for Hermes) orchestrating cloud based sub agents.
There is common wisdom: during gold rush, sell shovels.
Apple's shovel (ahem, Mac mini) is the highest quality.with Companies burning money left, right and center, Apple can dispense with advertising altogether
When you don't limit on-device AI to "can be used to run coding model", Mac minis are great. My M4/16gb been working on a long term research project using Qwen 2.5 8b for months now. The performance is good enough for processing a lot of small text prompts.
On-device isn't only cheaper/faster - for health data it's the whole point! The self-sovereign community desperately needs an affordable AI appliance and Apple is still the most likely candidate to deliver on the privacy promise.
219 comments
[ 693 ms ] story [ 358 ms ] threadI don't think I'm taking this out of context when I say this is unintentionally correct. Apple still doesn't know what to do about AI.
Luckily, it doesn't matter because it's a solution in search of a problem. Most consumers aren't using AI apart from google search.
Everyone else is using it as a content scraper and praying nobody will step in to end the piracy/fraud.
Others running the beta now on newer iPhones and enjoying it more so?
On my Pro 16 it has its ups and downs - I still can't get it to "play my running playlist on shuffle" whilst running (this is the only thing I used Siri for before the beta and it would improve my life immeasurably if it worked). But it responds to things like "how long will it take to drive to the AirBnb booking in my inbox", and "when is X playing a concert in Y - add a calendar entry with details" perfectly.
This is a beta and I have hopes, but I can imagine it will run better on a 17 and later
This is... a view.
Maybe I live in a strange sphere of strange ("normie"-ish) people, but the people around me are for sure using AI. Mostly chatgpt to be fair. They use it to compare products that they intend to buy, identify plants in nature, create travel plans, find interesting places to visit nearby, give movie suggestions based on what they have previously enjoyed and so on and so forth. AI is becoming a very integrated part of their reality. To "google" something and digging through the search results manually is very rapidly being replaced by asking chatgpt, for better or worse.
SOTA AI for "serious" work is in a different position, used by fewer people but with big pockets and sometimes a pathological dependence on it.
I can see this kind of low level usage as being perfect for local LLMs... So I can't see a market there for openai etc forever.
Are they using the free chatgpt or a paid one?
>"I can't imagine where we're going to be a year from now, three months from now, or even a month from now,"
I'd say he's making an accurate appraisal of his abilities
Apps like LMStudio, Ollama, Draw Things, etc do a great job of simplifying it but it's still a pain.
[1] https://github.com/antirez/ds4
MLX is fine, but the cross platform alternatives (ONNX) are terrible, GGUF often lacks and doesn't have the "easy convert" that some of the commenters below say MLX has.
> “He also described a shift toward running AI locally rather than in the cloud – a move motivated by privacy, security, and the rising cost of inference as agents consume more tokens.”
Classic Apple. No more just beating the “security and privacy” drum, now its “tokens are expensive!”
<neanderthal voice/> Cloud scary. Cloud expensive. Mac good. Buy Mac!
> “He also singled out what he calls ‘transparent AI’ on iPhone and iPad, referring to features scattered throughout the operating system and third-party apps that work quietly without announcing themselves as AI.”
<neanderthal voice/> Apple use AI, Apple just not say it. Apple smart, not lagging behind industry! Buy iPhone!
How about you invest in developing your own models, correctly? And provide a secure and private inference cloud service on your fancy Apple silicon? And integrate that into your platform so Siri gets smarter without you farming queries out to Google Gemini? Bill me for it in iCloud+ I’ll probably pay for those tokens.
Was that so hard?
Or phrase it in a very similar ask, why don't they invest in power plants? The model space is truly crowded, what do they gain or recover suppose they are SOTA? Across the Pacific they are pumping out free models that are only 6-12 months behind. What business sense does it make for Apple to develop their own models?
https://machinelearning.apple.com/research/introducing-third...
they just suck
I agree, apple shouldn't invest in their own models. But they should have close to the best inference + end user design.
AI features not being constantly shoved in my face and just selectively silently integrated where it’s most useful is preferred to what the rest of the industry has been doing, too. I think most of us are pretty sick of AI getting tacked onto things that don’t need it and then given prominent promotion and UI positioning, potentially at the cost of features we actually use.
They could be doing more, sure, but directionally this all seems fine?
It would need a path to a $2,500 machine. But this is a niche I don’t think another consumer-facing brand could do like Apple.
Apple makes product lines with assembly lines, its not a hand fab or custom build type of place.
I think "buy this $10,000 box and to easily grant every Macbook Neo on your team safe, private, free AI" could be a real winner.
I could absolutely justify having that machine for the actual work I do, and I'm not even doing any of the really hard AI things. If I were, I might prefer a Thelio Mega workstation from System76, which is $90,383.00 fully loaded.
Alas I can not afford a 10K personal computer right now, which is not the same thing.
If Apple released a machine that would let me run, say, DeepSeek V4 Pro or GLM 5.2 locally at 100 tok/s for $10k, I might hurt myself running to get my credit card.
But then, I'm also posting this sitting in my truck while my family attends an event, with a Vision Pro on my face so I can monitor 7 Claude Code sessions without constantly switching screens on my laptop.
It’s not necessarily out of reach or unusable over the course of time, the thing I keep hearing over and over is how usable many people find Mac’s very useable over the course of time, obviously software support plays a big part of that.
On top of that, AI models are getting more useful despite their smaller size. The future is a personal computer future, not a mainframe computer one. Yes, larger computers are useful, but most people will not be using that larger computers, which will be relegated to universities and larger companies.
See https://en.wikipedia.org/wiki/The_Cult_of_Mac
I am unsure that apple themselves understand why their hardware (top end & bottom end) has been so successful, without this understanding leaning into these use cases isn't really going to be possible.
You have a bold career as a technical journalist ahead of you!
With their apple finger right there on the pulse, they are going hard on the VR/AR glasses, cars and folding phones. By the end of the year (tm) we 100% will have all the features that were showcased and demonstrated 2 releases ago.
I trust they know more about their business model than some rando on the internet, sorry.
How you hold it is, of course, up to you.
It's just that Apple isn't really focused on software development professionals, and it's still fashionable to throw shade on them, so we hear a lot of kvetching about it, in communities like this.
There are many larger software companies that can support all three operating systems, they just make up excuses/reasons as to why they can’t perform over the years.
Apple simply cannot comprehend the ask.
Apple knows the market demand for this type of device.
You may have paid $50,000 for it, but you’re only one customer. At Apple scale they need to focus their finite resources on the products that serve the largest market demand.
$50,000 rack mount servers are not a large demand.
From a historic standpoint though Apple came back from near-death because they differentiated by focusing on the consumer first.
While other companies were recycling the same beige boxes meant to be tucked under desks for home use, Apple came out with products in candy colors.
While Microsoft was rolling out new business process tool SKUs, Apple came out with GarageBand and bundled it in for free.
Apple is not the company prioritizing going after a Fortune 500 company to replace their fleet with Macs. So their focus isn't going to be to design products and features to try to get that deal closed.
No peripherals except Ethernet, integrated compute (cpu+gpu+mem) and secondary storage (+mobo, psu). No accoutrements, just the minimum amount of hardware to run a model as a utility.
Even the appliance faceplate would be a display showing stats like an old HiFi stereo.
> Unified memory in Linux creates a single address space accessible to both the CPU and GPU, eliminating the need to manually copy data between system RAM and video memory. It is enabled via NVIDIA's CUDA, AMD's ROCm/HIP, or generic kernel-level Heterogeneous Memory Management (HMM).
So it does exist and is available for platforms that matter.
Intel and AMD had been doing this for years already, and had linux support for it from day 1.
Otherwise, AMD is quite close to what Apple has, and Strix Halo is honestly incredible.
Not sure what RDMA brings to the table.
All have unified memory. Linux runs just fine on all of those.
A bit too expensive for a home appliance though, isn't it?
95% of the price is going to be in GPU+CPU+RAM
https://taalas.com/products/
Unfortunately their chatbot, while amazingly fast, doesn't know anything about the company running it.
Anyway I wouldn't mind an ASIC running a diffusion language model locally. Even if eventually it would become dated. Beats outsourcing all that to a company that's running on VC money which in the future might either perish or worse - dominate the market and charge whatever they wish.
Definitely on the edge of what would make sense at home, but its interesting.
That said, Siri seems a little bit better now - my subjective opinion. It is a little bit less frustrating.
And the voice is still a poor text to speech model, very far behind GPT live.
They would even sell less than Windows Server licenses.
By the way, they are down the same path with the workstation market, now that they only top level answer is the Mac Studio.
Workstation market wants flexible towers that they can customise to their own liking and special use cases.
The main reason Swift exists for Linux, is that app developers need to have servers somewhere, and if they want to share Swift code with the backend, well it isn't going to be on macOS Server.
but Apple needs to change the licensing model, currently you are allowed to run only 2 macOS VMs for every physical one you buy
> Many AI tools are also Mac-first or Mac-only
I fail to recall AI tools Mac-only general purpose AI or agentic tools. Most of the claws, harnesses, studios and inference engines seem to be multiplatform. You can say you can run then in a Mac with a nicer UI wrapper or whatever, but "Mac-first" or "Mac-only"?
[1] https://omlx.ai/
So the ad free Apple on device experience will be welcome.
https://ads.apple.com/maps
This is mostly an US phenomenon, no Mac mini nor Mac Studio around here.
Only Thinkpads and Macbooks laptops talking to hyperscalers.
People are buying apple unified as electricity costs in many countries are very high, so cheaper to run than Nvidia setup.
As non-apple unified memory options increase, many people will have more choose those
Unless you go for the very expensive options, most of the Mac Minis really aren't suitable for running local LLMs, they're painfully slow with prefill/processing input, and the models you are able to run don't handle long context very well, which these sort of long-running agents perform very differently with when you can.
I'll agree with your latter point, hard to beat the value of using something like OpenRouter or similar remote inference.
Even with local models, you can run the agent software and the inference workload on different hosts, which is what I'm doing at home. Beefy server responsible for inference, tiny VM on other server is running the actual agent software + RPC + bridges and what not.
I use Claude Pro ($20/m) as a glorified search engine (no ads/SEO) plus simple hobbyist dev things (shell scripts, managing my Mac, apps etc.
I also use it for tasks like - “search the web for top ten selling EVs, put them in a table” and then iterate - pivot tables, charts, additional research”. It could be cars, it could be broccoli. Code Work has facilities to streamline this type of work, but I usually drop into the CLI.
How much if any functionality would I need to recreate if I switch to OpenRouter and would be match my costs with the API approach. I don’t want any cost overruns. With Codex or Claude, if I run of tokens, no big deal, I can wait.
Thanks!
OpenClaw supports all the mainstream (and free) chat apps like Discord, WhatsApp, Signal, Telegram... None of them requiring a MacOS machine.
Is it a lack of knowledge from the users or do they really value iMessage integration that much?
The relevant questions here are: will the person using this machine also conceivably be wearing a pair of $549 AirPod Max? Or a $399 base Apple Watch? Does that person expect to pay more or less for their largest-screen computing device than their headphones?
Framing that way points toward a $350 price point being a laptop for young children (younger than Apple Watch age, so lower elementary). That's a whole different software experience beyond just the hardware.
Anyone who wanted the OpenClaw use case that is comfortable with Linux probably already has several Linux machines (including a few Raspberry Pis) on-hand.
My understanding is that the barrier to entry to using iMessage makes iMessage a LOT more secure from spam. If you want to do mass iMessages you have to register as a business with Apple, go through all sorts of checks and attestations, etc.
At any rate, iMessages are a lot more trustworthy than SMS. So being able to spam people via iMessage is very desirable. I recall a few months ago a guy posting his little spam-iMessage-as-a-Service product here on HN. You could build your little iMessage spam army using a bunch of Mac Minis...
It seems like it's driven either by 1) people hearing Macs are good for AI, buying one, and using Claude for inference, not realizing that you interact with the anthropic API from an internet connected hair dryer. Or 2) people want their agents to have blue bubbles.
I find it hard to believe that enough normal people are doing on device inference is driving Mac Mini's out of stock. And even if they were the Mac mini is not actually a very good platform for it.
Neo-Siri in iOS 27 removes the need for a lot of this, but before then, if you want to ask a robot about information that is stored in Apple notes, or to send an iMessage, a Mac mini is your only practical option.
It has nothing to do with Macs being especially good at AI. It has everything to do with being one of the last 'cheap' devices being sold with that much unified RAM.
The second is that the puck is heading towards local models. The people running their own 'Claws are usually experimenting running their own services either to save money or to explore the future where 95% of requests are handled on device.
These execs are so out of touch they believe Apple hardware to be "a system that's under their control", how does it come to this? Besides, a VM without bi-directional sharing of data gives you pretty much the exact same thing.
Did hundreds/thousands of developers really go out there and bought Mac Minis just because one prominent technology semi-celebrity happens to have used a Mac Mini for the development of their thing? Seems bananas people would spend hundreds on monies on something they barely grasp how it works.
And all of that because Tim Apple fears any feature that could mean people could have less than one iDevice per person.
Host resolution automatically matches that of client, image quality is great, framerate is decent, latency is minimal. The host creates virtual screens for the connection so connected screens don’t light up and the machine remains locked to anybody accessing it physically too, which is a nice privacy assurance.
It’s not a huge niche but it’s an influential one. They’d get the engineers and CXOs of AI ventures and a lot of academics and hobbyists.
For the platform it would keep them cemented as the high end vendor. In the long term it would position them to take advantage of any software or training breakthroughs that deliver frontier model performance at that scale.
For as much as I dont like this aspect of modern computing, I understand why it is done from a technical perspective. Power, heat, and performance are all "better" when ram is on the motherboard vs in a "stick".
Just writing this down so I can be praised/mocked in 5 years.
The one thing that is marginally exciting: the Apple SoC or M series chips.
It's unfortunate they are locked behind crappy macOS and other proprietary apple crap.
Unsurprising. Apple seriously thought the iPad would replace computers and usher in a "post-PC" word during their "What is a computer?" ad campaign era. Now they are sticking phone chips in laptop chassis.
...for some users. See their "Mac is a truck" analogy.
And it has. My parents haven't owned a Windows or Mac machine in six years, since they got rid of the one I gifted to them a decade ago. Its all iPad and iPhone.
We are both late and early.
https://news.ycombinator.com/item?id=35527692
Buy an Nvidia Spark, then whatever cheap Mac you want to use as a thin client. There's no reason to force Apple Silicon's round peg into a square hole like AI inference.
I would imagine the time spent waiting is more of the issue.
I also imagine for a lot of people that macOS being easy and software-rich is half of the appeal. Waiting longer for prefill is potentially better than fiddling in the Linux command line for hours for a certain buyer.
A real dummy can walk into an Apple Store, buy a Mac, download LM Studio, drag it into the applications folder, and they’re up and running.
The other half of that equation is latency, predicated on prefill performance which needs a powerful GPU and ideally ALU-level optimization to build larger KV caches quickly. Even the M5 gets smoked in this department, the M5 Max has a 50% longer TTFT on Qwen's 27b dense model at only 16k of context, which is a pretty typical starting context to use for agentic editing in normal apps like OpenCode/Claude Code: https://raw.githubusercontent.com/Osmantic/MMBT-Messy-Model-...
For agentic, 50-256k token on-device coding sessions, the Spark will be faster and consume less power running larger models. Without an external GPU (which Apple doesn't support), Apple Silicon will always be bottlenecked during prefill. Apple's failure to address this with their GPU architecture is a big reason why Apple Silicon viewed as a waste of time and money for professional datacenter deployment.
I keep hearing people make this claim that TTFT is a problem, and… it just isn’t, if you’re running oMLX.
Folks in my camp keep saying this, and folks in your camp keep beating a drum we tell you isn’t resonating. Not sure why I keep bothering to argue; you can’t buy a high RAM Mac Studio like mine anymore.
They're not usable for deployment. They're perfectly fine for "enthusiast" low-end usage with 10-30B models, but the same goes for almost every dGPU made in the last 10 years.
You think this is a mistake...
Of course. Do you think this was on purpose? All part of Apple's brilliant master plan?
I'm fatigued by it all at this point. It's streamlining the interesting and fun parts out of my job (by practical necessity of use there), and if I used it half as much outside of work I'm sure it'd do the same there too.
This is the prevailing opinion of people even outside of tech.
Interesting. For me it's streamlining the tedious and attentionally taxing parts of my work tasks. I love solving problems, I don't particularly love shaving yaks.
Apple is doing something very different. Their AI experience for end users definitely has been a little behind.
Apple Silicon, however, has been quite unique for the last 4-6 years and it's increasing overlap with LLMS.
The model/chip optimizations are definitely improvements, the thing that is really standing out the past 2 years is how much the open source model community has been making possible, especially when you know a group of use cases.
1. Most useful LLM work is done in parallel. A Mac Mini can run one LLM inference thread at a time. The cloud can spool up dozens and spread that inference across efficiently batched operations over a fleet of hardware.
2. Faster inference hardware such as the chips from Cerebras and Groq cannot be run locally. But the advantages of running >5x the token throughput per thread can’t be overstated. Add in the multi-threading advantage and it’s a knock-out punch for local LLMs.
Local inference has a role: if you’re working with extremely private matters or you want an uncapped model that will talk dirty or generate NSFW photos, local is the only option. I think Apple and others will continue to also run a lot of useful workloads locally such as text editing suggestions, speech to text, text to speech, and image manipulation. As local hardware improves, these capabilities will get better too.
But, for most LLM work, the cloud will continue to dominate for a long time to come, if not forever.
I don't want to run any workflows on someone else's computers.
That’s not accurate. With MLX, at least, parallel inference is both possible and useful. Model serving tools like LM Studio and oMLX support parallel generation with continuous batching, and the total throughput increases with it.
Can I run a few inferences in parallel on my Mac Mini? Yes. But put 1,000 Mac Minis in a datacenter serving 1,000 copies of myself? That's going to be more efficient.
There is so much happening in that scene, where tokens/sec double or 10x
So I could see the same hardware doing 20 tokens/sec on a large model suddenly doing 200 tokens/sec in the future, a better device in the future doing 500 tokens/sec, while having vision models baked in, audio models etc
Users wont consciously switch to local, they will just have it and use it
Well just because they don't sell them. Doesn't mean that will be the case forever.
you’re far far far more likely to see a camry or equivalent in an americans driveway than you are a super car.
you’re also more likely to see an enthusiast with a corvette or equivalent muscle car spending way more than it’s worth to tinker on that car in their garage than you are a super car.
I guess what I'm doing is not considered that useful then? I usually only have zero, one, or occasionally two things actively doing inference at a time, be it claude code sessions or one of the chatgpt/claude web interfaces, and i bet that's true for like 95% of people using llms. And anyway i bet even the hardcore people using a bunch of parallel agents would appreciate having access to local, private inference for some things.
You're obviously right though that cloud inference isn't going away anytime soon
And yes, I agree. I find the experience better less for speed and more for context management. But it's far from necessary.
There is probably a middle ground though. Something like a main thread agent (say a local model for Hermes) orchestrating cloud based sub agents.
Apple's shovel (ahem, Mac mini) is the highest quality.with Companies burning money left, right and center, Apple can dispense with advertising altogether
https://www.thedeepview.com/articles/how-apple-s-decade-long...