What will happen once people are priced out of buying computers? I find this worrying.

I can't justify a more than 2x price increase for what I'm guessing are marginal improvements. I want to replace my home lab but as it stands my underpowered NAS will have to continue as it has the past few years

However the main reason I would upgrade is to run LLMs locally and RAM prices are making this a no-go. One day if RAM comes down to earth I'll upgrade. If not I'll stick to this old machine for a long time.

I just wait for practical real live numbers with real stacks i use.

And btw. they reduced the memory bandwdith on the M4 in comparision to its previous version by 40gb/sec or so, you need to wait to see.

Says 1.2x for multithreaded and doesn't specify (I presume smaller) for single-threaded. That's compared to M5 not M5 Pro, so M5 Pro is likely significantly faster.

For example: it is copy-on-write, and my corruption happened in the superblock checkpoint. There is no built-in command in Disk Utility to "mount" an older checkpoint (and accepting a small amount of dataloss). Also the "First Aid" command allegedly runs volume repair top-down when it is recommended to do it inside-out.

[1] https://www.ntfs.com/apfs-structure.htm

We should be buying European solutions anyway. Different psychological barrier that was broken long before.

I mean, local LLMs already are really good if you own Nvidia Cards, not sure what you are waiting for. 2x3090 will blow any current Mac out of the water in tok/sec

170GB/s memory bandwidth Configurable up to 32GB unified memory

307GB/s memory bandwidth Configurable up to 64GB unified memory

That is also the factor you want to take into account for AI inference, of course.

Not like Mac gaming is good anyway...they have "Game Porting Toolkit" but it's not free and open like Proton so nobody will use it unless Apple pushes them. Removing Rosetta doesn't help either.

They include game metrics in their marketing often, because they know their customers care about gaming. But no—for whatever reason, it never translates into Apple caring about gaming. I don't get it, but that's the way it has been for a long time.

IMO memory size is the most important since you can always wait longer for it to compute while a blazing fast but dumb model is much less useful.

I paid $499 for my M4 back in April with the academic discount since I predicted the next generation base model would cost more. Sadly my prediction was correct: $599 to $899 is steep. I understand why; the prices of DIMMs, GPUs, and SSDs have also skyrocketed. But $899 for a base Mac Mini is painful, even if the alternatives on the current market are also painfully expensive.

The m6 with same specs are $2000.

Also m4 Mac mini seem to now be in stock on their refurbished store.

It depends on what model(s) you want to run.

I went with a 64GB M2 Max a few years back even though it stretched my budget and I'm glad I did. I'm not sure what I'd pick in this case. Tough call.

If it persists for years people will have time to spin up more fabs and meet demand

My daily driver personal machines are incredible pieces of kit, but everything is always getting worse because majority of software is some SaaS living in the cloud, or even if on-desktop is constantly updating slop.

Why does it take 10 seconds for my streaming music app to open when my machine is 10-20 cores, 32GB+ RAM, high bandwidth NVMe storage, and I am connected to gigabit fiber internet? The thing is autoplaying videos, and has about 10 different panes clearly owned by different product managers that hate each other. Half the time its pushing me some pop-up dialog I don't care about.

Winamp in 1998 took about 2 seconds to load.

I’m sure some kind capitalist will be happy to rent them a thin client linked to a computer in the cloud.

Same with my M1 Air for being on the move, even notebooks would be more expensive than they’re worth.

What happens is that people start buying Chinese mini PCs, some of those even have spyware preinstalled along with another malware that is Windows 11

We've been through.. 4? 5? memory boom-bust-boom cycles already. Nothing is pointing to this one being any different, aside it probably becoming the mother of all busts since AI isn't adding much productivity and as businesses realize this, they'll stop paying top dollar for access meaning AI vendors cancel all their reserved fab capacity.

The fabs already exist at that point so it's less of a loss for the memory fabs to crank out memory chips and have the RAM OEMs sell them at bargain bin prices. No pseudo-cartel pricing possible either with CMXT becoming a major player. $50 for 32GB-64GB of DDR5 in 2029 here we come babyy

It will last longer and hopefully the GPU prices will recover in a few years.

I hope

I'm not sure I'd consider a gaming PC 'regular consumer' hardware especially with today's prices. But yeah, it's definitely doable.

I don't know what this means - game porting toolkit is free

They’ve also said when they pull Rosetta for intel compatibility, it will still be available for old games to use, which is a huge improvement from their usual stance of “that game studio that hasn’t existed since 2008 should dust off their code and update it to our new architecture.”

Most people simply don't care about using well-made software so long as it gets the job done because they don't even look at other options. I suppose this is how carpenters feel when they see ubiquitous IKEA furniture.

I have an old M1 Max with 32GB that runs inference models fine, especially those using MLX.

Dedicated graphics cards are faster but not to the point that it matters.

That’s a six year old machine.

My newer machine M5 max 128GB will far outperform your typical 32GB gaming card once the model exceeds memory.

To even get close to that on a dedicated graphics card you are paying upwards of $20K.

That said, models are getting smaller and faster which allows me to run multiple different models with ease.

[edit] Just checked the 5 models I use take up 57GB when all loaded at the same time.

/s

My son is using my old ~2011 Mac mini with Linux Mint on it. Slow but works.

- Lots of options competing on quality, price, user experience, user needs

- Every marketplace innovation that we talk about on this platform

So somebody might offer exactly that and if it's a good option, people will use it. If it's not, they'll use an alternative.

Government regulation will lead to:

- Giant entrenched companies that are difficult to disrupt thanks to insider connections

- Lack of competition, price increases, terrible customer service and pretty much everything that "anti-capitalists" today complain about in regards to capitalism.

Today development work is limited by compile times and test run times, particularly with UI tests in the iOS simulator.

Hardware on the development machine finally maters.

Maybe we'll see a return to the late 1990's, when companies would give you a free computer as long as you allowed it to constantly display an ad banner on the top of the screen.

Would love to know where this comes from. I am not sure about that. Could you point to data/anecdata?

no one remembers the last time things got better rather than worse. shitcoin-driven scarcity seamlessly transitioned into COVID-driven scarcity and now into hyperscaler-driven scarcity. the chipmakers' cabal decided to meet increased demand with four digit margins rather than expanding production.

To be honest though, my usage does not benefit from the speed. The only benefit is to the reptile part of my brain that says “speed. Must have the fastest possible, and since I’m paying for it, even though maybe I should downgrade it, I want my speed.”

https://www.imec-int.com/en

Don't give them ideas...

It definitely matters, if you're intending to run a Claude Code/OpenCode style agent workflow. Most of those harnesses start with 8-12k token contexts, which is a lot of prefill for a Mac but cheap for a CUDA GPU. The GPGPU compute on the fastest Macs is still trailing behind Nvidia's laptop GPUs; the highest-bandwidth Apple Silicon chip (now the M5 Ultra) has ~7x lower memory bandwidth than a single B100 card.

There's a good reason why Apple Silicon isn't to be found anywhere in the datacenter buildout. It's nonviable for training, and wastes electricity running real-world inference workloads.

The game gets a bad rap for the last-gen console performance, but even 2020's iGPUs could handle it: https://youtu.be/oUm5GmJcBPA

(250,000 pre-orders, vs now ~5000 sales/quarter)

I'm usually running about ~4 instances of it locally in separate Git worktrees, which (in our setup) means I'm also running 4 copies of the whole stack: web, frontend, Redis, Postgres, etc. These are rebuilt and torn down multiple times per day. Lot of frontend assets to compile.

Additionally, like many people in corporate environments, we do of course also pay a performance tax thanks to the security software we need to run.

So perf does matter for us.

We could, of course, alternatively... just run cheap shitboxes locally and have all our dev environments in the cloud. Certainly appealing in many ways from a cost perspective and you don't even need to give up DX really.

However, in the next ~5 years, as "actually useful" local LLM inference starts to become genuinely useful (for software engineering, this seems to roughly mean "the ability to run 28B+ models at usable speeds") then having local compute horsepower might start to look more attractive again. Even if it's not an absolute winner on cost/utility alone, the data privacy and cost certainty have value to many.

Companies will prioritize components for cloud computing and lease us thin clients...

Wait for the product to be publicly available if you want in depth testing.

Leasing not only allows Apple to keep "selling" people new devices, it will also mostly like accelerate the second-hand sales and resource recycling arms of the business. Consumers will be more willing to upgrade sooner and stay in perpetual payments vs having to buy out the device at lease end - and Apple gets back newer devices where the recycled tech is arguably more valuable.

I feel like a big part of macos slowing down is installing too much crap over the years that sits in the background.

Macs, as nice as the hardware is, are hostile to repair.

Then they sold out and most recently literally sold out to this software company from my neck of the woods. Every single article now feels like it's written by an LLM optimized to write like a crypto/AI bro.

I used to use it daily and now not at all.

Oh, you mean like Apple, Google, and Microsoft?

> - Lack of competition, price increases, terrible customer service and pretty much everything that "anti-capitalists" today complain about in regards to capitalism.

Have you looked around? This is already happening.

Literally everything listed for 'government regulation will lead to' can and has happened in unchecked capitalism. Insider connections rule everything right now in this current regime of chest-beating 'capitalists.'

Capitalism and socialism aren't real things outside of freshman year econ textbooks.

There's been a few studies. Most find productivity gain from 'AI' to be negligible.

It's Deja vu for NVMe and SSD drives this time as prices are way up there.

Particularly with Chinese companies pushing to get more out of smaller models. You can run really capable models on the top end macs right now, it just costs way more than buying credits.

Personal assistant (e.g. openclaw) in the mac ecosystem. You need external models to power it ofc, but cool to have all the Apple integration and local network in your assistant. The claw can be addressed in multiple ways and can command your local hardware if you so wish.

If you are running Steam in the background, one of the buttons triggers an HUD keyboard, where the left trackpad types keys on the left side of the keyboard and the right trackpad for the right side of the keyboard. My guess is the new controller works the same way.

Also, some media center software, such as Kodi, have gamepad support. So a gamepad will feel completely natural to control it without having to resort to using the mouse and HUD keyboard.

https://www.economist.com/business/2026/07/27/america-has-be...

And it leads to comments like the parent blaming "capitalism"

But yes, they all went out of business during when the dot-com bubble burst.

“Results are compared to previous-generation Mac mini systems with Apple M4, 10-core CPU, 10-core GPU, 32GB of unified memory, and 2TB SSD.”

Everything government involvement touches gets more expensive. See healthcare, college tuition, student loan debt as some easy examples.

The one notable exception is if you're a group of workers and you negotiate a contract with your employer: the state can and will interfere to void it if it prevents the workers from being properly exploited.

This goes back before AI, with computers not being found to be improving productivity numbers: Solow's paradox.

* https://en.wikipedia.org/wiki/Productivity_paradox

* https://www.brookings.edu/articles/the-solow-productivity-pa...

M1 Max: 400 GB/s memory bandwidth, ~20.8 FP16 Tflops M4 Pro: 273 GB/s memory bandwidth ~17.2 FP16 Tflops

One of the problems here is that things like this only count as "increased productivity" in the economic sense if it results in more income at the end of the funnel. Are you actually making more product faster in all your new-found free time?

Very few orgs I've worked in have actually been limited by the speed of programmers writing code. Usually it's communication overhead and management processes eating up most of the bandwidth - and those tend to expand to absorb any additional bandwidth.

Medicare is cheaper that US private insurance:

* https://www.kff.org/medicare/how-much-more-than-medicare-do-...

Also:

> A single-payer universal health care system could cover every American, save more than 100,000 lives a year, and still cost $1 trillion less than the system it would replace, according to a new preprint study led by researchers at the Yale School of Public Health.

* https://ysph.yale.edu/news-article/universal-health-coverage...

This strikes me as a problem with how we define productivity in this context. If I can do my job in 1 hour, when it took me 8 before, am I not more productive in a rate-based sense? Sure, I might only work 5 hours in a week instead of 40, so my net production is the same, but my rate of productivity was certainly much higher.

It is big. No dashboard. square steering wheel without turn signals or gear stalk. The rear-view mirror doesn't see out unless you lower the cover in the back. The side mirrors are strangely shaped and small. The front corners are a mysterious mostly unpredictable distance out the front window.

But then... It has cameras in all directions including the front bumper. There is a top-down sort of heat-map display when you park. All 4 wheels can be turned for a tighter turning circle. and it can self-park.

So... I would feel embarrassed to be seen in it, but it was definitely an interesting experience.

I may need to replace some parts at some point, but this far it's not been necessary.

I'd love to have the 16", which didn't exist when I bought the 13", but thanks to Altman, hardware prices are insane these days. :/

With larger (also really fast) unified memory on the chip, it could easily load larger weights, and the thing is that the quantized results on the M5s are really good, I believe that for most local inferencing, users would be using INT4 (maybe more bits per weight sometimes) quantized models, that might be where the Neural Accelerators kick in. In raw power, a desktop 5090 easily outperforms the M5 Max, but for this specific use case, I believe that the M5 Max is good enough.