zimmie

hmlongco said:

newisneverenough said:

Probably this is common knowledge, but why has Apple made so much Not user upgradable? Is it to sell more new machines ? Is it to eliminate technical problems when users upgrade? Conceptually, I want more control of the thing I pay thousands for. 

One thing is that all of the RAM in onboard the SOC and shared between the CPU, GPU, and various other systems like the NE. THat's where a lot of the performance comes from in that memory is on the chip and it doesn't have to go to an external bus to get it. Further, moving something like a shader or bitmap from the CPU to the GPU is basically just passing a pointer, as opposed to copying a bunch of bytes back and forth.

Adding external RAM kills off aspect of the performance.

This is incorrect. Apple could get the same capabilities with slotted RAM. They're just using multiple DDR channels. The base M2 uses two DDR5 channels at 50 GB/s each, M2 Pro uses four channels, M2 Mac uses eight channels, and M2 Ultra uses 16 channels. For the laptops, iMac, Mac mini, and Mac Studio, the reason for soldered RAM is space. You need a separate slot per channel. Reserving space for eight SO-DIMM slots for the MacBook Pro would be impractical. Eight or 16 SO-DIMM (let alone full-size DIMM) slots on the Mac Studio would be hilarious to see, but also impractical. It would probably double the volume.

For the Mac Pro, there's no technical reason they could not have used slotted RAM. It definitely has the space and cooling capacity to spare. They simply chose not to.

newisneverenough said:

Probably this is common knowledge, but why has Apple made so much Not user upgradable? Is it to sell more new machines ? Is it to eliminate technical problems when users upgrade? Conceptually, I want more control of the thing I pay thousands for. 

The SSD in the Mac Studio actually is upgradable. Apple released the repair manual last December, and it includes details about the flash carts which make up the SSD's storage. 512 GB and 1 TB storage options use a single cart (which limits performance), while the higher capacity options use two carts. The upgrade process involves removing all installed flash carts, installing a new set, then using Apple Configurator on a separate Mac to restore the SSD controller.

There are a few reasons for integrating the SSD controller and only swapping the flash the controller uses. One big one is that it allows Apple to guarantee the security characteristics of the SSD controller. Several SSD vendors have been caught lying to computers about their encryption capabilities. The machine tells the drive to encrypt itself with a particular key, the drive says "On it!", then the drive just ignores the key and leaves the flash encrypted with the default key (typically all-zeros).

It also allows Apple to target their own performance goals rather than simply accepting other vendors' market segmentation strategies (Oh, you want a lot of random IOPS? You'll need to go with an "enterprise" drive at 3x the price and 8x the power draw.)

The RAM isn't upgradable for space reasons, as I explained above.

dewme said:

I wonder if they’ll offer standard non-prescription reading glasses style magnifier lenses? Nothing fancy, just $1.5X - $3.5X magnifiers with high quality lenses. A large number of 40-ish and older folks have presbyopia and later in life,or due to other vision complications, far field corrected  monovision as a result of cataract surgery. Cataract surgery is one of the most, if not the single most, common outpatient surgeries performed in the US. 

Magnifiers would make the problem worse. These are among the sharpest head-mounted displays ever announced, but they're still 20/40 sharpness at best.

Instead of trying to magnify an already-grainy screen, use Dynamic Type to adjust the size of text.

ARM Mac Pro, including rackmount. Nailed it, multiple times.

$1k more expensive, lower RAM ceiling, and no word on MPX compatibility (they aren't offered as options when ordering). Disappointing, but not exactly surprising. They probably only released it because they plan to use the rackmount version themselves for Xcode Cloud, as I said.

Foveated rendering? Nailed it!

Roughly 5K screen per eye yielding 20/40 vision equivalent? Nailed it!

I'm glad to hear about the lenticular array for perspective correction on the outward-facing display. That will go a long way towards making it not creepy to interact with somebody wearing a set. The synthesized avatar for FaceTime looks pretty creepy, though. We'll see how it does in reality later.

larryjw said:

When first installed, the ChatGPT app highlights the warning that it may return false information, 
The May 10 Lancet published the following article, along with detailed supplemental addendum showing the interaction with ChatGPT which anyone can then reproduce to verify the results. 
The dangers of using large language models for peer review

ChatGPT responded with totally made up material, sounding quite authoritative. 

Banning ChatGPT would seem a good idea. It would give alternative AI systems which actually can tell the truth an opportunity to be developed -- if that can be done. 

Large Language Models like GPT-4 are inherently lying machines. 

GPT isn't strictly lying, you just aren't asking it the question you think you're asking. People think they're asking the question they put in the prompt. What GPT actually does is produce something which looks like what an expert would say in response to the prompt. For example, experts cite stuff when they're asked about legal or medical matters, so the response should have citations. If there are no high-quality real citations to include, make some up which look plausible.

This behavior becomes much more obvious when you realize GPT (and Bard, and so on) will "believe" anything you tell it is at the other end of something which looks like a URL. The domain doesn't have to resolve, and the path doesn't have to be valid. An expert would load the page and potentially respond with new information based on what they read there. GPT can't actually load a URL or read, but it can fake a response which looks like what an expert who did all that might say.

People have been doing things like telling Bard that Google discontinued Bard months ago, and it responds with something like "Oh. I'm sorry, according to the link you provided, you're correct. Bard was discontinued in March."

Edit: Put "believe" in quotes. As Larryjw pointed out, the model doesn't actually know anything, so it doesn't actually believe anything. Ultimately my point is people think of GPT as being built to produce correct output, when it's actually built to produce plausibly formatted output.