Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
WWDC 2026 to Showcase Apple's 'AI Advancements'
- Thread starter MacRumors
- Start date
- Sort by reaction score
Hey Apple, how about an Studio M5 Extreme with 160 GPU cores and 1tb of ram to kick some Nvidia butt?
That's AI advancement.
That's AI advancement.
mattwolfmatt
macrumors 65816
Mostly Adequate
macrumors newbie
SalisburySam
macrumors 65816
Any announcement about Siri has lost all credibility for me. The only thing I pay any attention to now is what Siri has delivered. And so far I’m not at all impressed. Based on past performance, I expect little from WWDC re: Siri.
obviouslogic
macrumors 6502
Yeah not true. LLM does not encompass all of AI. It processes text and based on statistics responds accordingly. This is extremely useful in conversational interactions… Such as a front end for Siri. Before spoken text gets to the LLM, it has to be converted to text (speech recognition) and the resulting text back to speech (speech synthesis). Those two things are also part for what is called “AI”. Apple’s devices already do both, no Google needed. There are many others areas of “AI” that Apple’s devices can already do, and have done for years, without the need for Google.
obviouslogic
macrumors 6502
You do know Apple Intelligence features already exist and are currently run on devices and have been for well over a year now? The only thing that’s missing is the revamped Siri.
avkills
macrumors 68000
Better yet, how about a M5 (Max, Extreme) Mac Pro with PCI slots that can house an Apple GPU card with a lot more than 160 GPU cores. 1TB RAM would be very good.
Hell I would not even care if Apple made their own proprietary slot for their GPU card; as long as there were also some regular PCIe slots for storage and capture cards.
Ok I just woke up.
You may not have. But some people have been doing truly astonishing things. Here's an example.
GitHub - danveloper/flash-moe: Running a big model on a small laptop
Running a big model on a small laptop. Contribute to danveloper/flash-moe development by creating an account on GitHub.
github.com
The goal is to run a large model on a device with a small amount of DRAM. The basic idea, suggested by Apple a year ago, is to stream the weights in from flash storage, so that the weights required every step keep being swapped in from flash rather than all weights persisting in RAM.
In about a day, after deciding to tackle this seriously this developer (with massive help from Claude Code) was able to put together a framework to try this idea and then run a sequence of successive experiments to optimize the idea. This would probably have been maybe three months of work a few years ago.
(And it doesn't end there. Various humans on twitter have suggested experiments that others, again AI-augmented have tried and are still trying, some have been pushing harder, like running a 400B model on an iPhone - right now at about 1 tok/sec, but proof of concept, etc).
It's simply no longer true to say that this stuff is either worthless or just glorified google. People who have ideas and *rapidly* want to explore them are being able to do so 100x faster than before.
