expectedwrong hindsight

Apple Releases Eight Small Language Models and Google Tells Someone to Put Glue on Their Pizza

The on-device compute bet is either the smartest play in AI or Apple just stumbling into the right position for the wrong reasons.

2 min read 294 words #apple #ai #openelm #google #on-device
hindsight — nailed it

apple's on-device strategy proved correct. AI overviews' glue-on-pizza moment became the canonical example of shipping AI too fast. the contrast between apple's careful approach and google's embarrassment held up as a case study in shipping discipline.

Google's AI Overviews launched this month and, within days, was confidently instructing people to add glue to their pizza to keep the cheese from sliding off. This is a real thing that happened. A product used by hundreds of millions of people told them to eat glue.

Apple, meanwhile, quietly dropped eight small language models — OpenELM — designed to run entirely on-device, and open-sourced the training framework alongside them. No fanfare. No pizza glue.

The thesis is straightforward: Apple's entire play is in the node. Not the cloud, not the data center — the device. The M4 chips are coming with up to 512GB of unified memory, which means at some point in the near future, the most computationally capable machine in your home will be the one running your screensaver.

That's not a metaphor. While Google is burning server farms to tell you the earth is 93 million miles from the sun (sometimes), Apple is quietly building the infrastructure to run inference locally, privately, continuously — on hardware that people already paid four grand for.

The counterargument writes itself: Apple doesn't get AI, they're late, they're playing defense, they're just packaging other people's research. And maybe. They've had hardware missteps — nobody's pretending the butterfly keyboard was a strategic masterstroke.

But "doesn't get it" requires Google to be the benchmark, and Google just told someone to put Elmer's on a Margherita. The bar is doing things on the floor.

Apple has to get it. The whole company is structured around the idea that the device is sacred — that computation should happen close to you, not in a warehouse in Iowa. OpenELM is small, quiet, and runs on your phone. That's either the point or it's nothing.

I think it's the point.