undefined | Better HN

0 pointsFridgeSeal2y ago0 comments

> But ML is different and I think they should care.

It’s totally this time I promise, just like, one more ~~lane~~ model.

I’m sure they do care. I wouldn’t be surprised if they land significant support for on-app processing of models, they’ve already got the chip, dropping in local models is a sensible next step, and if close to zero effort for them.

> LLMs are going to have some general applications that users are going to want, and will become the norm

I have yet to see anyone, in my personal or professional circles, use any LLM:

- for more than a week

- for anything more than cutesy trivial things.

I’m sure there’s people around stapling models into their toaster, but this is so far from the normal.

0 comments

smoldesu2y ago

Part of Apple's problem is that they're expected to vendor support for third-party stuff. Who accelerates Pytorch or ONNX for Apple silicon, if not Apple?

They've done an okay job of that so far, but their flagship library is diverging pretty far from industry demand. At best, CoreML is a slightly funkier Tensorflow - at worst, it's a single-platform model cemetery. No matter what road they take, they have to keep investing in upstream support if they want Nvidia to feel the heat. Otherwise, it's CUDA vs CoreML which is an unwinnable fight when you're selling to datacenter customers.

I think it's possible for Apple to make everyone happy here by reducing hostilities and dedicating good work where it matters. Generally though, it feels like they're wasting resources trying to compete with Nvidia and retread the Open Source work of 3 different companies.

dylan6042y ago

> an unwinnable fight when you're selling to datacenter customers.

didn't Apple pretty much throw in the towel in this market simply by choice of form factor for their computers? The sheer desperation of their users wanting a device in this space is shown in the "creative" ways to mount their offerings in a rack.

all of the user friendly things they've done by shrinking the footprint, making them silent, etc are all things a data center does not care about. make it loud with fans to keep things cool so they can run at full load 24/7 without fear of melting down.

so from that lead alone, we can make the next assumption in that Apple doesn't care about vs CUDA. as long as they can show a chart in an over produced hype video for a new hardware announcement that has "arrows go up" as a theme, that is ALL they care about.

smoldesu2y ago

I mostly agree, which is why I question their strategy of even "competing" at all. The existence of CoreML feels strictly obligatory, as if the existence of Pytorch and Tensorflow spurned FOMO from the C-suites. It's not terrible, but it's also pretty pointless when the competing libraries do more things, faster.

Users, developers, and probably Apple too would benefit from just using the prior art. I'd go as far as to argue Apple can't thread the AI needle without embracing community contributions. The field simply moves too fast to ship "AI Siri" and call it a day.

> The sheer desperation of their users wanting a device in this space is shown in the "creative" ways to mount their offerings in a rack.

Well you and I both know that nobody is doing that to beat AWS on hosting costs. It's a novelty, and the utility beyond that is restricted to the few processes that require MacOS in some arbitrary way. If we're being honest with ourselves, any muppet with a power drill and enough 1U rails can rackmount a Mac Mini.

1 more reply

j / k navigate · click thread line to collapse

0 pointsFridgeSeal2y ago0 comments

> But ML is different and I think they should care.

It’s totally this time I promise, just like, one more ~~lane~~ model.

> LLMs are going to have some general applications that users are going to want, and will become the norm

I have yet to see anyone, in my personal or professional circles, use any LLM:

- for more than a week

- for anything more than cutesy trivial things.

I’m sure there’s people around stapling models into their toaster, but this is so far from the normal.

0 comments

smoldesu2y ago

Part of Apple's problem is that they're expected to vendor support for third-party stuff. Who accelerates Pytorch or ONNX for Apple silicon, if not Apple?

dylan6042y ago

> an unwinnable fight when you're selling to datacenter customers.

smoldesu2y ago

> The sheer desperation of their users wanting a device in this space is shown in the "creative" ways to mount their offerings in a rack.

1 more reply

j / k navigate · click thread line to collapse