except that he's making the common mistake of confusing software engineers with framework operators.

Agile is kind of a late-stage communism implementation. It's easy to figure out why management has such a hard-on when it comes to the subject, but its huge popularity among (even senior) developers is just bewildering. From what I've observed over the years, it's always the same buch of people: Those that don't care much, don't achieve much and (most importantly) don't like to take any responsibility.

Agile takes the whole "team" idea to a completely new level where individuals simply cease to exist, just like a really stupid version of a hive mind. And so, it's the perfect framework for indolent bureaucrats who prefer self-inflicted micromanagement over having to bear any personal accountability for what they do. Every failure as well as every achievement falls back on "the team", so people are essentially shielded from any individual responsibility.

This! Maybe the pricing is quite a bit off here, but isn't that exactly what so many people having been calling for? If you're a user of their platform, Meta is going to expect some kind of compensation in return for the service it provides. Network effects or not, Facebook isn't exactly the only communication channel out there, so I don't buy into all those "free communication is a right" kinds of arguments. At the end of the day, they're a company and they won't exactly settle on making money by pulling fairy dust out of thin air.

Both the github and pypi repo appear to be completely detached from Apple as a company. There's at least one of their ML engineers involved, but that's about it.

What isn't being said here is the fact that Lattner's also trying to put the whole AI thing on solid grounds engineering-wise. The Python AI software landscape is just an unprofessionally lumped together hot mess right now and I welcome everyone trying to change things for the better.

I guess that would be fine, but why on earth does the sticker price include a music streaming service that I'm not going to use ever?

Indeed! Only very few people will. But this is Europe, so it's not completely unreasonable to envision a path where the cellular and landline providers pick up this model and start negotiating deals with Meta. In such a scenario, they'd expect a competitive edge from offering a less intrusive and ad-free access to Meta's platforms as part of their package. They were already doing something similar with their zero rating models, but those have shot down by the courts due to very obvious net neutrality issues.

So much of what's broken with today's internet is caused by the user-platform misalignment of interest that comes with the current ad-based business model, so I'm really glad to hear that someone is finally starting to at least try out different options. Having said that, this might as well just be a smokescreen that allows Meta to make a point to regulators around the world regarding user choices.

Let's look at this from the opposite point of view: Pytorch is the actual joke here, offering not much more than entry level programmer friendly data transformation pipelines on top of a haphazardly coded bridge between tensors and whatever CUDA has to offer this week. Other GPU architectures don't seem to get much attention at all and their support comes down to a series of badly maintained hacks that barely make things work (hence your frustration with MPS I guess). On the other hand, Apple has a rock-solid track record when it comes to supplying APIs and toolchains for their own Stack, so there's reason to expect something pretty decent once they start rolling out LLM functionalities for their own devices. As for things like FP16 support, it's still quite debatable whether this is a good idea for training in the first place (I'm sure there will be some kind of quanitzation for inference). Having said that, even with Apple's insane memory pricing, you get MUCH more RAM per buck compared to Nvidia cards, so whether you're using FP16, FP32 or FP64 becomes much less important.

the whole premise of the article is wrong. The great news about llama2 wasn't the "somewhat improved model" part, but rather the fact that it came with a new license that absolutely permits commercial use, no (unreasonable) questions asked. At least as long as you have less than 700M monthly active users, which in itself is a quite hilarious way of excluding just a handful of companies that should either have the standing to build their own model or to properly license one.

It may as well be the exact opposite: They WANT to be regulated, so the barrier of entry is raised. All that talk about safety is just there so they get regulations that require a degree of manpower and/or money that only they can bring to the table. Incidentally, Meta (who happens to have just released llama2 with a very liberal license for private AND commercial usage) doesn't seem to be part of this.

Maybe, just maybe, looking for an excuse for not having to learn how to code is all the reasons you need for not doing it. The huge crowd who got into coding just because they were told it pays well really is the bane of the industry

Not true. You can absolutely get it on Amazon Video, even from the old continent

Not sure modern Java still deserves this amount of hate these days. The language and its ecosystem are quite fantastic. Usually, when people complain about Java being shit, they're actually talking about the development culture that most Java shops have adopted. That one is completely beyond repair.

Oh, that's easy: Please remember that the end game here is not about compression. It's about gaining insights on how AGIs might actually work. Admittedly, there are some domains where NNs create marvellous results, but in essence, they're just very powerful parameter fitting machines for preconceived models. You don't gain any insights from that. Secondly, there's not a single instance where this approach has been demonstrated to provide anything that might be related to real semantics.