jmalicki

Are you training something so big you need that much unified RAM though?

If you can fit it on a GPU, and especially for training, it is so much quicker than a Mac.

It is a completely reasonable comparison for me as a consumer, since they're the costs and benefits that I'll actually get.

> They also have enough power to negotiate contracts with strong privacy provisions.

What privacy provisions would you want to add to AWS? Most of the reasonable strong privacy provisions you'd want are already there and/or available if you want to sign up for it, even including US Govt Top Secret data if you meet some approval.

How would this allow compile-time resolution of virtuals that was not possible before?

Gcc can already do link-time devirtualization in some cases.

Interestingly embedded 32 bit Linux is not tier 1 or 2 though :)

Sure, but I had to look:

Tier1: Windows x64/i686, Linux x64/ARM gcc, Darwin/ARM

Tier2: Linux x64/ARM w/ clang, Windows ARM, WASM, Darwin/x64

Tier3 is a pretty low support level, but tiers 1 and 2 is a pretty short list of major commercial platforms. Also even for developers to fix, availability of e.g. RiscV machines in the cloud to reproduce and fix on is still somewhat limited.

This is why I strongly like on-call.

A lot of people hate it, but if you have a rotation, that will force this person to deal with what they create during their rotations, and the operational issues will be an explicit part of your job when you're dealing with them.

When it kind of goes to random people, the conscientious people who fix the bugs and operational issues look unproductive, and the aggressive people throwing slop over the wall who never bother look great.

Forcing this through a rotating cast, and having other people not worry about it unless management decides it's too much for one person and they need help during this rotation, really avoids some of the most toxic outcomes around this.

On-call doesn't have to mean someone is getting paged at 3am. It does mean everyone rotates through, and it is clear who has the monkey, so you don't have passive-aggressive vs. conscientious behavior hiding issues like that.

There's also the executive's incentives, and "I'll be gone, you'll be gone": the payoff for that is very long, and it will be hard for them to take credit, and their own tenure might not be as long as that team's.

Nothing here says the data isn't relational. It strongly disagrees, with reasons, why it's not better represented as relational.

Personally I prefer the relational stance, and there are a lot of people who don't get it who say things like "this data isn't relational", but that's not the argument GP made.

Some of that is that EU providers need to up their game here.

Needing an EU native option is really the one and only reasonably objection I've heard against using LLMs from the cloud, the rest is tin-foil hat level unless you're actually intending to meddle with the inference or fine tuning or something beyond just querying.

I am thinking the first place this appears is on GPUs and maybe Apple with unified memory, it is not going to be on PCs.

The idea is to provide overarching guidance, but when there is devil in the details, leave that up to courts to consider when there is a real actual issue at play, rather than legislators in the ivory tower theorizing about how things might play out.

The statute gives the general intent, but with an understanding that until the shit hits the fan in an actual court case, noone will have truly considered the detailed ramifications, so leave some intentional ambiguity there for the courts to resolve on real, actual, grounded cases rather than theorizing.

And if the legislature disagrees with a courts opinion about the ambiguity it left, it is very free to clarify in statute, which will override the court! (Unless the issue is constitutional, in which case the legislature is very free to draft a constitutional amendment!)

Since when has vagueness been considered a bad thing for statutes? It's the foundation of the US (and anglo) legal system and why courts exist, to fill in. And vague statutes that can be interpreted by courts has long been preferred to precise legislated meaning that you seem to prefer.

You could live in Germany or France if you prefer a legal system more ruled by clear statute?

From the opinion:

"Because the definitions of event and occurrence do not resolve the issue before us, we must also consider the statutory context"

They talk about this and the exact definitions for basically the entire opinion. It's not just "meh, I felt like it". They are dealing with how these words have been defined in prior statutes.

The nice thing is that for GPUs you're rarely writing directly against the metal but instead pseudocode lowered by the driver; the driver could know to just lower it this way. On an NVidia GPU they control the whole stack and don't need to get a further ecosystem going there. Possibly the same for Apple.

Why would you want to read something someone was so sloppy at they didn't improve it with AI?

Yeah, that says pretty much nothing about what they do. It reads like an enterprise tech company webpage saying "we provide solutions to help transform your company into the digital age".

Even in C, if you want differently sized data to be indexable in O(1), you're stuck leaving them as pointers. You definitely could just have a variable-sized area for this, but that level of optimization is pretty seldomly done in C.

> Training a model for a specific task with the same amount of compute will outperform training a general-purpose model with the same amount of compute. This should be common sense, right?

I see arguments like this all of the time in ML. There is a large branch of statistical theory that can cast some doubt on this.

Stein's paradox: when three or more parameters are estimated simultaneously, there exist combined estimators more accurate on average (that is, having lower expected mean squared error) than any method that handles the parameters separately.

This goes deeply against "common sense", but is mathematically proven, empirically proven, etc.

If you are just taking a chess model vs. an LLM, you are right.

But if you want a chess model, a medical diagnosis model, and customer service chat model, at a high level the intuition of Stein's paradox is that one combined model will get you better performance than three task-specific models.

Most practitioners don't get this because what amounts to graduate-level stats theory is often brushed aside, and we keep relearning it the hard way.

https://en.wikipedia.org/wiki/Stein%27s_example

How ironic. The reason you can't read the comment:

"> The big milestone is when computers can read and understand information like humans do. There is a lot of work going on in this field - Google, Microsoft, Facebook, academia,... Right now computers don't know how to represent knowledge so they can't read a text book and pass a test."

Is because reddit is threatened enough by "computers that can read and understand information like humans do" the he's locked out of it!

I see it now, I misread.

Still, being able to take you laptop to a meeting to either project, or to look up related material to catch up and clarify before making a point in a meeting etc. is invaluable (no, working at a laptop during a meeting doesn't mean you're not paying attention, it often means you're cross-referencing discussions with other data live).

It's not an algorithmic issue, though, that's a systems issue. GGP was talking about the cyclic link nature which isn't an issue of size, it's an issue of convergence.

That is not at all a problem. Links dampen their effect and you run it until convergence - read the paper. It's no different than summing an infinite convergent series, $\sum_{i=0}^n a^{-n}$.

Also, WTF, it seems impossible to find a PDF of the original paper still on the web without a paywall.

Which is why, like often in these stories, there is probably more to it.

Interesting, because multiple accounts for the same person is explicitly allowed by Reddit:

https://support.reddithelp.com/hc/en-us/articles/204535759-I...

So a university should never teach a class unless at least 95% of its students would benefit from it? Yes, the one thing doesn't necessarily benefit everyone, but a university offers wide choices of interests.

You can get large under-laptop cooling pads with giant fans that can help this somewhat. It's not part of the laptop, but help if it's docked at home. It's not totally ideal since the bottom of your laptop doesn't have radiator fins, but does help.

Except for LLM inference speed, macs suck as a server and are pretty expensive, why wouldn't you get a Linux box?

You can dock your laptop at work, then dock it at home, take it with you to a meeting room at work to browse supplemental materials and/or project your screen n the meeting, and use it for the occasional trip (like using it on the train to work).

It's not like you have to be on a plane for the portability to be useful.

Why not a non-mac cheaper desktop/workstation and a Mac laptop for use?

E.g. what most major tech companies have - a laptop that is for VSCode via ssh/web browsing, and a beefy Linux dev box you ssh into for everything else?

It's way cheaper, and what I use at home too - a lot cheaper than a Mac studio for everything, especially with RAM and storage.