cbsmith

I think there are lots of advantages to running a model locally. Saving money is one of them, but that's only if you can keep the thing busy. You wisely put the "free" in quotes for a reason: you paid money for the hardware the model is running on, and you're paying for the electrical bill to power it too. Even if you pay a 100% markup to the cloud, unless you're keeping it busy 50% of the time, it's cheaper to rent.

So the idea is that it SHOULD cost OpenAI a trillion dollars to do what you can accomplish with a potato?

So if your AI compute needs are handled by an Apple TV, I'd be really curious how those same needs served by the cloud work out to a trillion dollars.

I didn't like how this essay misunderstood the design principles in Java's class files. With dynamic binding to the runtime, you can't know for certain the layout of a data structure in memory (e.g. what is the ideal memory alignment?). If the class file is untrusted, you can't even be sure you have a valid data structure in the first place. So allocating the array and then assigning elements one at a time is what you do.

> Nobody wants to pay for a trillion dollar cloud bill.

Buying dedicated hardware as a way to keep your AI bill down seems like a tough proposition for your average consumer. Unless you're using AI constantly, renting AI capacity when you need it is just going to be cheaper. The win with the on-device model is you don't have to go out to the network in the first place.

s/providers/insurers/

As an old Perl programmer, I knew immediately what the /o would do. ;-)

Is there a good way to lay off 150 people?

Yeah, but you need energy to desalinate so...

"sharding party"

Man, that made me laugh. I'm using that.

> We do get compared to Elasticsearch a lot. While we support some of its APIs, Manticore isn't a drop-in replacement.

Thank you for saying that up front. I read a description of your product and the first thing I thought was, "this looks like a potential alternative to ElasticSearch, but it is not a drop-in replacement for ElasticSearch".

Yeah, this title for the article is really terrible. The "why" that scientists are investigating is not why many lung cancers aren't in nonsmokers. The "why" they are investigating is "why are these non smokers getting cancer?". Once smoking stops being such a dominant cause, you put more energy into the other cases.

> What you’re describing is incidental complexity. It is not a good thing.

Yes.

/me feigns shock

> Really? Perhaps read the 'Principle of Least Astonishment' [1] to see why this is a problem. If I create a new object I would expect the 'init time' properties to be initialised.

The Principle of Least Astonishment definitely applies. It's a deliberate design choice that unfortunately violates the principle.

> Great insight. Let's not run the 'init time' properties for a newly initialised object, just in case it works as expected. This 'feature' can't even be manually resolved by doing post-`with` updates (because often the properties are init/read-only). It makes the whole init-property feature brittle as fuck.

? I'm not sure I follow what you are going with here, but yeah, in general you'd have to carefully limit all uses of "with" for objects with precomputed values to inside the encapsulation of said objects. Alternatively, as you mentioned, you could just not have precomputed properties.

> It isn't, that's why there's a blog article documenting how unexpected it is.

The behaviour is expected for the language design. Whether developers using the language expect it is a separate matter.

The with operator clearly allows someone to break encapsulation and as such should only be used in cases where you aren't expecting encapsulation of the underlying record.

> It's probably worth reading the article:

> "Note that because Value is set after the cloning operation, we couldn’t write a copy constructor to do the right thing here anyway."

It's probably worth reading the entire article, as that quote is followed by: "(At least, not in any sort of straightforward way – I’ll mention a convoluted approach later.)", which presumably was what was being referred to there.

In general, there's a whole ton of gotchyas around encapsulation of precomputed values. That's just life outside of a purely functional programming context.

Yeah, there's a long and tangled history with UNESCO that spans multiple administrations.

That said, the wording of the statement is... problematic.

Yeah, I remember using untagged unions because the code "knew" which type was the right one to use at the right time and you wanted to save space (and wanted your structures to fit well in registers --and later cache lines).

Just a wee bit.

I'm pretty sure when they got the gold medal they weren't allowed to judge themselves.

As guy named Chris Smith, I really appreciated this story.

Pretty sure Jon Udell remembers XUL.

So what if it is a fiber channel over ethernet drive?

I got bad news for you: they're always unreliable.

I've got a setup where the network filesystem is more reliable than the local filesystem. The local filesystem is running on a RAID-0 with fiber channel over ethernet and the network storage is a RAID-5 of a bunch of RAM block devices all linked over infiniband.

Your typical SSD these days is effectively a very complex distributed RAID, but it pretends to be a very simple local disk... and that's what most local filesystems are running on.

Honestly, NFS was designed to run over UDP originally. Failure was totally an option. When I was in school, the NFS filesystem was way more reliable than the local filesystems on the same computer. Yes, distributed filesystems are hard, but... NFS is comparatively straightforward for a lot of what passes for "reliable" these days.

I love the irony.

PGDv6 is still closed source, yeah?

Except the Windsurf team is already moving in that direction...

Canada is "much more socialist" in that it has socialized medical insurance. Aside from that, it's maybe a tiny bit more socialist, though one could argue it's not more socialist at all.

The systems are different, but saying they are completely different is really a stretch. There's a GST that the US doesn't have, which is, ironically, a regressive tax. If you ranked the tax code of countries by similarity to the US tax code, I'm not sure Canada would be at the top of the list, but it wouldn't be that far down.

I still don't understand equality in JavaScript. ;-)

That's an amusing perspective. I really struggle with watching any video at double speed, but I've never had trouble listening to any of his talks at 1x. To me, he seems to speak at a perfectly reasonable pace.

It seems counter intuitive that the reader would no longer have that authorial instinct due to lack of writing. Like, maybe they never had it, in which case, yes. But being exposed to a lot of different "writing opinions" tends to hone your own.

Let's say you're right though, and you lose that authorial instinct. If you've got five different proposals/PRs from five different models, each one critiqued by the other four, the needs for authorial instinct diminish significantly.