edude03

"yes" but that's a philosophical question. I think they're getting better at "early fusion" IE, training the model that "apple" and these visual tokens are the the same concept, but LLMs are fundamentally a pattern matching machine so even with perfect fusion I personally wouldn't call it understanding.

> Cross out stuff you disagree with, draw a big circles around the stuff that resonates. > Not only does it make consumption a lot more engaging, it makes a revisit of that book extremely rewarding

This is actually the biggest reason I own a Remarkable 2/Kindle Scribe, let's me "destroy" the book while also not having to worry about losing my notes.

You might be taking crazy pills or you've yet to come across an app that's just a webpage inside a web view (a lot of apps these days unfortunately)

If phones didn't have secure boot nowadays inevitably little Johnny would (unknowingly) install a rootkit on his mom's phone for the promise of free vbucks or some see through walls app.

I didn’t make the point I was trying to if

> Is multiplied by each "system" you add

Was still the take away. The idea is the lost a single instance should cause downtown so it’s not multiplied. IE three instances of the database case tolerate a loss of a single instance for maintenance since the other two will take over the load.

There is of course an argument against the cost of this and sure I’d even accept the complexity argument since you probably need to add more tooling to manage the hand over but again the point of this complexity is specifically to avoid having a single point of failure / to naturally handle failure such that the possibility of downtime isn’t multiplied

As always it depends but at least for me in my anecdotal experience as a Distributed systems engineer (I know, I'm clearly very biased here) - the problem with a single system is that mundane things cause unavailability like - software updates, hardware upgrades/failures, having a single fault domain (IE, someone writes a bad query that wedges the DB, now no one can use it) and since failure isn't built into the design, _when_ things fail the mean time to recovery tends to be rather high. For example, if the power supply explodes and takes out your server, how long would it take to procure a new server and restore from backup? On the flip side, if your system is designed around (for example) server less functions or spot instances- which become unavailable multiple times per day, you would have already engineered in fast recovery

A fair callout, though yes it was supposed to be a simplified analogy. I think you got it exactly right though that its temporary in the same that I'd argue the benefits of using AI are temporary

The article states:

> If you are on a current iPhone or Mac

Presumably if you don't trust apple you wouldn't purchase their products and even if you were for example forced to use it via work or something you wouldn't use this feature ... so it doesn't really change the calculus as presented by this article - IF you ALREADY HAVE a MODERN Mac (and trust apple) this is your best option

You likely don't need to disassemble the inference code, the weights are "just an array of numbers" in MLX format.

It's not open weight, but the point is to be an on device (and thus local, privacy preserving) option. The article mentions that as the caveat

> What this means if you just want good transcription

> If you are on a current iPhone or Mac, the best on-device transcription engine for English is already in the operating system, and the private option is no longer the compromise option

Not to "glaze" the author as the kids say but this has to be one of the best written musings I've read on HN in a long time. I'm likely bias because it's written in "my style" but I feel like it's a rather fair and balanced approach to a nuanced and socially difficult topic.

I liken it to dieting. If your only goal is to be certain weight, then learning how to cook, learning how to portion, learning to make a meal plan, learning about macro nutrients is all "friction" now that we have GLPs. And maybe using a meal prep service reduces some "unnecessary friction" but you still would have to learn a bunch of useful skills along the way.

Concretely, If your only goal is to produce "software" then learning about design, planning, project management, testing etc is all unnecessary friction when you can just ask an LLM to "make it so"

Other than batch jobs, I can't think of a problem that can be solved these days that doesn't also require high availability - at the very least they require a warm standby.

I wonder why they're removing support for encryption when clearly they have the code for it and still supporting the actual FS

I'm using "sneaky" here to refer to anything that's not very obviously stated but anyway

> That their actions make sense for their business isn't any reason for people to accept their deceitful, customer-hostile decisions.

While I agree it's a dangerous precedence to set, I think this is a "vote with your wallet" sort of situation. They shouldn't do it, but from their POV this is what they need to do to offer the product they do at the price they do. If the product wasn't compelling people wouldn't accept that they do this. However they've decided if you want their product you have to use their interface and whatever spyware it comes with, so it comes down to, is the value proposition good enough that people will put up with it? As of today, the answer is unfortunately yes

> maybe you don't understand this hypothetical situation

> I'm suspecting you just don't care about other people's privacy.

Quite a leap to assume I have neither basic reading comprehension skills nor care for privacy, but assuming I'm just misunderstanding you - I think this is the fundamental disconnect between security and privacy.

For one, most of this data is already collected openly by most apps and sites on the internet in countries all over the world, they just call it "analytics" and preventing tools like ublock from blocking them is an ongoing cat and mouse game.

Secondly - as someone who buys a bunch of electronics from companies headquartered in china (DJI, Insta360, Roborock immediately come to mind) they already have both normal analytics like in point one, and anti tampering/ anti forfeiting / anti reverse engineering features that are at least as, but often more, invasive than this.

Thirdly, and probably most importantly - as the author states, you're using a tool that by design and to be effective, uploads your private data to a third party for processing. You use it knowing that once the API request is made you have no idea what's going to happen to that data and this again is just fundamental to how (cloud hosted) LLMs work - the only privacy preserving option is to run your own LLMs at home or remotely on hardware you control

Let’s see how long until opus 5 comes out but to me this lends some credence to the rumour that fable/mythos was supposed to be opus 5

I don't understand the privacy concerns the author is trying to highlight. Granted, doing anything "sneaky" will always raise suspicious once caught, but on the other hand, there would be no point in implementing these "security features" if they were upfront about how they work.

And no, IMO stenography isn't security by obscurity, in the same that using RSA and keeping the private key private isn't security by obscurity - keeping the private thing private is part of the security model.

We can't make single cores any faster, so realistically multicore is the only "solution". That said, most languages have M:N event loops, where M tasks are distributed across N OS threads so even if your software doesn't directly use multiple cores, you end up using them indirectly for example for IO to the database or other APIs.

Could have sworn the author was a nix(os) user already. I know it’s a meme but what all the problems they’re describing literally is solved by nix. The nix sandbox even catches calls for time for example to replace it with 0 for determinism.

I think it can't be improved because it's measuring the wrong thing. A junior engineer becomes a senior when they stop being told what code to write and start solving business needs. Therefore often the highest paid engineers aren't the ones who would do the best on leetcode - or SWE bench pro verified.

Maybe AGI is possible and we'll have software defined human intelligence that's completely autonomous but that's not coming in the next slightly better RL trained LLM and if existed likely wouldn't be under our control anyway

Googles machine translation team wrote the Attention is all you need paper that introduced transformers specifically to solve the problem that you can just model language by mapping one word to another. I'd be floored if they weren't using the tech they invented for intended purpose

Almost the exact same thing happened to me when I first tried opus, one prompt no output cost $60 in additional usage

I think it's because they're running out of ideas too BUT the current generations of foldables (galaxy fold 7 for example) are essentially indistinguishable from non folding phones when closed. Yes, that means they could have made a thinner phone over all - the Galaxy Fold is the same thickness as the iPhone 17 pro max but both are twice as thick as the air - but I think consumers have gotten use to thick heavy phones - its why the SE and air don't sell as well IMO

I should blog about it but it's essentially two things

1) "A lot" of nodes, (1 42U rack is one cluster, with battery backup and redundant switching)

2) Hybrid cloud, a few nodes of this particular cluster run in GCP (kind of cheating :P)

It does, you need to drain the node before removing it otherwise kube assumes the node will come back

Rescheduler is impractical because scheduling is environment specific. You might have for example a database that needs three nodes and you have three servers, there's no where to reschedule those pods to in that case.

In the cloud you can use cluster autoscaler or karpenter to automatically handle the unhomed pods however.

I'm not personally trying to gatekeep kubernetes, everyone should do what works for them. However, if I'm putting my professional credibility and/or my sleep schedule on the line, I would not advise anyone to do this.

Even at home, I run stuff that needs to be highly available enough that I wouldn't go this route when there are better options.

Insert the "No god no" meme here - you really shouldn't be updating nodes in place and thus shouldn't be restarting nodes.

I'm aware bare metal exists and it's not always practical to just provision more servers, yet I think for most workloads you're not getting the benefit of Kubernetes if you have say 3 servers and lose 1/3 of your capacity to do software updates.

It makes sense when you consider LLMs don't generalize very well, so they're heavily dependent on how good (how varied as well as how high quality) the training data is