"This is what's happening to the parameters of models when they're quantized down to sizes that are possible to run on your laptop. Instead of floats, small integers are what get stored and loaded into memory. When the time comes to use the quantized values, to generate an answer to a question for example, the values are dequantized on the fly. You might think this sounds slower, but we'll see later on that this actually ends up being faster as well as smaller."
I thought that most GPUs supported floating point math in these quantized formats, like they can natively do math on an float4 number (that's maybe packed, 2 float4s into a single byte, or more probably 16 float4s in an 8 byte array or maybe something even bigger)
Am I getting this wrong - is it instead the GPU pulls in the quantized numbers and then converts them back into 32-bit or 64-bit float to actually run through the ALUs on the GPU? (and the memory bandwidth savings make up for the extra work to convert them back into 32 bit numbers once you get them onto the GPU?)
Or is it some weird hybrid, like there is native support for float8 and Bfloat16, but if you want to use float2 you have to convert it to float4 or something the hardware can work with.
I am confused what actually happens in the vectorized ADD and MULT instructions in the GPU with these quantized numbers.
And every couple of days we see a post or a tweet about "Google/Apple/Microsoft just nuked my account with no notice and no recourse" so trusting them to sync passkeys rightfully makes some people nervous.
It's frustrating that there's no pricing information. The tech looks cool and all, but without knowing how much it's going to cost there's no way to really evaluate it.
The National Science Foundation has been doing this for decades, starting with the supercomputing centers in the 80s. Long before anyone talked about cloud credits, NSF has had a bunch of different programs to allocate time on supercomputers to researchers at no cost, these days mostly run out of the Office of Advanced Cyberinfrastruture. (The office name is from the early 00s) - https://new.nsf.gov/cise/oac
(To connect universities to the different supercomputing centers, the NSF funded the NSFnet network in the 80s, which was basically the backbone of the Internet in the 80s and early 90s. The supercomputing funding has really, really paid off for the USA)
"For Woody and Jayson Thomas. From the local universe to the first galaxies, the brightest moments in space and time occurred during our brief epoch together. That light is unquenchable."
She had gone back to school as an adult to study physics, she was just finishing up her undergrad when her husband and child were swept away by a wave while walking on the beach.
Are there APIs to get the iCloud sync into my own app? I'm all for iCloud syncing to my devices I just want a way to also get a backup in a file so if Apple decides to delete my account on a whim, I don't lose everything.
The fun thing about BlenderBIM is that it's IFC-native. (IFC is the 'Industrial Foundation Classes' - a data model/structure for modeling buildings and the components, systems, and intangibles like construction schedules.)
BlenderBIM is internally managing everything with the IfcOpenShell library - all of the data uses the Python interfaces of IfcOpenShell (which internally has a lot of C) to keep the model state. Blender is more a rendering backend and nice UI to manipulate the state of the IFC model with IfcOpenShell - but basically everything you can do with the Blender GUI you can pop open a shell and just type in Python and do the same thing.
This means you'll occasionally see some Blender things that don't do what you expect to the model you're editing - there are ways to have Blender do state modifications that don't all get translated to the IFC data underneath, so sometimes doing things like selections or modifiers are surprising for Blender users. (I think over time the list of things that are like this has gotten a lot smaller, and BlenderBIM is now pretty good about keeping the state of what's displayed in Blender in sync with what the underlying IFC model is storing)
The main commercial player in this space is Autodesk Revit. There is a lot of thinking that perhaps Revit has reached a point as a platform where Autodesk can't keep building on it (i.e. it has so much tech debt that it's getting hopeless) - see https://letters-to-autodesk.com/ Autodesk has a number of other 3D modeling software packages and I sometimes think that for their next generation of Revit they should consider the BlenderBIM approach and maybe build on top of Maya or one of their other offerings.
I'd even be willing to pay the $99 a year, I just want the signature to last longer than a week, ideally forever. For years when I don't feel like updating the app, I won't pay the developer membership fee.
It's not so much that OPML is the interesting part here, it's that it's a file. A few weeks back Andrej Karpathy had a twitter thread[1] about blogging software and shared this link on 'File vs App' - https://stephango.com/file-over-app - and that really was great for ecosystem interoperability. I can download the file using whatever tool is appropriate, store it however I want, and then upload it somewhere else using whatever tools is appropriate. I have the OPML export I took of my subscriptions from the day Google Reader shut down and there's still a fighting chance that other services could actually import that file.
It's also worth noting that OPML is only the container format here. Agreeing on a container format is obviously important and we won't get very far for interop if we can't even agree on the container format, but OPML is supposed to be a generic tree of 'outline' format, and conveniently RSS subscriptions (and folders) look like a tree.
I sorta expected that there would be a second standard that says "here's how you use this generic OPML container format to represent RSS feed subscriptions" but oddly that's actually included right in the OPML spec[2]. In fact RSS subscriptions are the only application format defined in OPML - there's a 'type' field defined for <outline> element and if type is set to 'rss' then there's also a required xmlUrl of the feed and optional things like the html link for the blog, the version of RSS used. This is the data and part of the spec that makes the actual subscription list exchange work.
But again the only entry for 'type' defined in the OPML spec is 'rss'. If you want to use OPML as a container for something else, like Youtube subscriptions or Twitter followers, you of course can but you gotta find some way to get everyone to agree on how to interpret the 'type' you set for that <outline> element. And as far as I know, no one's done anything like that for any other domain.
So it'd be awesome if more domains defined 'type' fields and set out some specs so I can export my video streaming subscriptions or Amazon wishlist or whatever but without defining more 'type' fields OPML is really not any more interesting than a CSV of URLs.
This sentence is doing a lot of work: "Hypothetical S2 does a bit more to simplify the layers above – it makes leadership above the log convenient with leases and fenced writes."
It'd be awesome to have a bit more transactional help from S3. You could go a long way with 'only update this object if the ETags on these other objects are still the same'. I know AWS doesn't want to turn S3 into a full database but some updates you just can't do without having a whole 2nd service running alongside to keep track of the states of your updates.
One of the long-standing challenges in the federated identity space has been that most of the solutions are built around domain names, which are a pain for most users to create on their own. There's a sense that people would prefer email addresses as their identifiers, but without some server help that's hard to do. The WebFinger protocol works well for translating email addresses into something that could be used for federated data servers, but alas most of the big email providers (ala gmail) don't participate in WebFinger.
A while back Brett Slatkin and Brad Fitzpatrick built out a protocol called 'Web FistBump' that could bring WebFinger to people who's email providers to support it. It was a clever hack with DKIM - you emailed their webfist.org server with what you wanted to be your Webfinger info, and because Gmail signed the message with DKIM anyone could verify the message. The webfist.org server just proxied WebFinger requests into lookups for those signed emails. Even better, because it's just a signed email you can treat it as a blob and have a pool of different resolvers do the proxying, kinda like a blockchain. I think there was a post from Brad somewhere that estimated that the total data needed if everyone in the world used WebFistBump for storing a blob was in the low 100s of GB, which is pretty managable for a wider community to keep online.
I need to read up on DIDs but it feels like DIDs just standardize on what the message format should be that would come back from something like WebFinger/WebFistbump, but if WebFistbump were actually up and running, it could make WebFinger more widely available. (Alas, I think the webfist.org server has been shut down but maybe they could flip it back on!)
I wish there was a good nonprofit 'infrastructure cooperative' that could provide some of these core services but have a corporate governance that could be trusted. The place I most want it for is a domain name registrar but DNS and mail servers would be good additional services.
Bluesky is a different protocol than Mastodon, but conceptually I think it's fair to say they're similar.
I think Bluesky would clean up if it could also interop with Mastodon - people want to leave Twitter but it sucks having to choose between Mastodon and Bluesky and so I think some folks are just in a holding pattern waiting to see what wins. Bluesky feels more like Twitter so I think it would get a lot of folks, but I think people hesitate to bet on it just yet.
This is a dumb question but how hard is it to train the mamba models that are on huggingface? It looks like the largest one is 2.8b - how many GPUs for how long do you need to train that up using a dataset like The Pile?
I know this isn't an actual v3 of the SSH protocol, but if there ever is a version 3 of SSH, it really needs some kind of (encrypted) SNI or at least a standardized metadata block that can be passed to any jumphost without having the know the specifics of the ProxyCommand on that middlebox.
Live data would be great, but I'd be interested even in historic data. Not like 1880s but a timeseries of 15-minute granularity of train positions for last year. It'd be interesting to see what the bottlenecks were and what other routing possibilities there could have been.
The article is about what UI toolkits are used to build content apps for TVs and TV devices like FireTV/Apple TV, but I wish for a world where using those apps weren't my only options.
It'd be nice if TVs and streaming service apps were better participants in the open-protocol smart home. I'd like it to be possible to call an API on my TV from say Home Assistant to tell my TV to turn on, switch to Netflix, and play the next episode of a show - though honestly I'd settle for just control over linear TV channels and to be able to say 'switch to channel 55' at an API level and have it work cross-vendor without having to use their native app. Maybe you can do this with CEC and stick a little raspberry PI and stick it into the TV with HDMI, but the CEC content controls seem very limited.
Apple's got some kind of a start for this since they've got a way to feed data into Siri and their internal guide, including telling the AppleTV device what's currently playing and what should be considered 'up next' but they're all private APIs, I think, and only work if the content providers sign deals with Apple directly.
Maybe a future version of the Matter standard will incorporate media playback, right now I think it's just changing volume and pausing whatever's playing on a speaker, and so a long way from being able to control what the actual content being played is.
An important part of this is to eliminate the idea that as a society we all need to have the same 2 weekend days off. (Which we don't anyway because a huge percentage of the population does work on weekends, just not most "professionals".)
Let my dentist's office be open 7 days a week, with some of them working M-Th and some working Thur-Sun and some working Sat-Tues etc. Maybe some of them mix it up and do M/T/Th/Fr or M/W/F/Su or whatever.
K-12 schools are the tough nut to crack in this scenario.
"This is what's happening to the parameters of models when they're quantized down to sizes that are possible to run on your laptop. Instead of floats, small integers are what get stored and loaded into memory. When the time comes to use the quantized values, to generate an answer to a question for example, the values are dequantized on the fly. You might think this sounds slower, but we'll see later on that this actually ends up being faster as well as smaller."
I thought that most GPUs supported floating point math in these quantized formats, like they can natively do math on an float4 number (that's maybe packed, 2 float4s into a single byte, or more probably 16 float4s in an 8 byte array or maybe something even bigger)
Am I getting this wrong - is it instead the GPU pulls in the quantized numbers and then converts them back into 32-bit or 64-bit float to actually run through the ALUs on the GPU? (and the memory bandwidth savings make up for the extra work to convert them back into 32 bit numbers once you get them onto the GPU?)
Or is it some weird hybrid, like there is native support for float8 and Bfloat16, but if you want to use float2 you have to convert it to float4 or something the hardware can work with.
I am confused what actually happens in the vectorized ADD and MULT instructions in the GPU with these quantized numbers.