HNHacker News
TopNewBestAskShowJobs

darknoon

682 karma · joined May 12, 2012

deep learning, GANs, art, graphics ML for creativity at diagram
submissionscomments
darknoon··on Sam Altman is still trying to return as OpenAI CEO
Whisper v3, just a couple weeks ago https://huggingface.co/openai/whisper-large-v3
darknoon··on DALL·E 3
Stable Diffusion + fine-tuning is quite powerful for creating specific art styles that aren't easily described to DALL-E / MJ
darknoon··on OpenAI GPT-3.5-turbo-instruct released
It's a completion model, which can work a lot better for certain use cases.

Eg I can ask it to generate a huge chunk of code and it doesn't try to give an "example", it generates a realistically long bit of code.

This is great particularly when you want to generate an entire webpage, vs having a chat with an agent that's going to tell you how to build a webpage yourself with small snippets of example code.

darknoon··on Meta announces its Quest 3 VR headset
They did a price drop on it, seems like it will stick around until they can make the Quest 3 a more consumer-friendly price.
darknoon··on Numbers every LLM developer should know
you can access the tensor cores from cuda, in practice you might generate the code with something like openai triton
darknoon··on Numbers every LLM developer should know
afaict OpenAI's instance is massively overloaded, you can see with the 32k context model actually being faster in practice rather than slower
darknoon··on Rare corn can self-fertilise [video]
In the video it said something about modulating the amount of nitrogen fixation based on the plant's needs. Could mean that it's able to scale it up if its needs are higher, but I imagine this is part of the breeding work they're doing.
darknoon··on OpenAI Tokenizer
We noticed that this webpage is out of date for recent models so we (diagram.com) commissioned a better one that lets you pick any of OpenAI's models including chats:

https://tiktokenizer.vercel.app/

darknoon··on Amazon warns employees not to share confidential information with ChatGPT
They seem to be updating the model roughly weekly, so it would be somewhat expected to see it start to include some more influence from what people have been entering into the chat window.
darknoon··on Unearthing a long-ignored African writing system
Having spent a few months in Senegal, basically if it's on a sign it's probably Wolof, if it's in a Koran or religious school it might be Arabic (though many people can only read the Koran phonetically, they don't understand the language).
darknoon··on Removing the JPEG XL code and flag from Chromium
The comments on the bug are quite rude and condescending, as happens often when HN links to a "surprise" decision in chromium. Would hate to be on the other side of this, though perhaps being more transparent could make their argument sound better.
darknoon··on Apple security bounty upgraded
Apple support can push a screen share to your device. Iirc, you have to accept it before it starts sharing.
darknoon··on Dark Sky iOS app will no longer work from Dec 31
So sad, I use this app every day!
darknoon··on Instagram can track anything you do on any website in their in-app browser
Safari webview (WebKit) is what it's describing
darknoon··on Growing food in hard clay
She tried this in the video… you have to watch it but basically it just bounced off the clay layer. Her soil is exceptionally poor.
darknoon··on Applebee’s exec urges using high gas prices to push lower wages, sparks walkout
I appreciate the quote from corporate: “We are still scratching our head about what this gentleman was thinking.”
darknoon··on Mozilla's Firefox Relay to be added to disposable-email-domains blacklist
Thank goodness that iCloud temporary emails are on the same domain! They wouldn't dare block all iCloud users right?
darknoon··on An ex-ARM engineer critiques RISC-V
Sounds pretty similar to thumb mode on ARM, no?
darknoon··on Ask HN: Good C++ code bases to read?
Personally, I think web browsers are good examples. You see integration with various platforms, abstractions on top of them, and general algorithms, generally with a consistent code style.

- WebKit - Chrome

darknoon··on A trip through the Graphics Pipeline (2011)
Not even SIMD on recent GPUs, it's SIMT.
darknoon··on Building a $5k ML Workstation with Tiitan RTX and Ryzen ThreadRipper [video]
Now is a particularly bad time to build a rig, since new NVIDIA cards are launching in a couple months. The value of a used 2080Ti (Turing) will tank, because Ampere cards will be available with similar performance for half the price.
darknoon··on Intel Initiates EOL for the VCA2: Three Xeons on a PCIe Card
"AVC transcoding at 30 FPS" sounds a lot less useful than what you can get from an NVIDIA consumer or Quadro card via NVENC. Very weird, and multiple Xeons doesn't sound cheap.

Am I misunderstanding the product?

I know Intel is coming out with a discrete GPU, which will probably have plenty of video encode hardware to compete with NVIDIA (esp since game streaming at 4K is quite popular).

darknoon··on Stripe records user movements on its customers' websites
As an end user, can you provide me with all of the event data you have collected about me? Or does Stripe help sites that integrate stripe.js to respond to data requests?

Ie, CCPA / GDPR says I have the right to see and correct it if I live in one of those jurisdictions.

darknoon··on Swift: Google’s Bet on Differentiable Programming
This is not the same thing.

CACurrentMediaTime() / CFAbsoluteTimeGetCurrent() are first of all not deprecated (just check CFDate.h / CABase.h) but return a time interval since system boot so they are guaranteed to be increasing. It's just a fp64 representation of mach_absolute_time() without needing to worry about the time base vs seconds.

Date() / NSDate returns a wall clock time, which is less accurate and not guaranteed to increase uniformly (ie adjusting to time server, user changes time etc)

darknoon··on What's so hard about PDF text extraction?
The first link on your demo gives me an error (mobile safari) https://www.appblit.com/pdfreflow/viewdoc?url=http://arxiv.o...
darknoon··on Apple to Pay Up to $500M to Settle U.S. Lawsuit over Slow iPhones
What they did is pretty logical imo, though it would have been better to tell people at the time. I'm a little disappointed by this judgement.
darknoon··on Siamese Network on MNIST Dataset
How does this qualify as semi-supervised? Labels are being passed in for all data points during training—isn't that full supervision? https://github.com/leimao/Siamese_Network_MNIST/blob/master/...
darknoon··on PHP in 2020
It's a shame that Hack didn't replace PHP in general use. I didn't see a mention in the article, but it was created to be exactly this safer/faster/more elegant version of PHP. For example, it has a broad type system including generics.

My first decent sized codebase was web site for my high school in vanilla PHP. It was a total mess (mostly due to my lack of experience and the need to support Netscape 4.7). I understand that some PHP apps are good, though I got burned by WordPress vulnerabilities enough to swear it off.

I wrote a little code in Hack while at fb, and the incremental typing was slowly evolving a messy codebase to some kind of order. (https://hacklang.org)

darknoon··on Real-Time Ray-Tracing in WebGPU
Hey, I really appreciate your work on these bindings. I have done a lot of work with Metal on iOS, and found it frustrating to start over from scratch when trying to combine graphics and deep learning (CUDA on Linux). It would be awesome to see a future where you can write high-performance GPU-driven apps in a cross-platform way with js/webgpu/wasm and slap in platform-specific UI, without bundling all of Unreal or Unity. Running in a browser would be an optional convenience.

Anecdotally, I was trying connect PyTorch's CUDA tensors to the GL textures that Electron/Chrome uses to render in a Canvas without going through CPU memory, but couldn't figure out where to inject my code. Chromium's GPU code is quite a maze. Perhaps a smarter person will be able to accomplish that.

darknoon··on The problem with parallelisation
This guy really misses the point. You don't put 1024 x86 cores on a chip because the programming model of individual threads being scheduled ad-hoc doesn't scale that well (and imposes a lot of overhead on the control apparatus).

But on a GPU, it is typical to have well over 1000 cores using a different programming model that accounts for memory access and grouping threads together to accomplish shared work and shared memory accesses.

So, the same work is now being done with slightly different algorithms that exploit this programming model.

← PreviousPage 2 of 4Next →