ioedward

AMD didn't submit results, despite geohot/Lisa Su's plan to get AMD on MLPerf: https://twitter.com/realGeorgeHotz/status/166980346408248934...

8 H100s would have enough VRAM to finetune a 70B model.

Not at all, PyTorch uses sub-classing (nn.Module) and other Pythonic features heavily that are not supported with Mojo.

Trying the fuse feature, seemed the most interesting:

> spaghetti ≈ trumpet

> Both spaghetti and a trumpet can be difficult to eat without making a mess—spaghetti with its long, slippery noodles, and a trumpet with its wide, flared bell.

ok google

Related: 1 year represented in a photo. https://eirikso.com/2011/01/04/one-year-in-one-image/

The technical term is GPGPU, general-purpose computing on GPUs, but I like to call them GPUs for short.

Nvidia's enterprise GPUs are surprisingly unreliable. Working on a 128 GPU A100 cluster on AWS, 1 would fail every few days. I didn't have any insight on whether it was a hardware or software failure.

You also need the optimizer (e.g. Adam)'s state, which is usually double the parameter's size. So if using fp16, one parameter takes up 6 bytes in memory.

TPUs are mostly hoarded by Google Research (including Deepmind) and Ads. Very few are being used by external people.

The twitter thread is worth reading, Yann responds to some of the questions raised here.

No it won't. Large language models are trained on 1,000 - 50,000 GPUs. No one's going to buy hundreds of Mac pros to mount them in a datacenter for training ML models.

Normally people split up the model across multiple GPUs, i.e. model/tensor parallelism.

Doesn't seem to be affiliated with OpenAI. Domain was registered in 2017, 2 years after OpenAI was created. Trolling/domain squatting?

> I understand that PyTorch is an awesome tool for researchers, but it doesn't necessarily fit into a prod environment.

Some of the largest deployments of ML are using PyTorch models, e.g. OpenAI, Meta, Microsoft.

Where did you buy the Instinct hardware and what's a good price?

Archive.org: https://web.archive.org/web/20150120162544/http://www.sbb.co...

So are you still shipping Python/Django as an executable with atom-shell? Or is this purely client-side javascript?

I'm getting stuck on the sign in page. After authorizing the app with Github, there's a black box that appears and the Sign In button shows up again.

Edit: Looks like the server is 500ing, according to the Javascript console.

dang killed it because he didn't like it. (I'll go look for the source)

edit: source - https://news.ycombinator.com/item?id=7693262

> Ok, when people start racing to post these at midnight, and beg for upvotes on top of it, this experiment has officially jumped the shark. I'm going to bury this post and ask you all not to post any more of them. Only one account (whoishiring) is allowed to make regular feature posts that we don't kill as duplicates. (That's for the obvious reason of preventing karma sweepstakes and race conditions.) Should we make this "Idea" thread a regular feature? I've thought about it quite a bit. I think the answer is no. Experiments are worth trying, but this one has gone on for a month now and I don't think it has cleared the bar [1]. Something about having all these ideas in one place makes the whole less than the sum of its parts. The threads seem to me to have gotten less interesting as they've become more regular. I'm sorry to disappoint those of you who disagree. But our job is to optimize HN for quality and I don't think the quality is high enough here. Ideas are better in the wild. Let's discuss them as they come up organically, rather than try to organize an idea-fest. 1. https://news.ycombinator.com/item?id=7682938

Would love to hear more about the terminal as well - how did you input commands to it, and how does it compare to using a phone or computer?

I emailed you, but also wanted to express in the comments that I would be definitely interested in idempotent json patches.

Sure! I would _not_ advise getting the Nexus 5. As noted below in other comments, the image quality and autofocusing speed has much to be desired. The Camera2 API that supports DNG is only available on Android 5.0 phones.

http://edward.io/raw_nexus_5.zip

There's nothing technical that's stopping them from doing it. I expect third party apps to do something similar once Lollipop has a higher market share.

All hand held, shot from 1/35th of a second to 1/1000th of a second. HDR Pro does some basic auto aligning and ghost removal, AFAIK.

I've been messing around with the Google Camera app and the new Camera API provided by Android 5. It turns out that you can actually obtain a lot higher quality images by using the DNG (digital negative) of the photo, instead of using the JPEG. By "HDRing" images yourself, you can actually outperform Google Camera's HDR+ functionality.

Here's the JPEG, non HDR+ shot on a Nexus 5: http://i.imgur.com/So44muL.jpg

Here's a similar image shot with HDR+: http://i.imgur.com/QFS3ZYd.jpg.

As you can see, the dynamic range is increased greatly; however there's strange black spots in the shadows.

Here's the same photo that I took in DNG format, edited in Lightroom:

http://i.imgur.com/VRFsnf5.jpg

And here's my HDR photo, combined 5 DNG exposures inside Photoshop HDR Pro's functionality:

http://i.imgur.com/RTT6ULz.jpg

Okay, I misunderstood you then. I haven't found any words ending with "io" that are still available.

No real words? I just ran my script through a list of the most common english words, and it returned 4k .IO domains. I'll post them if anyone's interested.