The Ai thread

[dada_dave]

Nvidia and Apple: Frenemies are putting grudges aside to supercharge AI performance

Apple and Nvidia have announced a surprise collaboration targeted at improving LLM inference performance, combining Apple's open-source ReDrafter with Nvidia's raw silicon grunt. The result? A 2.7x improvement in performance over auto regression.

www.notebookcheck.net

This is interesting, Apple actually teaming up with Nvidia to improve Apple’s LLM performance on Nvidia’s GPUs. As the article states, collaboration between the two has been very rare over the last 16 years.

(EDIT) Apple’s post:

Accelerating LLM Inference on NVIDIA GPUs with ReDrafter

Accelerating LLM inference is an important ML research problem, as auto-regressive token generation is computationally expensive and…

machinelearning.apple.com

Here’s Nvidia’s more technical blog post:

NVIDIA TensorRT-LLM Now Supports Recurrent Drafting for Optimizing LLM Inference | NVIDIA Technical Blog

Recurrent drafting (referred as ReDrafter) is a novel speculative decoding technique developed and open-sourced by Apple for large language model (LLM) inference now available with NVIDIA TensorRT-LLM.

developer.nvidia.com

 

[Eric]

 

[Eric]

Almost all of my authentic video/reel posts get accused of being AI now and I've given up on defending it.

 

[Herdfan]

Almost all of my authentic video/reel posts get accused of being AI now and I've given up on defending it.

What gets me is I am a member of a Sedona visitor’s FB page and we don’t get many issues there, but other travel pages put up these clearly AI photos claiming to be Sedona that aren’t.

But the kicker is when someone responds that they have been to the place in the photo. No you haven’t because it doesn’t exist.

 

[Eric]

What gets me is I am a member of a Sedona visitor’s FB page and we don’t get many issues there, but other travel pages put up these clearly AI photos claiming to be Sedona that aren’t.

But the kicker is when someone responds that they have been to the place in the photo. No you haven’t because it doesn’t exist.

A lot of new pages were popping up that were featuring other people's art posted without credit and a lot of it was mine. They had no way to report copyright without going through a huge process so I gave up on it, it was one of the driving factors for me leaving the platform, Zuckerberg was just the final straw. They can steal my work but at least I no longer have FB recommending the pages of said thieves.

 

[Herdfan]

Here is a perfect example:

Can’t prove the photo is AI but I can’t determine what kind of car is on the road.

But the comments. SMH. This “supposed” road is 15 miles from my house and looks nothing like this.

 

[Nycturne]

That looks a lot more like it fits near Yosemite to me.

But there are elements of the car and the road markings that don't look right, but it's hard to tell from a screenshot like this.

 

[Herdfan]

That looks a lot more like it fits near Yosemite to me.

Here is a reel of the actual drive.

But there are elements of the car and the road markings that don't look right, but it's hard to tell from a screenshot like this.

Yes, the car seems too wide and too short. And I know cars pretty well and can’t even begin to identify the make or model.

 

[Yoused]

I know cars pretty well and can’t even begin to identify the make or model.

It looks a bit like an Oldsmobile from the '70s or early '80s. The length could be distorted by a long lens. But that is some damn sloppy driving, going over a double yellow line like that.

 

[throAU]

The quality of some of the stuff I've been playing with and getting out of chatgpt is pretty good.

ChatGPT - Fresnel Zone Explained Simply

Shared via ChatGPT

chatgpt.com

 

[ArgoDuck]

Yeah, it’s tricky.

I’ve found it handy for some short Swift/SwiftUI coding problems where I know it could take me a few hours digging around to access the right material and put it together myself. I would say it’s been correct and relevant in 80% of cases (4/5 in the last 12 months), where it’s clearly absorbed and (accurately) summarized a lot of material from other coders.

I do not and would not trust it for any code longer than a few lines.

Mathematics? Good at finding a well-covered formula and describing how to apply it, per your Fresnel example. Asking it to reason its way through a problem? Not good; as i think openAI acknowledge, o3 notwithstanding.

As for anything around core creative reasoning work - the kind i do anyway - definitely not. It interpolates well - barring the risk of hallucination - by summarizing the middle or a spread of positions in an area already well covered in its training material.

But anything novel…no!

An area i expected it might work well relates to its summarization of large blocks of text. One use case is that of extracting personality metrics from biographical texts. Because (validated) psychometric instruments - those that use self-report questionnaires - are simply a series of statements that represent a type and degree of behavior, i speculated ChatGPT and Gemini might do a good job of deriving ‘Five Factor Model’ numbers from biographies of a few hundred words.

And they both do, with some caveats. Regarding prominent aspects of a person they agree well; on other dimensions, agreement is poor (ie, inter-observer reliability low).

It could have been an interesting side project supplementing my other work, but not wishing to feed google and openAI others’ biographical work product i abandoned the idea.

Overall, i’ve found it occasionally a great time saver for some kinds of things given careful prompting (but so far my use has steadily declined FWIW).

 

[throAU]

Yeah, supposedly the new o1 model is better at reasoning. I have a paid account and am starting to play with it. Pretty sure the above was with the 4o model tho.

You're on the money with summarising or querying large blocks of text. I've used local models in lmstudio for that a bit (not for personality metrics, but for things like contract terms - as a test), haven't with GPT yet as I'm less willing to upload files to them. Generally you'll get answers to queries about the text with a citation. Helpful for finding references, etc.

 

[Eric]

Here is a perfect example:

Can’t prove the photo is AI but I can’t determine what kind of car is on the road.

But the comments. SMH. This “supposed” road is 15 miles from my house and looks nothing like this.

View attachment 33461View attachment 33462

Sounds like you guys are doing the necessary sleuthing but the first comment here is dead on. I spend a lot of time planning some of my shots, time of day, travel, angle, zoom, etc. just to have a bunch of yahoos online tell me it's AI. It can be both frustrating and deflating, but the lines have become so blurred that everyone is questioning even the most authentic work now.

You guys probably saw my post awhile back where Snopes.com took up someone's claim that my photo was fake when it wasn't. The cat's out of the bag and it is what it is for now but I'm hoping they come up with ways to easily and genuinely authenticate art, leaving it up to the community notes from the clueless and opinionated masses isn't the way to go about it.

 

[throAU]

So... creative types.

I made a thing.

Patch to the Heart by @turbulentroadie8613 | Suno

industrial black metal song. Listen and make your own with Suno.

suno.com

(yeah I'm a Microsoft patch administrator amongst other things)

 

[Eric]

So... creative types.

I made a thing.

Patch to the Heart by @turbulentroadie8613 | Suno

industrial black metal song. Listen and make your own with Suno.

suno.com

(yeah I'm a Microsoft patch administrator amongst other things)

Spend your life learning an instrument and refining your craft, OR, take ten seconds and type in your idea of a cool song into an AI generator.

I'm glad to have lived most of my creative life before this shit came around.

 

[Herdfan]

Spend your life learning an instrument and refining your craft,

My BIL is a music producer in Nashville. The is also one of the top mixers.

He has limits on how much he will tune a singer. One or two in a song, fine. But he won’t fix a bad singer.

 

[Eric]

Just wanted to share this, look at the effort they went through for this real shot. Nothing but respect.

 

[Eric]

Have to hand it to the TikTok algorithm, when I posted this the only location I added in the description was San Francisco but somehow it got to a ton of people who actually live in this specific neighborhood and they've all been commenting. From a technical standpoint the only way I can figure it's possible is visually geo targeting it.

 

[Agent47]

Have to hand it to the TikTok algorithm, when I posted this the only location I added in the description was San Francisco but somehow it got to a ton of people who actually live in this specific neighborhood and they've all been commenting. From a technical standpoint the only way I can figure it's possible is visually geo targeting it.

Maybe Tiktok is reading the image meta tags?

 

[KingOfPain]

Let the grift begin (although I'm guessing most of AI is some kind of grift anyway):

Couple allegedly tricked AI investors into funding wedding, houses

FBI claims GameOn founder forged six years of financial records in brazen scheme.

arstechnica.com