AI General Thread

[the-pi-guy]

Tokens too expensive for you too? 

Nah I just use the free stuff. 

Just haven't felt like it's been as needed lately. 

[the-pi-guy]

Me: "I got a 90% reduction in file size."

Copilot: "Oh that's really great. It's shaving off a lot of unnecessary data."

Me: "Oh I was looking at the wrong line, it was actually an 80% reduction."

Copilot: "Yeah that makes more sense. It would have been a red flag if it was a 90% reduction."

[Legend]

Me: "I got a 90% reduction in file size."

Copilot: "Oh that's really great. It's shaving off a lot of unnecessary data."

Me: "Oh I was looking at the wrong line, it was actually an 80% reduction."

Copilot: "Yeah that makes more sense. It would have been a red flag if it was a 90% reduction."

I hate these types of convo quirks in ai so much lol.

There is zero reason to trust anything they say cause they will happily say the opposite ten seconds later!

[Legend]

OK this AI world on instagram is just too much.

Every ukulele girl has the same background.
 

These are just two examples. And the top one I know is a real girl cause I have previously watched her on youtube: https://www.youtube.com/watch?v=Vg1rl65u19Y

But is this a real video of her in this room? Who is the actual owner of this room!?!?!?!?

[the-pi-guy]

OK this AI world on instagram is just too much.

Every ukulele girl has the same background.
 

These are just two examples. And the top one I know is a real girl cause I have previously watched her on youtube: https://www.youtube.com/watch?v=Vg1rl65u19Y

But is this a real video of her in this room? Who is the actual owner of this room!?!?!?!?

That's just Green Screen Gary's room. 

[the-pi-guy]

Apparently there is an all in one model MiniMax-H3. Does video, text, image, audio apparently.  

Apparently it's incredible, even uncensored and all that jazz.


And apparently it's against the license agreement to use it in the US/Europe....  

[Legend]

Apparently there is an all in one model MiniMax-H3. Does video, text, image, audio apparently. 

Apparently it's incredible, even uncensored and all that jazz.


And apparently it's against the license agreement to use it in the US/Europe.... 

How big is it?

[the-pi-guy]

How big is it?

I stole someone's answer here:

int8convrot models are 34gb, bf16 models are 66.3gb.

there's also pruned variants that are compacted to 21gb.

the text encoder: bf16 at 51.5gb, int8convrot at 27.1gb, nvfp4_awq at 15.7gb

the vae's are also massive, 605mb for audio and 5.21gb for video.

[Legend]

I stole someone's answer here:

Wow that's pretty small!!!

Well, compared to kimi and deepseek lol.

What's the biggest one you've downloaded yourself?

[the-pi-guy]

Wow that's pretty small!!!

Well, compared to kimi and deepseek lol.

What's the biggest one you've downloaded yourself?

Like ~30 GB.  

[Legend]

Like ~30 GB. 

One-shot MiniMax H3 generation 🤯

Prompt was "Jim and Dwight from The Office discuss autonomous coding agents" pic.twitter.com/G7CnSo6SUd

— Justine Moore (@venturetwins) August 3, 2026

Not bad!

[the-pi-guy]

One-shot MiniMax H3 generation 🤯

Prompt was "Jim and Dwight from The Office discuss autonomous coding agents" pic.twitter.com/G7CnSo6SUd

— Justine Moore (@venturetwins) August 3, 2026

Not bad!

It seems to do wildly better with scene transitions than any other local model I've seen.

[Legend]

Google is having issues. Deepmind CEO just stepped down to chair, 3.5 pro is missing in action.

[the-pi-guy]

I was able to get the model running on my computer. 

It's quite a bit harder to run.
With LTX-2.3 I was making 10 seconds videos in about 10 minutes, at 720p.

Here I'm making 7 second videos in like 13 minutes at 480p. 

You still see the funky AI stuff. Like I generated a video of someone going under a blanket, and the blanket very obviously phased through them.

I don't think LTX-2.3 was as capable of getting completely different views of the same scene. It could do smooth turning the camera around for sure. But I don't think it could swap back and forth on two characters. I had it generate in a single go, where it swaps back and forth between two characters twice, and it does really well. 

I also got my pipeline working for making extended videos. And I had it basically make 3 7 second videos that could be stitched together as a seamless video. Although I didn't listen to the audio back and forth to see how close that was. I just had it make more from the image.

There is a reference pipeline that would probably make it even better.

[Legend]

I was able to get the model running on my computer. 

It's quite a bit harder to run.
With LTX-2.3 I was making 10 seconds videos in about 10 minutes, at 720p.

Here I'm making 7 second videos in like 13 minutes at 480p. 

You still see the funky AI stuff. Like I generated a video of someone going under a blanket, and the blanket very obviously phased through them.

I don't think LTX-2.3 was as capable of getting completely different views of the same scene. It could do smooth turning the camera around for sure. But I don't think it could swap back and forth on two characters. I had it generate in a single go, where it swaps back and forth between two characters twice, and it does really well. 

I also got my pipeline working for making extended videos. And I had it basically make 3 7 second videos that could be stitched together as a seamless video. Although I didn't listen to the audio back and forth to see how close that was. I just had it make more from the image.

There is a reference pipeline that would probably make it even better.

Mind sharing your best vs worst results so far?

13 minutes for 7 seconds isn't bad if most generations are useful. Could let it render overnight