Max

🇪🇺eu/acc, compressing reality (pied piper)

[Max @max_pe2002]

@engnr_george @niccruzpatane for collection and logging this would be absolutely overkill

[Max @max_pe2002]

@DavidSHolz dont give them ideas

[Max @max_pe2002]

@teslaeurope @yunta_tsai the EU is actively killing people by now allowing Tesla FSD

[Max @max_pe2002]

Thought my ImageNet model was brokens until i finally increased the CFG. CFG=[1.0, 1.5, 2.0, 2.5, 3.0, 4.0]

[Max @max_pe2002]

@ph_reinhardt can you link the paper?

[Max @max_pe2002]

@___Harald___ @antferdom how do typical robotics models work for self driving? like simple diffusion policy which is pretrained on all data and finetuned on good data? in my mind this would already completly solve self driving. (with high context lenght)

[Max @max_pe2002]

@JDVance @elonmusk the USA saved Europe from the Nazis when will the USA save Europe from mass migration?

[Max @max_pe2002]

@MozarellaPesto @ludocomito so FD Loss?

[Max @max_pe2002]

@MozarellaPesto do you think you could get rid of the transforms if adding a small image space mse loss?

[Max @max_pe2002]

why arent LLM people using things like REPA? i think SRA could be usefull. aligning an early layer with a late layer

[Max @max_pe2002]

@tokenpilled65B @madebyollin @Shauray7 maybe its a eval issue. JiT atleast for me doesnt produce good fine details for far away objects which might not impact FID much but which might impact perceptual losses

[Max @max_pe2002]

@tokenpilled65B @madebyollin @Shauray7 oh okay thats very interesting isnt this a bit then what JiT says. like JiT does ps64 with dim 768 and stuff

[Max @max_pe2002]

@tokenpilled65B @madebyollin @Shauray7 im still working on the arch before scaling. i did a small test scale with 1.4b params total. with the decoder having a dim of 1920

[Max @max_pe2002]

@tokenpilled65B @madebyollin @Shauray7 yep convs seem to work better than just ViTs but i dont like convs

[Max @max_pe2002]

@tokenpilled65B @madebyollin @Shauray7 im also training tokenizer right now and LPIPS and Dino loss cant help my large patch size issues a little but not much

[Max @max_pe2002]

@madebyollin @Shauray7 JiT allows large patch sizes to work which doesnt mean that large patch sizes are equally good. also JiT really only works well for Imagenet as subjects are close. As soon as a face is only handled by like 1 patch it just cant model it properly

[Max @max_pe2002]

@ostrisai why arent you using FD loss for this?

[Max @max_pe2002]

@elonmusk @SawyerMerritt @AnthropicAI @SpaceX when can individuals or smaller labs rent them?

[Max @max_pe2002]

@iScienceLuvr isnt FID also measuring coverage?

[Max @max_pe2002]

@HyperTechInvest i remember a time i could get 8x B200 for 8$