Blaesmo

alt acc for browsing the internet, a ml researcher on computer vision and downstream applications, opinions are own, i do not provide any consultant services.

[Blaesmo @blaesmo]

@song_jin73347 @K2keUtBetSkPmlI @by_yueeeer blender, 3d渲染软件

[Blaesmo @blaesmo]

@Minecraftxfj 如果你和你的导师还有老板还有上司频繁聊论文和实现和工程就不会有这种情况了

[Blaesmo @blaesmo]

@jecdohmann @skalskip92 have you tried instead using something like tipsv2 for vision encoder which have cls token for text and image space alignment when input an image? that might benefit more from cot as it have text cls, and then pad the image cls token back into the end of the cot, wonder how it go

[Blaesmo @blaesmo]

@skalskip92 however this all depends on how a model treat image, is it through vision encoder, is it through feature extraction model like dino or patch method like tipsv2, the dino or tipsv2 path is less explored

[Blaesmo @blaesmo]

@skalskip92 so if you have cot on for detection you are feeding unrelated text to be used to pad groundtruth of the image and llm obviously goes into more hallucination because of this, image and text is not the same, you dont send image multiple time to think, you do it once

[Blaesmo @blaesmo]

@SnozakiSakura 哦然后别用来干自己私事，会监控

[Blaesmo @blaesmo]

@SnozakiSakura 前提是你学业上有这个需求并且学校受理

[Blaesmo @blaesmo]

@georgioguinta @unusual_whales what im afraid is that this lawsuit go into the wrong direction and go from forming a formal foundation of licensing business data into simply blackmail companies and profit which only benefits the record labels

[Blaesmo @blaesmo]

@georgioguinta @unusual_whales as someone who works on ai myself i think artist should be the one to be paid respect to, not the label company, or at least artist should have a cut of the cake unlike current situation and they should have public licensing with clear price, rn its not transparent at all

[Blaesmo @blaesmo]

@unusual_whales so, if i ask claude and other llm to search lyric online and i toggle help improve model so they train on the conversation and learn lyric of a song, then i can call the song copyright owner to sue the ai company for 150k per song? this is dumb.

[Blaesmo @blaesmo]

@georgioguinta @unusual_whales i know this becuz ppl have been feeding gov data to llm and believe me many of them dont know to toggle off the improve, they are dum user. there are policy for this yes, but they never sue the company, just the person who use ai, this why there are 0 trust policy

[Blaesmo @blaesmo]

@georgioguinta @unusual_whales like what happen if i(user) simply upload million of lyric or asked claude search the lyric of song then i toggled help improve model, model would read the lyric online and then that process get feed as training data, then i do it with every ai, sony can sue any and everyone

[Blaesmo @blaesmo]

@Pregory1 @InForTheFun_ @unusual_whales so chances are the lyric come into their train data from books, online forums like reddit etc. which then the thrid party platform is to blame instead of anthropic(reddit or other place that hosted the lyric which Antr scraped), scraping in such scale hv no knowledge of all data

[Blaesmo @blaesmo]

@Pregory1 @InForTheFun_ @unusual_whales what i find very problematic is that they claim lyric as music piece even tho anthropic llm do not perfectly recall all lyrics, and that they have no dead proofs on how anthropic obtained the data for the lyrics, all we know is anthropic scan and burn books which is legal n legit