beep@piefed.world to Not The Onion@lemmy.worldEnglish · 6 days agoAnthropic calls for pause of global AI developmentwww.rte.ieexternal-linkmessage-square13fedilinkarrow-up151arrow-down11file-text
arrow-up150arrow-down1external-linkAnthropic calls for pause of global AI developmentwww.rte.iebeep@piefed.world to Not The Onion@lemmy.worldEnglish · 6 days agomessage-square13fedilinkfile-text
cross-posted from: https://piefed.world/c/onion/p/1174854/anthropic-calls-for-pause-of-global-ai-development
minus-squareelgordino@fedia.iolinkfedilinkarrow-up13·6 days agoSurely not because Deepseek V4 is 95% as good for 3% of the price. An unrelated concern I’m sure.
minus-squareIncogCyberSpaceUser@piefed.sociallinkfedilinkEnglisharrow-up3·6 days agoHow do they manage that, btw? Are their models so efficient or are the subsidizing their service like no other?
minus-squareelgordino@fedia.iolinkfedilinkarrow-up6·6 days agoThey have a new attention mechanism, that in short, compresses the context. https://dasroot.net/posts/2026/04/deepseek-v4-hybrid-attention-massive-contexts/ And yes probably subsidies too
minus-squareboonhet@sopuli.xyzlinkfedilinkEnglisharrow-up2·6 days agoWouldn’t be surprised if the big players have something similar, given they both updated to 1M context window recently Still, this is great news for everyone whose name isn’t Dario or Sam. Assuming they have open source it or at least publish the whitepaper
Surely not because Deepseek V4 is 95% as good for 3% of the price. An unrelated concern I’m sure.
How do they manage that, btw? Are their models so efficient or are the subsidizing their service like no other?
They have a new attention mechanism, that in short, compresses the context.
https://dasroot.net/posts/2026/04/deepseek-v4-hybrid-attention-massive-contexts/
And yes probably subsidies too
Wouldn’t be surprised if the big players have something similar, given they both updated to 1M context window recently
Still, this is great news for everyone whose name isn’t Dario or Sam. Assuming they have open source it or at least publish the whitepaper