Paul
210 posts
Mistral is 100% ahead of DeepSeek
Recipes are commodified. Now, compute is everything.
Pack it up, Huawei bois.
(but tbh I feel a little bad how we're mocking this model. Maybe it's actually great, like has short outputs, amazing diversity etc. so enough)
stochasm@stochasticchasm
if we suspect 30T tokens this is 2.3e25 flops in pretraining, not to mention how many flops would have been spent in RL inference
English
@teortaxesTex But, but, I have a lot of questions about Mistral AI. I really can't figure it out. Maybe they're working on AGI too.
English
@LoveMePaul I am not talking about Mistral
English
happy day for GPU-poor enterprises I guess
4chan bros driving GLM in Q1 on old Supermicro will be ecstatic too
the Swan Song of the Dense
Lisan al Gaib@scaling01
Mistral Medium 3.5 is out and it's a dense 128B model
English
@teortaxesTex Of course, you're right—for companies facing chip shortages, today is definitely a day to breathe a sigh of relief.
English
@teortaxesTex Large-scale fundraising, large-scale capital expenditures, large-scale hiring, and large-scale AI infrastructure development！😼
English
Redditors speculate
I think DeepSeek won't have to worry about instabilities very soon, and this should be quite scary for their competitors
English
…they're not out of rabbits, huh. Seeing as vision was spotted in the website HTML, I guess "character mode" and "coder mode" will also be released like this. They can do a trick a week at least 2 more weeks.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
DeepSeek's tactics are amusing "OK, you don't think reliable 1M context, fast, for pennies, at top-tier open source intelligence, is cool enough? Have a 75% price cut. Not entertained? Have a ≈free cache. Still meh? How about vision?" I wonder if they're out of rabbits
English
@teortaxesTex @lipeng0820 @zephyr_z9 Since I wasn't selected for the beta test, I can only use the test results they sent me!😭
English
@LoveMePaul @lipeng0820 @zephyr_z9 is it any good? Can you prompt it in English. What does it call itself?
English
Deepseek is currently conducting a large-scale rolling beta test of its image recognition mode!👀 @teortaxesTex @lipeng0820 @zephyr_z9
English
We’ve just received more test results regarding the image recognition mode🤔🤫🤫.@teortaxesTex @lipeng0820 @zephyr_z9
English
Here comes the big one! Over the past six months, Deepseek hasn’t just been focused on developing V4—more importantly, it’s been reshaping the landscape of AI competition from the ground up! @teortaxesTex @lipeng0820 @zephyr_z9
CN Wire@Sino_Market
🇨🇳BIG CHINESE TECH FIRMS RUSH TO SECURE ORDERS FOR HUAWEI CHIPS AFTER DEEPSEEK PREVIEW - SOURCES SCRAMBLE FOLLOWS RELEASE OF DEEPSEEK V4 MODEL USING HUAWEI CHIPS #CHINA #TECH #AI #DEEPSEEK #HUAWEI (mktnews.com/flashDetail.ht…)
English
接上一条，聊聊中微跟国际一线大厂的差距和其主要客户。
CCP差距最小，局部已经追平。5nm节点上进了国际一线客户的量产线，深宽比做到90:1，跟泛林同代产品在同一档。但泛林在3D NAND领域积累太深，200+层NAND的超高深宽比（200:1以上）目前还是泛林的绝对领地。泛林一年刻蚀收入100亿美元级别，中微2024年全部刻蚀收入约73亿人民币，体量差接近十倍。ICP差一个身位，进了产线但最复杂的几道关键工序还在验证。LPCVD去年才开始放量，ALD刚进客户端验证，这两个品类跟东京电子、ASM的差距最大。
真正的分水岭在3nm及以下的GAA架构。GAA相比FinFET的刻蚀工艺步骤从5道增加到9道，刻蚀设备用量占比从20%升到35%。中微在这个节点上有没有成熟量产方案，公开信息里还看不到。5nm站稳了，3nm/GAA能不能接住，决定了中微到底是"国内龙头"还是"全球第一梯队"。
客户端的情况跟技术差距对得上。海外，台积电、SK海力士、联电、格罗方德都在名单里，台积电CCP从7nm就入选了，5nm也在用，但同一道工序台积电通常认证两到三家供应商，中微拿到的是部分工序的份额。国内，中芯国际、长江存储、华虹宏力、长鑫存储，逻辑和存储两大阵营的头部基本覆盖，CCP在国内逻辑器件两家头部厂商中份额达到39%，还在提升。海外有标杆但份额有限，国内是基本盘且渗透率在加速。出口管制持续加码，国内晶圆厂验证国产设备的意愿比以前强得多，这条线的确定性很高。
Macro_Lin ｜ 市场观察员@LinQingV
很多人知道中微做刻蚀设备，但不太清楚CCP、ICP、LPCVD、ALD分别是什么。简单聊一下。 刻蚀是把材料去掉，薄膜沉积是把材料加上去。芯片制造就是在这两个动作之间反复循环，制程越先进，循环次数越多，精度要求越高。 CCP，电容耦合等离子体刻蚀，两块平行电极板之间产生等离子体，离子能量高，专门刻介质材料。3D NAND里那些又深又窄的沟槽主要靠CCP，中微Q1报出的90:1超高深宽比就是CCP的活。ICP，电感耦合等离子体刻蚀，用线圈产生等离子体，能量相对柔，适合刻金属和硅，逻辑芯片的栅极、接触孔这些精细活归ICP管。140:1的3D DRAM深宽比就是ICP的成绩。 LPCVD，低压化学气相沉积，在晶圆表面镀一层均匀的薄膜。ALD，原子层沉积，一层原子一层原子地往上镀，精度到埃级别，GAA架构那种纳米级多层堆叠只有ALD能做到。 中微从CCP起家，现在同时推ICP、LPCVD、ALD，本质上是往平台化方向走。
中文
