seekinganythingbutalpha (@ivanalog_com) header
seekinganythingbutalpha (@ivanalog_com) profile picture

Focus on 1) Technology and Dividend Blended 2) Drawdown protection. Check highlights and articles. No investment advice. Now building ModelDock

Iceland
Joined October 2011
Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)Photo by seekinganythingbutalpha (@ivanalog_com)
seekinganythingbutalpha retweeted
Amazon and Google sold ~600M+ Alexa and Nest devices that could barely do anything except playing music and setting timers (and even that worked reliably maybe 30% of the time). Now we have effectively AGI-level intelligence, but nobody seems to be building smart speakers around it. Why?
608
101
56
6,095
492,803
seekinganythingbutalpha retweeted
$GTLB 最近走的真好。 其实SaaS大面积回暖了,很多慢牛股,买一栏子安心守着就好了。 Deepseek这个API性价比,对SaaS公司和业务AI化非常友好。
$GTLB Cup & Handle形态突破, 图看着不错 GitLab 算是这轮软件股里, 数没被 AI 抢走生意的公司之一 最新一季营收 2.64 亿美元,同比增长 23%;ARR 超过 10 万美元的大客户增长 18%,净收入留存率还有 117%,未来一年待确认收入增长 24%。 至少从数据看,大客户不但没有离开,还在继续增加订阅,继续付费。 很多企业从代码仓库、Merge Request、CI/CD、测试、安全扫描,到权限、审计和部署,整个软件开发流程都在 GitLab 里面。 现在 AI 可以直接在里面生成代码和测试、Review 代码、修复 Pipeline、处理安全漏洞。Claude Code、Codex 等外部 Agent 也能接进来,由 GitLab 提供项目上下文、权限和审计。 大模型越商品化,对 GitLab 反而越有利。 模型以后可以随便换,难替代的是企业积累多年的代码库、Pipeline、权限系统、历史记录和合规流程。AI 写出来的代码越多,企业越需要一个控制平台,知道谁改了什么、有没有跑测试、程序员协作。 $GTLB 这种典型被之前被错杀的软件股买起来
3
1
1
10
13,072
seekinganythingbutalpha retweeted
苹果的M5新品在本地推理上,表现非常突出 根据现有的M5芯片表现估计 作为不需要动手安装的产品,M5 Max 64G和 M5 Ultra 96G ,无论是速度还是价值,都达到一线水平, 本地模型输出成本仅高于5090,速度也是仅次于5090为代表的消费顶卡。 详细参见 modeldock.net/benchmarks/
The new Mac mini is here. Small in size. Big on performance. From everyday productivity to all things AI, it can help you do it all.
34
9
2
53
25,989
seekinganythingbutalpha retweeted
估计苹果的人看到我这个帖子, 当时都是会心一笑。 而为什么其他厂商,面对如此大的市场需求,无动于衷呢? 最终,新时代的苹果 还是苹果。
目前市场上,哪家芯片/集成商/OEM先造出: 128GB到256GB GPU-accessible memory, 500GB/s到1TB/s以上带宽, 足够好的INT4/FP4算力, 200到400W,安静稳定,Linux/Windows,Ollama/vLLM开箱即用, 一万美元以内, 就会成为AI时代的苹果。
20
1
1
27
22,803
seekinganythingbutalpha retweeted
真不知道有多少人真的天天用Opus Fable,这么踩deepseekv4flash 捧A社。 Opus 4.6以后一直在倒退,优化感觉都是为了节约资源而设计的。本身这个模型幻觉也相当严重(而且有非常强的不诚实,reward hacking,说大话的倾向),砍了上下文以后更是变身弱智。Opus写作能力也差于Sonnet,所以最近几个月一直是闲置状态。 Fable是很强大,但是它的主要问题并不是Token不够,而是速度实在太慢。Fable从启动,到研究完一个问题,到开始工作,一般需要一到两分钟,有这个时间,deepseek已经交活了。 Deepseek的指令遵从性,诚实度都很高,reward hacking的倾向非常低,速度奇快,这导致它在适当的Harness记忆和指令优化下,特别适合作一个工作AI。
26
9
2
139
28,230
seekinganythingbutalpha retweeted
年初有多嫌弃大科技,追芯片 现在就有多FOMO大科技,软件 瞬间故事线没人聊Anthropic破坏微软了😂新高见
76
4
6
266
98,525
seekinganythingbutalpha retweeted
Please answer honestly and accurately
🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta! 🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇 🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex! Check out the configuration details in our official API docs: api-docs.deepseek.com/quick_…
22
80
11
1,170
138,263
seekinganythingbutalpha retweeted
Jukan这篇文章写的非常好,推荐阅读。 过去,采购存储追求的是最低单位容量成本($/TB);AI时代,更重要的是最低系统总拥有成本(TCO)。 如果更快的Flash能够让价值极高的GPU少等待几秒钟、几分钟甚至几小时,那么增加的存储成本很可能远低于节省下来的GPU成本。在这种情况下,企业的采购决策就会从“买最便宜的存储”,转向“买能让整个AI集群效率最高的存储”。 这意味着,高性能企业级NAND和AI存储系统的价值,不再只是容量,而是能够提升整个AI基础设施的投资回报率(ROI)。这也是为什么越来越多的云厂商和AI公司愿意投入更多预算建设高性能Flash存储层,而不仅仅追求最低的每TB成本。
Replying to @chnyfng127674
If you push this argument to its logical extreme, you arrive at the following conclusion: GPUs are getting three times faster every two years, while storage is failing to keep pace. “The criterion for selecting storage media has expanded from $/TB to $/GPU-hour.” Traditional high-capacity storage has primarily been evaluated in terms of $/TB—in other words, how cheaply it can store large volumes of data. But in AI training infrastructure, where GPU availability, power, and latency are critical constraints, that metric alone is no longer sufficient. If data arrives too slowly from storage and leaves GPUs idle, the cost of wasted GPU time may exceed any savings achieved on storage. In simplified terms: Incremental cost of flash < GPU idle time eliminated by flash × cost per GPU-hour For workloads where this condition holds, using flash is economically rational despite its higher cost per terabyte. GPU-adjacent storage tiers should therefore be evaluated not only by $/TB, but also by the total system cost per effective GPU-hour. In other words, up to a certain ceiling, it could make economic sense to spend several times more on flash than we do today.
53
10
67
22,555
seekinganythingbutalpha retweeted
Replying to @ivanalog_com
其实应该是模型的能力上去了,需要的算力减少了,我自己在用很明显,同一个任务现在基本上一次就完成,减少了好几轮的运算。所以这个也是 token 收费的不合理之处,因为这个 token 它是厂商定义的,调一个智能度,可以多收你好多钱
1
3
1,453
seekinganythingbutalpha retweeted
Samsung is reportedly in talks with AMD to manufacture CPUs through its foundry business, with mass production expected in 2028. LFG!
Samsung sees rising chip production requests from BYD, Google, AMD- Nikkei Asia BYD, Google, AMD, Tesla and others are looking increasingly to Samsung Electronics for contract chipmaking services as surging demand for AI infrastructure strains the advanced chipmaking capacity of market leader TSMC, Nikkei Asia has learned. Samsung has seen a surge in inquiries from existing and potential global customers, including those in China, looking to use its advanced chip manufacturing capacity, according to six people with knowledge of the matter. BYD, China's top electric vehicle maker, has been in talks with Samsung about building future generations of autonomous driving chips, while Google is engaging with the South Korean company about potentially producing its next-generation Axion processors, set to launch around 2028, two people familiar with the matter told Nikkei Asia. Google is also considering asking Samsung to build part of its key Tensor Processing Units (TPUs) for AI computing workloads, as early as 2028, one of the people said. "TSMC is prioritizing advanced-node production not only because it reinforces its tech leadership and long-term strategy, but also because those nodes are more profitable and remain in short supply," said an executive from a Chinese autochip designer, who asked not to be named. "Samsung's yields still lag behind TSMC's, but capacity availability has made it an increasingly attractive option." Yield is the industry term for production quality, measured as the percentage of chips produced that meet the required specifications and performance. TSMC, Samsung and Intel are the world's only chipmakers that can offer cutting-edge chip production at a large scale. TSMC has secured the majority of such orders, as well as orders for advanced chip packaging services, from the likes of Nvidia, Apple, AMD, Broadcom, Marvell and MediaTek. These clients take up a large share of the Taiwanese chipmaking group's advanced chipmaking capacity. Intel, which mainly manufactures chips for in-house use, is working to attract more customers to adopt its advanced chipmaking and packaging capacity. The Chinese auto industry executive said many chip designers in the country are now pursuing a dual-sourcing strategy, spreading orders between TSMC and Samsung rather than relying on a single foundry, with exact allocation depending on the product. "Geopolitical factors also push some U.S. clients to use multiple contract chipmakers, if available out there, if they can," another chip industry executive with knowledge told Nikkei Asia. "TSMC's advanced chip production capacity is really full, which is hard for Chinese clients who have smaller volumes to place new orders, so several have turned to approach Samsung for future possibilities," said an executive-level chip industry source who is familiar with the matter. Google has been a driving force in the multi-foundry strategy, sources told Nikkei Asia. MediaTek, Taiwan's top chip developer, is helping Google to work with both TSMC and Intel for advanced chip packaging, Nikkei Asia earlier reported, while Qualcomm, the U.S. mobile chip king, has long made use of both TSMC's and Samsung's services. However, using multiple foundries for advanced chip production requires substantial investment in adapting chip designs to different chipmakers' manufacturing processes, along with greater R&D and supply-chain coordination. Such a strategy is typically viable only for deep-pocketed chip developers, given the additional engineering and workforce costs involved. Meanwhile, capacity constraints at TSMC are pushing other existing customers to potentially place more orders with Samsung, sources said. Tesla works with both Samsung and TSMC on various products, including its AI5 chips used in its cars and robots. The upcoming AI6, however, will be made by Samsung in Texas. Tesla founder Elon Musk also announced a partnership with Intel for its future terafab projects in Texas, aiming to bring more critical chip production back to the U.S. and help meet growing demand. AMD and Nvidia, key clients of TSMC, are also working with Samsung on some chip projects due to capacity constraints. AMD, for example, is in talks to build some future CPUs with Samsung from 2028, one source with knowledge of the situation said. Nvidia's Groq language processing units (LPUs) are produced by Samsung, and the U.S. chipmaker has not decided whether production of the next generation of these highly specialized chips will be given to other contract chipmakers. The majority of Nvidia's chips are made exclusively by TSMC. The increasing requests for leading-edge chip production for Samsung would mark a victory for the South Korean chipmaker's foundry business, which used to mainly serve its own consumer electronics and home appliance divisions. The growing foundry business comes as Samsung is also riding the AI investment boom, enjoying exponentially growing demand for its memory products, including high-bandwidth memory chips, DRAM and NAND flash memories. Samsung said that it does not comment on customers. New orders would help Samsung's foundry unit rebound from a recent downturn. The company said earnings for the segment declined in the first quarter due to it being the off-peak season, but it did not elaborate on how big the dip was. Samsung's foundry and system large-scale integration business units posted a combined 6.9 trillion won ($4.6 billion) in sales in the first quarter, up from 6 trillion won a year ago. Samsung does not break down its earnings for the two units, the latter of which makes an array of logic and analog chips, including image sensors and display chips. BYD declined to comment. Google said it does not comment on the details of its supply chain.
16
94
18
796
143,158
seekinganythingbutalpha retweeted
港股智谱股价盘中一度涨超45%,目前涨36%。 在Anthropic最强模型受美国政府管制的背景下,智谱周末宣布旗下能力最强的开源模型GLM-5.2面向GLM Coding Plan用户全量开放,并将于本周按MIT协议开源,以吸引全球开发者。
亚马逊云服务(AWS):Anthropic已要求我们撤销所有用户在全部区域范围内对Claude Fable 5和Claude Mythos 5的访问权限。 此前美国商务部长卢特尼克致函Anthropic,将Fable 5和Mythos 5纳入出口管制,限制境外及境内外籍人士访问,理由是存在越狱安全风险。Anthropic虽遵令下架,但公开反对。
58
17
7
180
112,937
seekinganythingbutalpha retweeted
The data center default cycle is going to be like nothing we’ve ever seen
44
56
5
872
66,380
seekinganythingbutalpha retweeted
Replying to @Kenntnis22
北大马会不就这样干了吗?但很快就被中共“镇压”了。科技化的中共极权政府的暴力能力比朴、全年代的韩国威权政府要强力的多的多。想要完全靠内部努力来解决中共是不可想象的。所以我说世界如果不能解决中共问题,那么中共将有一天会“解决”全世界。
2
5
134
6,201
seekinganythingbutalpha retweeted
OpenAl is offering private-equity firms a guaranteed minimum return of 17.5%, as well as early access to models not yet in public release.
312
252
603
3,041
4,341,621
seekinganythingbutalpha retweeted
1925年上海的川菜馆 “菜甚美而价奇昂” 。 而国民革命(国民党北伐)胜利以后,从1920s年代后期到1930s年代后期这十年间上海川菜馆的歇业开业信息来看,大多因贵而歇而廉而兴。teapawt.org/sophyzhang2003/status/…
据1925年上海世界书局出版的《上海宝鉴》中记载,当时川菜馆的菜码相当丰富:炒肉片、椒盐虾糕、辣子鸡、炸八块、凤尾笋、松子山鸡丁、米粉牛肉、米粉鸡、白炙脍鱼、奶油广肚、红烧大杂烩、酸辣汤、清炖鲥鱼、红烧春笋、叉烧肉、火腿炖春笋、白汁冬瓜方、清炖蹄筋、鸡蒙缸豆、锅烧羊肉、蟹粉蹄筋、
1
5
2,226
seekinganythingbutalpha retweeted
事実としては、米国が世界最悪の侵略国家なのに、日本ではなぜか中国ということになっており、それを頑なに信じ込む人間が非常に多い。これもまた信仰宗教と呼べるものだろう。
101
669
57
3,270
90,850
seekinganythingbutalpha retweeted
🔥Seedance2.0制作的广告,顺利登陆福州地铁+福建电视开屏广告 刚回到福州就来看看自己制作的广告上墙效果,这真是一个神奇的体验,AI让我在西班牙给家乡制作一条广告,地域的服务限制被打破,一人公司的终极形态不就是到哪里都可以工作,我现在外加到哪里都可以随心所欲拍摄!
14
11
2
120
16,327
seekinganythingbutalpha retweeted
GPU availability falling across the board. Of note, B200 availability making new lows. Less availability = bullish demand.
22
88
23
617
129,302
seekinganythingbutalpha retweeted
上星期不是回了趟香港吗,说实话,还是很喜欢。 甚至都动了十年后回香港养老的念头。 我看到酒楼里面的老人家拿张报纸从早上看到中午,这就是我向往的惬意生活嘛。香港医疗水平很高的(当然要钱这另说)。 再说,我一个平凡老头带着老伴过平静日子,不惹政治,好想真可以。 谁能够给我泼点冷水,否则我开始着手准备了哈哈
69
3
2
98
42,873
seekinganythingbutalpha retweeted
I bet most people aren’t even using the basic Clawdbot ( now Moltbot) commands. the problem is, if you prompt everything and let the bot guess, you’re just wasting tokens and running in circles. there are a lot, but these are the ones you need, added some combos too below..
36
51
6
487
48,480