San Francisco Bay Area
2K followers
500+ connections
San Francisco Bay Area
2K followers 500+ connections
Activity
2K followers
Recommendations received
View Nitishโs full profile
Join to view full profileOther similar profiles
Explore more posts
-
Apala Guha
Microsoft โข 2K followers
Ruokai Yin, our summer intern from Yale to Microsoft Azure, worked with Xuan Zuo, @Sattwik Deb Mishra, Hokchhay T. and Preyas Shah and me. We published a paper based on his project: https://lnkd.in/eHuG3BSt Summary of paper: Finding optimal sharding is hard especially for large LLMs. A large number of devices have to be divided into TP, PP, EP dimensions. For TP, we have to decide the tensor dimension e.g. head, hidden size to shard for every op. Additionally, TP can be one-dimensional/two-dimensional etc. This makes for a large sharding space that is hard to cover exhaustively. SoTA methods use heuristics to select sharding e.g. GSPMD, Sequence Parallelism, Helix etc. There are significant time gaps, months or years, between the discovery of each of these methods. We want to fast-forward the process of discovery by 1) using a formal representation of the sharding space, and, 2) using reinforcement learning to search the sharding space. This paper demonstrates that our method is 3.5x better than random search, and on par with hand-tuned heuristics when looking only at 1D sharding and without allowing sequence parallelism. We aim to extend the work to multi-dimensional sharding as well as sequence parallelism. Beyond that, we wish to combine the problem of sharding with fusion, tiling and collectives as well.
-
๐ง๐ต๐ฒ ๐ฏ๐ฒ๐๐ ๐๐ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐น๐ฒ๐๐๐ผ๐ป ๐ฐ๐ผ๐๐ ๐ผ๐๐ฟ ๐ฒ๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ $๐ฒ๐ฌ๐ฌ ๐ฎ๐ป๐ฑ ๐ฎ ๐บ๐ฎ๐ฟ๐ฟ๐ถ๐ฎ๐ด๐ฒ ๐ฎ๐ฟ๐ด๐๐บ๐ฒ๐ป๐. Our VP of Engineering, Xiaofan(James) Luan, was supposed to buy his wife a Dior bag for their anniversary. Instead, he bought three Claude Code subscriptions and spent the holiday trying to cross-compile 2 million lines of C++. Every fix on one platform broke two others. $600 later, the only output was "git reset --hard" โ and a very cold dinner table.๐ "Make it compile on Windows" is a trap. The real goal was "compile everywhere without hacks" โ no AI is going to figure that out for you at 2 am. What worked: constraints before code, review tests not code, bottom-up, one layer at a time. Same task, two days. Then he ran six parallel Claude sessions across three machines with git worktree. The bottleneck stopped being intelligence and started being how fast one person can alt-tab. AI solves exactly the problem you give it. Engineering is in knowing which one to give. His wife is still waiting for that bag. Full story: https://lnkd.in/gtsW_Wvk โโโ Follow Milvus, created by Zilliz, for everything related to unstructured data
-
Milvus, created by Zilliz
13K followers
๐ง๐ต๐ฒ ๐ฏ๐ฒ๐๐ ๐๐ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐น๐ฒ๐๐๐ผ๐ป ๐ฐ๐ผ๐๐ ๐ผ๐๐ฟ ๐ฒ๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ $๐ฒ๐ฌ๐ฌ ๐ฎ๐ป๐ฑ ๐ฎ ๐บ๐ฎ๐ฟ๐ฟ๐ถ๐ฎ๐ด๐ฒ ๐ฎ๐ฟ๐ด๐๐บ๐ฒ๐ป๐. Our VP of Engineering, Xiaofan(James) Luan, was supposed to buy his wife a Dior bag for their anniversary. Instead, he bought three Claude Code subscriptions and spent the holiday trying to cross-compile 2 million lines of C++. Every fix on one platform broke two others. $600 later, the only output was "git reset --hard" โ and a very cold dinner table.๐ "Make it compile on Windows" is a trap. The real goal was "compile everywhere without hacks" โ no AI is going to figure that out for you at 2 am. What worked: constraints before code, review tests not code, bottom-up, one layer at a time. Same task, two days. Then he ran six parallel Claude sessions across three machines with git worktree. The bottleneck stopped being intelligence and started being how fast one person can alt-tab. AI solves exactly the problem you give it. Engineering is in knowing which one to give. His wife is still waiting for that bag. Full story: https://lnkd.in/gtsW_Wvk โโโ Follow Milvus, created by Zilliz, for everything related to unstructured data
-
H-1B Workers AND Layoffs Senators Want Answers From Big Tech on H-1B Workers, Layoffs Letters were sent to Amazon, Apple and other big companies from top lawmakers on Judiciary Committee The H-1B system, created in 1990, is the primary pathway for foreign professionals to work in the U.S. Roughly 700,000 people currently live in the country on H-1B visas, according to a National Foundation for American Policy analysis of government data, with technology firms among the biggest users. Most H-1B holders are from India and China. Amazon, the largest H-1B sponsor in the U.S., won approval for more than 14,000 new hires on the visa in fiscal 2025โthe most of any companyโeven as it has announced layoffs affecting tens of thousands of jobs in recent years. โWith all of the homegrown American talent relegated to the sidelines, we find it hard to believe that Amazon cannot find qualified American tech workers to fill these positions,โ the senators wrote to Amazon CEO Andy Jassy. Aside from Amazon, Apple, and JP Morgan, the other companies receiving letters included Deloitte, Alphabetโs Google, Meta, Microsoft, Walmart, Cognizant Technology Solutions and Tata Consultancy Services. The firms were asked to respond by Oct. 10.
-
Early Adopters
632 followers
If you work in tech, this weekโs layoffs might have hit you or someone close. Entire teams across startups and big companies were cut, from engineers at Intel to product and voice teams at Google. Even AI-first companies like Scale AI are shrinking operations and restructuring fast. But when you look at whoโs being let go, a pattern starts to emerge. The roles affected are often tied to legacy systems, manual workflows, and older product lines. Hardware teams, generalist PMs, and support ops are being replaced or reorganized. At the same time, hiring is picking up around model infrastructure, agent teams, and internal tooling that can scale with fewer people. Founders and CEOs are rethinking how their companies are built. This shift is not just short-term pressure. It reflects a change in how businesses are choosing to grow and who they are planning to grow with. What are you noticing? Follow Early Adopters for more content on founders, startups and business leaders. We will talk about which roles are rising in a second next video.
-
Phanindra Reddy
MAGICTEAMS.AI โข 9K followers
As a founder who want to be in the US for building my company, this is what it meant there is insane panic over the h1b issue, here are the updated facts you should know about - this only applies to the new applications, so if you already in h1b , you are good to go and your renewal will not be changed to 100k, it is the old pricing - if you are already a student studing with f1 visa with otp, you are under old pricing -this fee increase is not annual but is charged per approved petition. - for people who change jobs with transfer, you get the previous pricing, even if you re enter the us , there is no possible change - but if you are coming to study after september 21, aka today, when your employer applies h1b, you are in the new slab , paying 100k usd - same with people coming over directly via h1b and starting a new h1b in the usa from a new employer tldr: people already in the us, either in job or student , there is nothing to worry about, as you are in the old slab for new ones, it is hard as it gets what about founders and startups who want to be in SF for building in AI bootstraping - i want to self fund my h1b , but as a new application, we need to pay the 100k - the better one is O1, which is hard to get, and need to show abilities and funding for the startup - L1, but there is significant infra and process needed for it - or give a 800k - million dollars for EB 5 the only option i see here is get biz upto a million or more, and self sponser this might work for established players, but for startups , if there is a better option, that will help us to build
-
Patrick Tammer
Google โข 6K followers
Micron, SK Hynix, and Samsung stocks are soaring but few understand why. 1. The structural reason: - Memory, in particular High-Bandwith Memory (HBM) has become crucial to run LLMs for billions of users - Running LLMs is mostly is memory-bandwidth bound, not compute-bound - During decode, GPUs spend more time fetching weights and KV cache than doing math, making HBM the primary bottleneck 2. The supply chain reason: - As demand soared, the major players shifted production capacity to high-margin HBM - That led to undersupply of other memory types (SRAM, DRAM) which are still needed for AI What this means forโฆ 1. Business leaders Memory cost will drive up GPU pricing Even if you don't buy chips, AI infra costs will likely rise as supply chain players will pass on costs 2. Entrepreneurs After decades of silence, there is massive opportunity in innovating memory Its still overlooked by many but we will soon see more high valuation memory startups which will become attractive acquisition targets for the 3 big incumbents ๐ท: FT โฆ Did you find this helpful? โป๏ธ Repost this to inform your network ๐ Follow me for more AI insights ๐ Subscribe to my newsletter
-
Vikas Rana
PPace โข 2K followers
๐ฏ From IIT Kanpur to Metaโs AGI Frontier Trapit Bansal, an IIT Kanpur alum with impressive stints at Google, Microsoft, and OpenAI, has joined Metaโs newly launched Superintelligence Labs, sharing his excitement on X: โThrilled to be joining Meta! Superintelligence is now in sightโโฏ Bansal helped build OpenAIโs reinforcement-learning and reasoning foundation models before switching to Meta, signaling the depth and pace of Zuckerbergโs AI talent pushโฏ. ๐งฉ Why this matters: Meta is doubling down on AGI, rebuilding momentum through high-impact hiresโฏ. It's a clear wake-up call for competitors (OpenAI, Anthropic, DeepMind): the war for top AI minds is intensifying. ๐ Points for discussion: Will adding star researchers translate into actual AGI breakthroughs, or are we just inflating a talent arms race? How will this influx shape Metaโs AI culture and speed of innovation? Source: https://lnkd.in/gGcqeUat
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content