Saikat Banerjee

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.

58

My first job was writing bidding algorithms for the Clearing House of India. In C. On Unix. I was never great at remembering syntax. But I understood the systems — the clearing logic, the bidding flows, how the pieces fit together. Today I use Claude Code with a team-of-agents framework. The syntax problem? Gone. What used to take days takes hours. But here’s what nobody is talking about: AI made building faster. Building was never the real bottleneck. I manage supply chain transformation programs across 4+ P&Ls at GE. What I see every day: teams in silos building overlapping algorithms independently. Systems that technically connect but don’t cohere. Rising opex. Painful integration. AI doesn’t fix this. AI makes it worse. Each silo now produces more divergent solutions, faster. McKinsey’s research confirms it — 89% of organizations still run industrial-age operating models. Only 1% operate as decentralized networks. And 42% of companies scrapped most AI initiatives in 2025, up from 17% the year before. Conway’s Law told us this in 1967: siloed teams will always produce siloed systems. So I wrote a whitepaper proposing a fix — the CORE Framework (Coherent Organizational Redesign for Enterprise AI). Five layers: 1. Shared Contracts — common data models before anyone builds 2. Human Orchestrator — cross-cutting authority, not just visibility 3. AI Governance — because AI-generated code loops, duplicates, and over-solves 4. Team Redesign — organize by value stream, not technology layer 5. Opex-Aware Delivery — measure production health, not shipping speed The organizations that win with AI won’t adopt fastest. They’ll restructure to absorb properly. Full framework on Medium: https://lnkd.in/df76HJUf Would love to hear — are you seeing AI accelerate the silo problem in your organization? #AI #ConwaysLaw #OrganizationalDesign #SupplyChain #DigitalTransformation #AgenticAI

11

By enabling locally trained large language models (LLMs) that work seamlessly across India’s diverse tongues, this initiative unlocks AI’s potential where it matters most—e-commerce, government services, and financial transactions—the daily touchpoints of hundreds of millions of Indians. This isn’t just an infrastructure push; it’s a path to mass inclusion, scaled innovation, and AI adoption that directly benefits the people.

16

A new study from Carnegie Mellon, “Does AI-Assisted Coding Deliver? A Difference-in-Differences Study of Cursor’s Impact on Software Projects” that looks at a more recent period, tracking code quality in projects using GenAI tools up to mid-2025, finds that AI briefly accelerates code generation, but the underlying code quality trends continue to move in the wrong direction. https://lnkd.in/gm6dxb8W

2

5 Comments

From writing code to leading global teams in the AI era 👨‍💻➡️🌍 Just published a deep dive into the leadership journey of Suresh Ganjigunta, an Engineering Leader at Amazon, whose path from individual contributor to strategic leader offers invaluable insights for anyone navigating tech leadership today. Key takeaways that resonated with me: 🔹 Leadership found him before he found leadership - Sometimes the best leaders emerge naturally by taking initiative and influencing outcomes as peers 🔹 Thriving in ambiguity is a superpower - Whether re-architecting payments platforms or building multimodal AI experiences, the ability to create frameworks for uncertainty is crucial 🔹 Innovation belongs in the everyday - Rather than treating innovation as separate initiatives, integrate it into daily problem-solving and team workflows 🔹 Go deep AND go wide - Technical expertise must be paired with cross-functional relationships across product, design, and program teams What struck me most: His approach to AI leadership - from early computer vision experiments to complex LLM-powered systems - shows how curiosity and structured experimentation can drive breakthrough innovations. His philosophy that "AI isn't about replacement, it's about amplification" feels especially relevant as we all navigate this rapidly evolving landscape. For aspiring tech leaders: The intersection of customer understanding and technical expertise is where the most impactful leadership happens. What's your experience transitioning from IC to leadership? What challenges have you faced in scaling teams across different cultures and time zones? Read the full story: https://lnkd.in/g8uK85jV #TechLeadership #EngineeringManagement #AI #Amazon #Leadership #Innovation #CareerGrowth #TechCareers #GlobalTeams #ArtificialIntelligence

17

3 Comments

𝗕𝗲𝘆𝗼𝗻𝗱 𝗧𝗲𝘅𝘁 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻: 𝗘𝗺𝗽𝗼𝘄𝗲𝗿𝗶𝗻𝗴 𝗟𝗟𝗠𝘀 𝘄𝗶𝘁𝗵 𝗘𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗙𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝗖𝗮𝗹𝗹𝘀 𝗞𝗶𝗿𝗮𝗻 𝗣𝗿𝗮𝗸𝗮𝘀𝗵's insightful article on function calls sheds light on how LLMs can be instructed to construct data structures that describe external function calls. Crucially, the LLM doesn't execute these calls directly. Instead, it intelligently generates the blueprint for the action, paving the way for seamless integration with external tools and APIs. This indirect execution mechanism is key for security and control, allowing developers to orchestrate complex workflows and extend the capabilities of LLMs in a safe and structured manner. Definitely worth a read for anyone interested in the cutting edge of LLM applications and their potential to drive real-world impact! #LLMs #AI #MachineLearning #FunctionCalls #ExternalSystems #Innovation

3

Couldn’t agree more. In any technology spectrum, there are always two extremes—and the real value for most people often lies in finding that optimal midpoint. In AI, that’s where the benefits of decentralized intelligence truly shine.

5

I recently found myself exploring about an interesting AI system design question: If Sanskrit and English had equal amounts of digital content available to train large language models (LLMs), would Sanskrit be the better language because of its well-defined grammar and logical structure? Sanskrit is often praised for its systematic grammar, explicit case markings, and structured sentence construction, where grammatical roles like subject and object are clearly encoded in word endings rather than inferred from word order. Intuitively, that feels like it should make reasoning easier for AI systems. However, modern LLMs are not rule-based systems that “apply grammar” the way humans learn it — they are statistical models that learn patterns from massive data. English, despite its ambiguities and looser structure, has simpler morphology and fewer surface variations per word, which aligns well with today’s Transformer-based architectures and tokenization methods. Sanskrit’s rich inflections and combinatorial forms, while logically elegant, can increase modeling complexity unless the architecture is designed to leverage that structure explicitly. This leads to a deeper realization: language structure alone doesn’t determine AI capability — architecture design and representation strategy matter just as much. It also raises a bigger question for the future of AI: as we move toward neuro-symbolic and reasoning-focused systems, could structurally explicit languages offer new advantages?

22

2 Comments