Manish Singh

🔥 Real Madrid spent $80M on Ronaldo, and Meta spent $100M to land Jiahui Yu from OpenAI. But the real game-changer? Trapit Bansal, an IIT Kanpur grad who went from Accenture analyst to IISc researcher to UMass Amherst PhD. He’s worked at Google, Facebook, Microsoft, and OpenAI, helped build the o1 reasoning model, and now joins Meta’s Superintelligence Labs with a rumored $100M payday. Meta didn’t stop there. They’ve poached a wave of core OpenAI talent: - Shengjia Zhao (ChatGPT and GPT-4 co-creator) - Jiahui Yu (o3, o4 mini, GPT-4.1) = Shuchao Bi and Hongyu Ren (GPT-4o creators) Meta is building its own $100M dream team. This isn’t just a hiring spree. It’s a strategic knockout move to drain OpenAI of its key people and leap ahead in AGI. 🧠 What this means: Meta is rewriting the roster with superstar engineers OpenAI risks serious talent loss even as it focuses on retention The AI battlefield now includes billion-dollar sign-on bonuses Bottom line: From an Accenture cubicle to Meta’s Superintelligence Labs, this high-stakes transfer saga could reshape the future of AI. Who’s on your fantasy team?

22

2 Comments

Soumith Chintala, 𝗰𝗿𝗲𝗮𝘁𝗼𝗿 of PyTorch and one of the most respected builders in 𝗺𝗼𝗱𝗲𝗿𝗻 𝗔𝗜, has joined Mira Murati ’s fast-rising startup, Thinking Machines Lab. After 11 years at Meta (Facebook), where PyTorch grew from a 𝗻𝗶𝗰𝗵𝗲 𝗽𝗿𝗼𝗷𝗲𝗰𝘁 𝘁𝗼 𝗽𝗼𝘄𝗲𝗿𝗶𝗻𝗴 𝗴𝗹𝗼𝗯𝗮𝗹 𝗔𝗜 𝗶𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻, Chintala’s move signals a new wave in human–𝗔𝗜 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻. Thinking Machines Lab isn’t just gathering top talent from Meta, OpenAI, Anthropic, and leading universities—they’re offering 𝗴𝗮𝗺𝗲-𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝘀𝗮𝗹𝗮𝗿𝗶𝗲𝘀 (up to $500K), 𝗮 𝗺𝗮𝘀𝘀𝗶𝘃𝗲 $𝟮 𝗯𝗶𝗹𝗹𝗶𝗼𝗻 𝘀𝗲𝗲𝗱 𝗿𝗼𝘂𝗻𝗱, and are aiming for a $50 billion market valuation. Their new product, Tinker, is already in use at Princeton, Stanford, and by major enterprises. Meta, meanwhile, is regrouping under the 𝗻𝗲𝘄 𝗦𝘂𝗽𝗲𝗿𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗟𝗮𝗯𝘀 division led by Alexandr Wang, with other AI leaders reportedly on the move. Chintala’s reasoning: time to pursue “𝙨𝙤𝙢𝙚𝙩𝙝𝙞𝙣𝙜 𝙨𝙢𝙖𝙡𝙡, 𝙨𝙤𝙢𝙚𝙩𝙝𝙞𝙣𝙜 𝙣𝙚𝙬, 𝙨𝙤𝙢𝙚𝙩𝙝𝙞𝙣𝙜 𝙪𝙣𝙘𝙤𝙢𝙛𝙤𝙧𝙩𝙖𝙗𝙡𝙚.” His legacy—the PyTorch tool now thriving in classrooms from 𝗠𝗜𝗧 𝘁𝗼 𝗿𝘂𝗿𝗮𝗹 𝗜𝗻𝗱𝗶𝗮—𝗶𝘀 𝘀𝗲𝗰𝘂𝗿𝗲. #AI #TechNews #PyTorch #Meta #ThinkingMachines #Startup #AIJobs #TalentWar #Innovation #BigFunding

8

I gave nine closed-source language models the brand-new JEE-Advanced 2025 paper—the toughest under-grad Physics-Chem-Math exam in India—and many of them out-scored the human AIR-1 (332/360)!! Why JEE-Advanced is a clean STEM benchmark - Fresh content → Paper was held on 18 May 2025, so it isn’t floating around training corpora. - No biology, no trivia → Pure PCM reasoning across 96 high-pressure questions. - Uncontaminated → GPQA and annual AIME are excellent, but their problems can leak into model training sets. JEE-Adv 2025 avoids that. Knowledge Cut-offs - OpenAI o3/o4-mini models → 1 Jun 24 - Gemini latest models → Jan 25 - Anthropic latest models → Mar 25 - JEE Advanced 2025 Paper date → 18 May 25 Only Gemini 2.5-Pro/Flash released after the exam date, so 7/9 models had zero chance to memorise these questions Implications - Deterministic PCM Q&A at JEE level is essentially **solved** for frontier models.   - Latency and unit-economics are now the real bottlenecks. - I'm expecting inference routers -- that hit a cheap, fast model first and fall back to a stronger model only when its confidence score is low -- to become the norm

665

68 Comments

Publication update from MISN Lab @ IIT Delhi! Our paper, "𝐌𝐨𝐝𝐮𝐥𝐚𝐫𝐢𝐭𝐲-𝐀𝐢𝐝𝐞𝐝 𝐂𝐨𝐧𝐬𝐢𝐬𝐭𝐞𝐧𝐭 𝐀𝐭𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐝 𝐆𝐫𝐚𝐩𝐡 𝐂𝐥𝐮𝐬𝐭𝐞𝐫𝐢𝐧𝐠 𝐯𝐢𝐚 𝐂𝐨𝐚𝐫𝐬𝐞𝐧𝐢𝐧𝐠," has been accepted at TMLR 2025 (@Transactions on Machine Learning Research)! This work is especially close to our hearts as it was: An outcome of the ✅ course project for Optimization for Machine Learning taught at IIT Delhi, ✅ Led by undergraduate students Yukti Makhija Samarth Bhatia of Indian Institute of Technology, Delhi and mentored by course TA and my PhD student Manoj kumar. 💡 This is a powerful testament to what motivated undergraduates can achieve when given the right platform, mentorship, and academic rigor. Huge congratulations for the outstanding work! Summary: This work makes 𝐟𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐜𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐢𝐨𝐧𝐬 to the theory and practice of graph clustering by introducing a modularity-driven coarsening framework with 𝐩𝐫𝐨𝐯𝐚𝐛𝐥𝐞 𝐜𝐨𝐧𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐠𝐮𝐚𝐫𝐚𝐧𝐭𝐞𝐞𝐬 under the 𝐃𝐞𝐠𝐫𝐞𝐞-𝐂𝐨𝐫𝐫𝐞𝐜𝐭𝐞𝐝 𝐒𝐭𝐨𝐜𝐡𝐚𝐬𝐭𝐢𝐜 𝐁𝐥𝐨𝐜𝐤 𝐌𝐨𝐝𝐞𝐥 (𝐃𝐂-𝐒𝐁𝐌). The proposed method integrates structural and feature-based information through a carefully designed optimization objective and seamlessly plugs into GNN and VGAE architectures. Extensive experiments across benchmarks demonstrate state-of-the-art performance, scalability, and theoretical consistency—pushing the frontiers of unsupervised learning on graphs. Link: https://lnkd.in/gWEcTZGg #TMLR2025 #IITDelhi #MISNLab #UndergraduateResearch #MLTheory #GraphClustering #GraphML #OptimizationForML #GNN #CourseToConference #AIResearch #StudentLedInnovation

139

9 Comments

Adaptive test models are known to give disproportionate weight to early responses. A momentary lapse of concentration, a misread question, or even a technical glitch can push a student onto a permanently different trajectory. In a traditional exam, one wrong answer costs one mark. In an adaptive system, it can silently reshape the entire test. #jeeadvanced #iit https://lnkd.in/dJp4y7md

8

This post is about a third paper in ICLR led by Roman Belaire, Ph.D. student of Pradeep Varakantham. Key insights in this paper came out from discussions at AAAI, which is anecdotal evidence of why in-person conferences are so useful. Title: On Minimizing Adversarial Counterfactual Error in Adversarial Reinforcement Learning This work addresses a fundamental challenge in adversarial reinforcement learning (RL) with state perturbations: the partial observability of the state due to adversarial perturbations. Unlike prior approaches, we directly account for partial observability by introducing a novel objective—Adversarial Counterfactual Error (ACoE)—that balances value optimization with robustness. To ensure scalability in model-free settings, we propose a theoretically grounded surrogate objective, Cumulative-ACoE (C-ACoE). The mechanism quantifies the discrepancy between the agent’s belief state and the true state by providing a belief structure over expected adversarial outcomes, and that also makes C-ACoE satisfy the Bellman optimality structure and hence C-ACoE can be incorporated in any standard RL algorithm. Our empirical evaluations on standard benchmarks including MuJoCo, Atari, and Highway demonstrate that our method is the state-of-the-art in adversarial RL with state perturbations. https://lnkd.in/eKBGkkui

41

Needless to say, state-of-the-art LLMs are fantastic at code generation. Benchmarks like HumanEval, MBPP, APPS, and SWE-Bench evaluate these capabilities. In the hardware design space, we have VerilogEval and RTLLM. While they’re a good start, the design problems are fairly rudimentary. We recently released ArchXBench, a new benchmark suite with complex problems such as ML accelerator designs, crypto engines, and more. This work will be presented at #MLCAD2025. Paper: ArchXBench: A Complex Digital Systems Benchmark Suite for LLM-Driven RTL Synthesis → [arXiv link](https://lnkd.in/g45zSpP4) Benchmark suite: [GitHub link](https://lnkd.in/gpyvG3WH) #MLCAD2025 will be in Santa Cruz, CA, on September 8–10. I’ll be in the Bay Area from the 7th to the 11th—would be happy to catch up with any #IIIT alumni. PS: Learning to write posts on LinkedIn and experimenting with tagging, and simple things like making text bold. Please excuse me if this one looks a bit clumsy. Wouldn’t it be nice if there were a preview feature? Any suggestions on this are welcome.

42

1 Comment

“Bhai, tune physics kyun li?”... I bumped into an old acquaintance today at IIT Bombay. We used to prep together for JEE, and he had scored well enough to grab a mechanical engineering seat here. After a two-minute nostalgia trip, he dropped the line: “Bhai, tune physics kyun li? Tu to accha padhta tha..” (Bro, why did you choose Physics? You were good at studies..) Two things hit me. One, the casual judgment hidden in those words — as if my choice to study something I genuinely loved was somehow a misstep. And two, the absurdity of how society has trained us to measure intelligence and ambition only by the degree or career path you choose. If I am not pursuing a degree that doesn't have B. Tech or MBBS written in front of it, then is it worthless? People often forget that engineering is an applied form of physics. Yet somehow, pure sciences get tagged as “impractical” or “less marketable.”  Meanwhile, the very breakthroughs that define our world — AI, quantum computing, and quantum information theory — are rooted in these pure sciences. These fields promise to revolutionize industries by solving complex problems beyond the reach of classical computers. As AI and machine learning continue to evolve, the need for quantum-enhanced algorithms becomes more pressing. Consider the 2024 Nobel Prize in Physics, awarded to Geoffrey Hinton and John Hopfield for their groundbreaking work in artificial neural networks, foundational to modern AI. Hinton, often dubbed the "Godfather of AI," developed the Boltzmann machine, a concept rooted in statistical physics. Their work has propelled AI advancements, from chatbots to medical diagnostics.  According to a report by NASSCOM, the Indian IT industry is experiencing a slowdown, with growth rates dipping below 10% in recent years. Software development, web development, and all other coding jobs are becoming crowded fields. Prestigious branches like CSE are seeing saturating job markets, partly because everyone chases the same “hot” degrees, and partly because AI is automating a growing chunk of these roles. Soon, machine learning and Generative-AI might face the same fate. And who will come to the rescue? The scientists, the researchers, the curious minds who never just followed the mainstream path. Choosing pure sciences isn’t a retreat from opportunity — it’s an investment in a future that no one can yet fully imagine. It’s in these fields that revolutions are born. While engineering applies existing knowledge, pure sciences explore the unknown, laying the groundwork for innovations that can reshape our world. So, the next time someone questions your choice of studying Physics, remember: In a world obsessed with career metrics, sometimes curiosity and genuine interest are the boldest choices we can make.

587

41 Comments

A team of researchers from IIT Bombay led by Prof. Gopal Dixit have found a way to use light to control and read tiny quantum states inside atom-thin materials. The simple technique could pave the way for computers that are dramatically faster and consume far less power than today’s electronics. https://lnkd.in/dARyH2rm

341

5 Comments

⚡ ML inference continues to fascinate me, and I've been diving deep into LLM inference optimization techniques - something I genuinely enjoy working on! Recently, I've been exploring the intricate challenges of serving large language models efficiently, and it led me to build TorchWeave LLM - an open-source inference compiler that achieves 2-5x throughput improvements through some compelling optimization strategies. What makes ML inference so interesting is how it sits at the intersection of systems engineering and machine learning theory. Every optimization decision has cascading effects on memory, compute, and user experience. 🔄 Continuous Batching: One technique I found particularly rewarding to implement - merging concurrent requests into shared decode steps to maximize GPU utilization. The scheduling algorithms here are genuinely fascinating. ⚡ Per-Request KV-Cache Management: Exploring how attention masks can enable efficient memory management for variable sequence lengths opened up entirely new optimization possibilities. 📡 Real-time SSE Streaming: Building Server-Sent Events with time-to-first-token metrics taught me so much about balancing performance with user experience. The exploration paid off: 98% throughput improvement with 4 concurrent requests Average TTFT of 0.860s Production-ready architecture that scales What I love about working on inference optimization is that every bottleneck you solve reveals three new interesting problems. The rabbit holes of memory management, scheduling algorithms, and distributed systems design never get old. The technical implementation spans everything from PyTorch model interfaces to FastAPI async patterns - each component was a learning opportunity that deepened my appreciation for the complexity of production ML systems. 🔗 Check it out on GitHub: https://lnkd.in/gCZdZ5Pm I'm curious - what aspects of ML inference optimization do you find most compelling? Always eager to learn from others tackling similar challenges! #MachineLearning #LLM #MLInference #PyTorch #Optimization #OpenSource #MLOps #AI #SoftwareEngineering #TechnicalExploration Built with Python 3.12, FastAPI, PyTorch, and genuine curiosity about making models faster ⚡

49

6 Comments

Many of my IITB friends who went for PhDs in the US had years of study while all of us started earning at work Our incomes grew while they went along on small student researcher compensations One of them just called me to ask what I think about a $10M offer from a large AI lab Patience paying big time

2,236

69 Comments

LLMs face a mathematical limit that no amount of scaling can fix. A paper by Vishal Sikka (former SAP CTO, McCarthy's student) and Varin Sikka makes the case: https://lnkd.in/gUh2dxUi Here's the argument: - One forward pass of a transformer costs O(n²·d) operations - There exist problems requiring MORE than O(n²·d) to solve - Therefore, LLMs cannot correctly solve those problems But how do we know harder problems exist? The proof relies on a 130-year-old idea. In 1965, Hartmanis and Stearns proved the Time Hierarchy Theorem using diagonalization — a technique dating back to Cantor. The proof (see animation): Imagine a table where rows are machines and columns are inputs. Each cell shows what machine Mᵢ outputs on input n. The construction: for any time bound T(n), we can enumerate ALL machines that halt within T(n) steps — there are only finitely many distinct programs of a given size, and we only care about those fast enough to finish in time. Now construct a new sequence D that differs from machine Mᵢ on input i (just flip the bit on the diagonal). D cannot be computed by ANY machine in the table. It's designed to disagree with each one somewhere. Give the machines more time? You can always construct a new diagonal sequence that escapes them too. This creates an infinite hierarchy of complexity classes, each strictly more powerful than the last. O(n) < O(n²) < O(n³) < ... < O(2ⁿ) < ... Each jump unlocks problems impossible at the previous level. Matrix multiplication, all-pairs shortest paths, CFD simulations — their complexity is above O(n²). What this means for AI: - Agentic systems doing payments, logistics, industrial control often hit >O(n³) subproblems and hallucinate solutions - Verification can be harder than solving (checking a TSP tour optimal = comparing against (n-1)!/2 alternatives) - "Thinking" traces don't escape — they just add tokens, not computational class (https://lnkd.in/gCbpEGrb) The implication: hybrid architectures — LLMs + symbolic modules + external solvers. Progress may come from models that know when to delegate to algorithms that CAN solve the hard problems. Terminal 3 #ai #llm #complexity #computerscience #agents

13

2 Comments

I will be presenting at the NeurIPS LLM Evaluation workshop this weekend with Mohit Singh Chauhan, Ph.D., sharing our recent work on uncertainty quantification for language models 🤖 The work is based on our paper “Uncertainty Quantification for Language Models: A Suite of Black-Box, White-Box, LLM Judge, and Ensemble Scorers,” recently published in Transactions on Machine Learning Research (TMLR) 📄 We outline a general framework for closed book hallucination detection in real world use cases by adapting black box, white box, and LLM judge style UQ methods into standardized 0 to 1 confidence scores, and introducing a tunable ensemble that lets practitioners combine and optimize these scores for their specific use case. In experiments on several QA benchmarks, the ensemble typically surpasses its individual components and outperforms existing hallucination detection methods. The full suite of scorers is implemented in our open source Python library UQLM (link in comments). If you will be at NeurIPS and are interested in hallucination detection, reliability, or LLM evaluation, feel free to stop by our poster or say hi! The TMLR version of the paper is attached below. #NeurIPS2025 #LLMEvaluation #HallucinationDetection #TMLR

148

11 Comments

Full article: https://lnkd.in/gyFXKeXD World-class research shouldn't be held back by hard-to-get tools. That’s the belief behind InSituMicron founded by Mayank Mishra, Kousik Papakollu, and Lalith Bhaskar who are building compact material testing systems that bring precision into every lab, no matter the size or budget. We got a chance to interview Mayank Mishra CMO of InsituMicron. In this story, we uncover how InSituMicron is tackling the barriers to material testing in Indian research. Journalist: G Reshma #deeptech #indianstartups #materialscience #researchinnovation #prodwrksinsights #hardwarestartups #founderstory

1

A team from MIT and UCSD including MIT Mathematics Assistant Professor Adityanarayanan "Adit" Radhakrishnan have devised a way to test whether LLMs hold hidden biases, moods, and abstract concepts, marking a major step towards creating models which are safe yet effective. Read more at MIT News: https://lnkd.in/ejTsqSzU

9