Arvind Yadav

In this 20-minute talk, PolyMage Labs founder Uday Reddy Bondhugula talks about why and how one should build effective compilers for AI programming frameworks: https://lnkd.in/gWUQNUsF

79

Everyone’s racing to optimize LLM inference. Quantization. Pruning. Speculative decoding. PagedAttention. Dynamic batching. KV-cache optimization. Continuous batching. Mixture of experts. Chain-of-thought efficiency. Brilliant engineering. Also: all workarounds. We’re squeezing efficiency out of hardware that wasn’t built for this workload. Here’s the pattern: take probabilistic, time-evolving computation (what LLMs actually are), force it into Von Neumann’s static state paradigm, then spend engineering cycles converting it back into something useful. Collapse. Process. Reconstruct. Every single cycle. Each conversion costs power. Each conversion costs latency. Each conversion costs thermal budget. ATOM-iK asks a different question: what if the hardware aligned with the computation instead? What if register locations were treated as a state array where only deltas move between cycles? What if probabilistic, fuzzy inference ran natively—not as an optimization target, but as the natural language of the architecture? That’s not a trick. That’s not a technique. That’s alignment. No quantization needed. No speculative decoding workarounds. No chain-of-thought inefficiencies. Because the hardware was built for exactly this: time-evolving, probabilistic, spatially-distributed computation. You get: ∙ Faster inference — no conversion tax ∙ Lower power — no wasted state collapse cycles ∙ Better accuracy — fuzzy logic runs native, not quantized ∙ Scalable — thousands of cores per chip, minimal thermal load The inference optimization race is real. The solutions are clever. But they’re optimizing the wrong thing. The next generation of AI infrastructure won’t optimize inference on Von Neumann. It will replace Von Neumann with architecture that is inference. That’s what we’re building. #ATOMiK #LLM #AIInference #ComputeArchitecture #Hardware #PhysicsAligned #Quantization #EdgeAI #Optimization #VonNeumann

5

Sarvam-M: A 24B Parameter Multilingual AI Model is finetuned Mistral-Small with Indic language desi nuances covering 11 major languages from india 🇮🇳 Key Features: • Dual-mode interface: Quick "non-think" responses + detailed "think" mode for complex reasoning 🧠 • Strong performance on math (GSM-8K) and coding (SWE-Bench) benchmarks 📊 • Supports both native Indic scripts and romanized text ✍️ • Built for real-world multilingual conversational agents 💬 I think it's a good start , instead of training model from scratch Sarvam focused on building indic dataset and model like in video shows good translingual conversational capability . Try it out via the Colab link - https://lnkd.in/gzcW4hru Playground - https://lnkd.in/gXijZrkY

27

1 Comment

Quantization in Large Language Models (LLMs) As LLMs scale, intelligence is no longer the bottleneck — memory, latency, and cost are. Quantization turns out to be one of the most powerful (and misunderstood) techniques enabling real-world deployment of large models. In this blog, I break down: • Why LLMs are surprisingly quantization-friendly • INT8 vs INT4 vs NF4 — what actually works in practice • How modern architectures absorb low-bit noise • Where quantization fails and how to design around it Quantization is no longer an optimization step — it’s foundational infrastructure for production LLM systems. Link: https://lnkd.in/gqGje75G #LLMs #AIEngineering #Quantization #MachineLearning #GenAI #MLOps

9

2 Comments

📝 Announcing our EMNLP 2025 paper that introduces the Alignment Quality Index (AQI), a metric leveraging latent geometric representations and clustering indices to diagnose hidden misalignments in LLMs. 🔹 “𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐈𝐧𝐝𝐞𝐱 (𝐀𝐐𝐈) – 𝐁𝐞𝐲𝐨𝐧𝐝 𝐑𝐞𝐟𝐮𝐬𝐚𝐥𝐬: 𝐀𝐐𝐈 𝐚𝐬 𝐚𝐧 𝐈𝐧𝐭𝐫𝐢𝐧𝐬𝐢𝐜 𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐃𝐢𝐚𝐠𝐧𝐨𝐬𝐭𝐢𝐜 𝐯𝐢𝐚 𝐋𝐚𝐭𝐞𝐧𝐭 𝐆𝐞𝐨𝐦𝐞𝐭𝐫𝐲, 𝐂𝐥𝐮𝐬𝐭𝐞𝐫 𝐃𝐢𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞, 𝐚𝐧𝐝 𝐋𝐚𝐲𝐞𝐫 𝐰𝐢𝐬𝐞 𝐏𝐨𝐨𝐥𝐞𝐝 𝐑𝐞𝐩𝐫𝐞𝐬𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧𝐬” 🔹 In collaboration with Manipal University Jaipur, LinkedIn, Indian Institute of Technology, Kharagpur, Indian Institute of Information Technology Guwahati, Texas A&M University, VIT Chennai, Evalueserve, New York University, Birla Institute of Technology and Science, Pilani (Pragya Lab) 🔹 Published in EMNLP 2025 Main 🔹 Paper: https://lnkd.in/g54bEBCu ✍🏼 Authors: Abhilekh Borah, Chhavi Sharma, Danush Khanna, Utkarsh Bhatt, GURPREET SINGH, Hasnat Md. Abdullah, Raghav Kaushik R, Vinija Jain, Jyoti Patel, Shubham Singh, Vasu Sharma, Arpita V., Rahul Raja, Aman Chadha, Dr. Amitava Das ➡️ 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬 𝐨𝐟 𝐨𝐮𝐫 𝐀𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐈𝐧𝐝𝐞𝐱 (𝐀𝐐𝐈): 🧬 𝑰𝒏𝒕𝒓𝒊𝒏𝒔𝒊𝒄 𝑳𝒂𝒕𝒆𝒏𝒕 𝑮𝒆𝒐𝒎𝒆𝒕𝒓𝒚 𝑴𝒆𝒕𝒓𝒊𝒄: AQI measures alignment by assessing how distinctly safe vs. unsafe prompts cluster in latent activations, using combination of Davies‑Bouldin, Dunn, Xie‑Beni, and Calinski‑Harabasz indices—making it invariant to decoding style and resilient to alignment faking. 🧠 𝑳𝒂𝒚𝒆𝒓𝒘𝒊𝒔𝒆 𝑷𝒐𝒐𝒍𝒆𝒅 𝑹𝒆𝒑𝒓𝒆𝒔𝒆𝒏𝒕𝒂𝒕𝒊𝒐𝒏 𝑳𝒆𝒂𝒓𝒏𝒊𝒏𝒈: we use a sparse learned pooling mechanism across transformer layers to capture alignment‑relevant abstractions without modifying the base model—enabling internal safety diagnostics. 🚨 𝑬𝒎𝒑𝒊𝒓𝒊𝒄𝒂𝒍 𝑭𝒂𝒊𝒍𝒖𝒓𝒆𝒔 𝒐𝒇 𝑩𝒆𝒉𝒂𝒗𝒊𝒐𝒓𝒂𝒍 𝑴𝒆𝒕𝒓𝒊𝒄𝒔: AQI uncovers misalignments that behavioral metrics like G‑Eval or refusal rates miss—especially under jailbreaks, safety‑agnostic fine‑tuning, and stochastic decoding, proving its value as an early‑warning alignment audit tool.

117

2 Comments

🚀 How do we speed up AI-powered scientific discovery without giving up control? At #TPC25, ORNL’s Prasanna Balaprakash brought bold answers to the table, showcasing how AI can guide scientific research with unmatched speed and precision. Read more in HPCwire ⤵️

7

Every time you use ChatGPT, Claude or Gemini… remember this: The Transformer that powers them was co invented by Ashish Vaswani & Niki Parmar 🇮🇳 The PyTorch framework most LLMs are trained on? Soumith Chintala 🇮🇳 Scaling deep learning? Anima Anandkumar & Inderjit Dhillon 🇮🇳 Multilingual NLP foundations? Kalika Bali & Pushpak Bhattacharyya 🇮🇳 Indian origin researchers aren’t just “involved” in AI. They are the architects of the revolution. So next time someone calls Indians “outsiders”… Remind them of the Indian contribution. 🇮🇳🔥 👆 Posted by Sidharth via the 𝕏 platform Most claims check out from reliable sources like Wikipedia and academic pages: - Historical inventions (zero, algebra, trig, chess, shampoo, Ayurveda, yoga, plastic/cataract surgery, buttons, ruler, diamond mining) originated in ancient India. - Fiber optics pioneered by Indian-origin Narinder Kapany; it's optical tech, not a direct WiFi precursor. - AI: Vaswani & Parmar co-authored Transformer paper; Chintala created PyTorch; Anandkumar & Dhillon advanced DL scaling; Bali & Bhattacharyya foundational in multilingual NLP. All Indian-origin. The more you try to take credit for collective effort the more they'll hate you. All the authors of the transformers architecture used in IILMs and the authors of PyTorch are listed below: The "Attention Is All You Need" paper authors: Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, Illia Polosukhin. PyTorch original developers: Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan.

3

🚫 Stop Writing Monolithic ML Scripts, Do This Instead ✅ Too many ML projects start as “quick scripts” and end up as 1,000+ lines of tangled code that’s nearly impossible to debug, reuse, or scale. In PRATHAMESH JOSHI's latest video at Vizuara's ML in Production Series , He has tried to explain: 1)Why monolithic ML scripts are a nightmare for real-world projects 2)How to debug smarter (and avoid “walls of text”) 3)Designing modular ML code with clean folder structures 4)YAML configuration files for reproducible experiments 5)Test-Driven Development (TDD) in ML Instead of just theory, He rebuild the Iris dataset project in a modular way separating data loading, preprocessing, model training, evaluation, and config management so you can see exactly how it works in practice. 📺 Watch here: https://lnkd.in/gXCZ3tpF 📖 Read the full article: https://lnkd.in/g5sTACCE Let’s move from “it works on my machine” to production-ready, maintainable ML code. 🚀 Attaching Notes below for reference #MachineLearning #MLOps #DataScience #ModularCode #Python #IrisDataset #vizuaraAI #Vizuara #Development #Deployment #DevOps #MonolithicCode

14

From Hours to Minutes: Go-powered ML that moves fast! 💨 Join Vaidehi Thete at #GopherCon as she shares how her team supercharged a machine learning pipeline—cutting inference times from hours to just 10–15 minutes using Go. 🚀 In this talk, you’ll discover how #golang brings ML to life in production by:  🔸 Leveraging shared memory to move millions of scores efficiently  🔸 Using interfaces to prototype quickly without sacrificing safety  🔸 Decoupling operational logic from model ranking for better architecture Whether you're scaling ML systems or just love elegant backend design, this session shows how Go can be the engine behind high-performance inference at scale! 🛠️ 🎟️ www.gophercon.com  #GopherConOrBust #GophersTakeManhattan

3

Learn how TiDAR (“Think in Diffusion, Talk in Autoregression") works — Moulik Gupta's latest explainer walks us through this innovative approach to speeding up LLMs' outputs. https://lnkd.in/eSth96Ey

5

What if agents could coordinate without a bottleneck? Three months of research into routing mathematics, graph memory architectures, and peer-to-peer communication patterns. Sharing what I learned building Kitana and Kontext. https://lnkd.in/gKzxHMwn

9

As an ML Engineer, I get the most questions asked from NLP, Gen AI, and LLMs. So, I have prepared this structured roadmap which covers end-to-end NLP for your interview prep. From my personal experience, I have seen interviewers don’t just care about theory. They want to see: - Intuition - Project experience - Awareness of evaluation metrics, trade-offs, limitations, and production readiness Follow #careerlabconsulting for more daily tips like this...! #careerlabconsulting #cfbr #interview #

3

2 Comments

🚀 RBMs & DBMs : I explored some of the foundational generative models in deep learning Restricted Boltzmann Machines (RBM) and Deep Boltzmann Machines (DBM) through a blend of NPTEL tutorials and seminal research papers by Geoffrey Hinton. 📚 Highlights of My Learning Journey: 🧠 RBM – Understood the energy-based architecture, Gibbs sampling, and Contrastive Divergence for unsupervised feature learning. 🔗 DBM – Learned how stacking RBMs enables hierarchical feature extraction, using greedy layer-wise pretraining followed by joint fine-tuning. 🛠 Hands-on Implementation – Built RBM and DBM models in PyTorch, trained on MNIST dataset, and visualized reconstructions. 📄 Research Insights – Studied how these probabilistic models can capture complex dependencies in data. 📚 Learning Resources 🎓 NPTEL RBM & Graphical Models Lecture 📄 RBM Paper by Geoffrey Hinton (2007) 📄 Deep Boltzmann Machines Paper by Salakhutdinov & Hinton 💡 These models may be classical, but they’re incredibly valuable for understanding the mathematical foundations behind today’s deep generative approaches like VAEs and GANs. Your Turn: How are you planning to explore or learn these foundational models? Drop your learning approach or resources in the comments let’s make it a knowledge-sharing thread! #MachineLearning #DeepLearning #RBM #DBM #GenerativeModels #AIResearch #PyTorch #GeoffreyHinton #NPTEL #UnsupervisedLearning #KnowledgeSharing #connections

17

2 Comments

#AI, #machineleanring #llm #GenAI, #AgenticAI Indore Institute of Science and Technology, The LLM Family Tree: Open-Source vs Closed-Source (Who’s Building What?) Large Language Models are evolving fast — but not all LLMs are built the same. Some are open for innovation, others are closed for control & scale. Let’s break it down 👇 🔓 Open-Source LLMs (Community-Driven Power) Best for: research, customization, cost control, on-prem deployment 🧠 LLaMA / LLaMA-2 / LLaMA-3 — Meta 🧠 Mistral / Mixtral — Mistral AI 🧠 Falcon — Technology Innovation Institute (UAE) 🧠 BLOOM — BigScience (Hugging Face) 🧠 Qwen (open variants) — Alibaba 🧠 Phi (small models) — Microsoft Fine-tuning freedom Transparent architecture Ideal for Agentic AI, RAG, and enterprise-specific models 🔒 Closed-Source LLMs (Enterprise-Grade & Polished) Best for: reliability, scale, production readiness 🤖 GPT-4 / GPT-4o / o-series — OpenAI 🤖 Claude — Anthropic 🤖 Gemini — Google DeepMind 🤖 Cohere Command — Cohere 🤖 Amazon Titan — AWS Strong reasoning & alignment Optimized infrastructure

50

Kharagpur Data Analytics Group, IIT Kharagpur is partnering with NYU Tandon and Pathway to bring you a sharp conversation on what comes after Transformers, straight from the people shaping the next wave of AI. Especially for you if you’re: → 2nd/3rd year exploring research internships → Final year planning MS/PhD applications abroad This is part of a larger push to connect Indian AI talent with top US research labs, and we made sure IIT KGP has a seat at the table. Speakers: → Martín Farach-Colton (Chair CSE, NYU Tandon) → Julian Togelius (Professor NYU, Author Artificial General Intelligence) → Adrian Kosowski (CSO, Pathway | PhD at 20, 100+ papers) → Zuzanna Stamirowska (CEO, Pathway, Moderator) Date: Feb 6, 2026 | 9:30 PM IST Register: https://lnkd.in/dQMwHfCi Limited registrations with exclusive merch for select participants!

91

Realistic, emotional AI voices at sub-100ms latency -- built by two 23-year-old Indian students with zero budget. Maya1 is here. It's outperforming even Google’s best model with: - 3B parameters, runs on a single GPU - 20+ emotions, from laughter to rage - Sub-100ms latency, perfect for real-time applications What makes it stand out: - Natural language voice design: Just describe the voice like briefing a voice actor - Inline emotion tags: <laugh> <sigh> <whisper> <angry>… control emotions precisely where you need them - Streaming-ready: SNAC codec enables real-time voice generation at 24 kHz - Production-ready infrastructure: Single GPU, vLLM integration, WebAudio compatible Use cases are massive: podcasts, audiobooks, AI agents, game characters, accessibility tools, and more—all fully open source under Apache 2.0. Check it out and hear it in action: https://lnkd.in/gsJBEUVM ♻️ Share it with anyone building AI voices and apps :) I share tutorials on how to build + improve AI apps and agents, on my newsletter 𝑨𝑰 𝑨𝒈𝒆𝒏𝒕 𝑬𝒏𝒈𝒊𝒏𝒆𝒆𝒓𝒊𝒏𝒈: https://lnkd.in/gaJTcZBR #AI #GenAI #AIAgents

6

2 Comments