Vamsi Adari

People often ask why India still doesn’t have many global-scale LLMs of its own. It’s not talent. It’s not ideas. It’s mostly infrastructure + economics. Training frontier LLMs needs: • Massive GPU clusters (tens of thousands of GPUs) • Extremely expensive compute (hundreds of millions of dollars) • Huge proprietary datasets • Long-term patient funding without quick ROI In the US/China, Big Tech can burn billions before seeing profit. In India, most companies still optimize for services revenue and short-term ROI. So instead of building foundation models, we became very good at applying AI. We build copilots. We build AI products. We build AI services. But foundation model research needs deep tech funding + risk appetite + infra. That’s slowly changing now. But frontier models are a capital game first, talent game second. #learning #goals #Motivation #technology

174

20 Comments

Bait and Switch? How ZEE5 is scamming FIFA fans in broad daylight. Yesterday, I paid ₹799 for the ZEE5 FIFA World Cup subscription pack. The terms were crystal clear and explicitly advertised: Access on 3 devices. Today? The app has suddenly slashed that access down to just 1 device. This isn't just a "policy update"—it is a classic bait-and-switch scam. You cannot sell a consumer a product under one set of rules and then secretly change those rules 24 hours later after pocketing their money. That is unethical, illegal, and a blatant violation of consumer trust. If a local shopkeeper sold me a box of 3 items and then snatched 2 back out of my hand after I paid, it would be called theft. Why does a multi-million dollar corporate entity like Zee Entertainment think they can get away with the exact same behavior online? I bought this plan expecting to share the tournament with my family. Now, we are locked out because of corporate greed. ZEE5 Global, fix this "glitch" or whatever excuse your PR team is cooking up, and restore the 3-device access your customers actually paid for. This is absolute robbery. #ZEE5 #FIFAWorldCup2026 #ConsumerProtection #Scam #UnfairTradePractices #ZeeEntertainment

8

Google dropped Gemini 3 and AntiGravity. Everyone’s busy comparing benchmarks, talking about speed, parameter counts, performance curves. but here’s what’s actually interesting 👇 we’re watching Google slowly turn AI from a “model you query” into an infrastructure layer. Gemini 3 for reasoning + generation. AntiGravity for training, distillation, and squeezing huge capabilities into tiny devices. And the pattern’s becoming clearer: AI isn’t just getting faster, it’s getting denser...same intelligence, smaller footprint, lower cost. That changes who gets to build. But here’s the part that caught me: we’re moving into a world where model size won’t be the moat anymore. Instead, the moats will be: -how cheaply you can train -how efficiently you can compress -how safely you can deploy, and >how much “agency” you’re comfortable giving these systems Gemini 3 is a strong model. AntiGravity is a strong signal. Google wants to make “training big models on small hardware” normal. that sounds empowering… until you realize it means a lot more people will be deploying agents they don’t fully understand. and if the last year taught us anything, it’s this: >capability scales faster than safety. >compression scales faster than control,and >agency scales faster than alignment. AI won’t be dangerous because it’s powerful. AI becomes dangerous when it’s powerful, cheap, and everywhere. Gemini 3 shows where the models are headed and for me AntiGravity shows where the world is headed. the question is the same as always: are we building systems we can steer? or systems we can only hope behave? still our call. if you read till here, i'm sure you'll read this too: https://lnkd.in/guQdBDzV

8

Heading to PyTorch India (Bangalore) — excited about vLLM × NVIDIA inference Large-scale LLM inference has quietly become a systems problem, not a model problem. As PyTorch India approaches, I’m particularly excited about discussions around: • vLLM internals (PagedAttention, continuous batching) • GPU-aware inference scheduling • Memory bandwidth + KV-cache pressure • Where PyTorch abstractions end — and hardware realities begin With sponsors like NVIDIA, IBM, and Red Hat, this event sits right at the intersection of: PyTorch → vLLM → GPU kernels → production reliability. Questions I’m carrying into the event: 1.What breaks first in real production — scheduler, KV cache, or memory bandwidth? 2.Where do PyTorch users *overestimate* batching gains? 3. How far can GPU-native optimizations take us before architectural changes are unavoidable? Looking forward to learning, debating, and connecting with folks building inference at scale.

18

𝗔𝗜 𝗟𝗲𝗮𝗱𝗲𝗿𝘀𝗵𝗶𝗽 𝗥𝗲𝗾𝘂𝗶𝗿𝗲𝘀 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲, 𝗡𝗼𝘁 𝗝𝘂𝘀𝘁 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻 If India wants to compete with global AI powers like the USA and China, we must confront the infrastructure reality. Today, India operates at roughly 1 watt of compute per person, while China operates at ~20 watts per person, and the USA at ~150 watts per person. This is not just a statistic. It represents capacity, speed, research advantage, and geopolitical influence. AI is powered by compute. Compute is powered by energy, chips, fabs, and data centers. The scale required to compete is enormous: • 50 million CPUs • 7 million GPUs • 3 billion memory devices • 1 million storage systems • 3nm semiconductor fabs operating at million scale monthly • High Bandwidth Memory scaling from current capacity toward industrial-scale production — with at least 10 memory fabs required This is a national infrastructure challenge, not just a startup problem. The discussion also centered around the five strategic pillars shaping the AI economy: 🔹 Compute 🔹 Content 🔹 Communication 🔹 Cognition 🔹 Community India’s GCCs (Global Capability Centers) were the previous growth engine. The next wave could be sovereign compute, semiconductor capability, and AI-native infrastructure. The question is no longer whether India will adopt AI. The question is whether India will build the backbone that powers it ? #IndiaAI #AIInfrastructure #IndiaAiImpact #ComputeEconomy #Semiconductors #DigitalSovereignty #AIFuture #artificialintelligence

9

Wrote a blog explaining "LoRA Fine-tuning Efficiency Under Different Loss Functions." LoRA made fine-tuning cheap, but not always efficient. This post explores how loss functions like Cross-Entropy, Label Smoothing, and Focal Loss change convergence, stability, and generalization in lightweight LLMs. Read here: https://lnkd.in/gh_vm2kG

2

𝗜𝗻𝗱𝗶𝗮 𝗜𝗰𝗲𝗯𝗲𝗿𝗴 𝗜𝗻𝗱𝗲𝘅: 𝗛𝗼𝘄 𝗜 𝗺𝗮𝗽𝗽𝗲𝗱 𝗔𝗜 𝗲𝘅𝗽𝗼𝘀𝘂𝗿𝗲 𝗮𝗰𝗿𝗼𝘀𝘀 𝟭.𝟰 𝗯𝗶𝗹𝗹𝗶𝗼𝗻 𝗽𝗲𝗼𝗽𝗹𝗲. A few weeks ago, I shared a map showing where AI might actually impact the Indian workforce. That visual went viral, but it was just the output. Today, I am releasing the engine behind it. We often see AI exposure studies for the US and Europe because the data infrastructure - specifically the O*NET database - is designed for Western economies. But the Global South works differently. The job titles, classifications, and labor structures do not map 1:1. For the past few months, I have been building 𝗜𝟯 or 𝗜𝗻𝗱𝗶𝗮 𝗜𝗰𝗲𝗯𝗲𝗿𝗴 𝗜𝗻𝗱𝗲𝘅 to bridge this data gap. The full methodology is now live on SSRN. The challenge was not the analysis; it was the translation. As the system diagram below shows, I had to architect a crosswalk between two massive, disconnected datasets: 𝟭. 𝗧𝗵𝗲 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗕𝗿𝗶𝗱𝗴𝗲: I could not just match keywords. I built a Hierarchical Semantic Matching Algorithm to map 3,445 Indian National Classification of Occupations (NCO) codes to 1,016 US O*NET codes. This allowed me to "import" the AI exposure scoring frameworks (like Felten et al.) used in US research and rigorously apply them to Indian microdata. 𝟮. 𝗧𝗵𝗲 𝗗𝗮𝘁𝗮 𝗘𝗻𝗴𝗶𝗻𝗲 (𝗣𝗟𝗙𝗦 𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴): I ingested the Periodic Labour Force Survey (PLFS) 2023-24 microdata—processing both person and household files—to establish the ground truth of the workforce structure. 𝟯. 𝗧𝗵𝗲 𝗜𝗰𝗲𝗯𝗲𝗿𝗴 𝗜𝗻𝗱𝗲𝘅: The final output isn't just a raw score. I calculated a 𝗦𝘂𝗿𝗽𝗿𝗶𝘀𝗲 𝗜𝗻𝗱𝗲𝘅 to identify districts where high AI exposure correlates with low urbanization - identifying the "hidden" risk that traditional economic indicators miss. This is the plumbing behind the map. It is my attempt to turn a qualitative debate about "AI taking jobs" into a quantitative framework for policy. You can read the full paper or explore the data yourself on the app I built below. 𝗣𝗮𝗽𝗲𝗿 (𝗦𝗦𝗥𝗡): https://lnkd.in/gFKmxKbk 𝗟𝗶𝘃𝗲 𝗔𝗽𝗽: https://lnkd.in/gfCBBVUq #AIResearch #Economics #India #LaborMarkets #DataScience #ProjectIceberg #SSRN

13

1 Comment

Was going through the usage of Autoencoders for Code migration, Anomaly detection etc. As we know that basically the auto encoders, encode-compress and then decode-regenerate the original data. Hence for images, the dimensionality reduction and noise reduction are few of the merits and advantages of using the encoders. Behind the scenes they utilize AST trees, where the nodes are identified, embeddings are generated by the embedding matrices which contains the nodes with similar or identical functionality placed nearby in the vectorized matrices - This provides a better structural mapping in case of Code migration tasks. Further parsing of these trees using the LSTM or GNN lead to the formation of single vectors. Hence with reduced dimensions, the model can be trained using this single vector effectively and it is easy to regenerate or de-parse this vector and recreate the code or detect any deviations. So very interesting indeed.

8

if Ai is commodity then commodity is all about three dimentions. - Cost per unit - Speed of throughput - Quality of outcome -- All thin wrappers of all commodities are fade. Only ones that survive long run are the ones who optmize for atleast one the above 3. -- Creators mistake complex wrappers as non thin wrappers. However wrappers dont need to be complex. Wrapper needs to create its on asset with every use. An. asset created by one user is useful to all user. Hence non thin wrappers are the ones with network effect. Network effect is not new but what is new is Ai's ability of accelerate the growth of network effect.

33