Manoj M

Shweta Shah, Ph.D. has suggested a concept that she calls "Content Quality Management". Shweta and I have been writing together for a while, and I am really impressed by this idea, that I think addresses exactly what the data community needs to consider, going forward in an era so indisputably dominated by AI. Shweta is adding to an ecosystem of thinkers like Bill Inmon (with his absolutely wonderful concept of the "textual warehouse") and Jessica Talisman (coiner of the "ontology pipeline"), and a wealth of people in the records- and information management space like tech entrepreneur Rachael Greaves, and skilled practitioners like Michael Upton, Andy Potter (deeply involved in ISO standards on AI records), and Stephen Clarke MRIM that are trained in thinking in the intersection of language and data. If you want ideas in how to succeed with AI, I suggest you follow Shweta Shah, Ph.D. #contentquality #contentqualitymanagement #ai #data #language #dataquality #dataqualitymanagement

42

6 Comments

In conversation with Dr Manish Gupta, Senior Director at Google DeepMind who has been the core behind the latest advancements in Google AI Infrastructure and advancements. I got a chance to connect with him and talk about: - How is Gemini solving for multilingualism and resource constraint as in India? - Long term research efforts V/s Real-time updates that we see - how are efforts aligned? - Advice for young folks looking forward to leverage Gemini models for building in India. - Building Responsible AI in India considering the landscape and dialects in India. Also, Gemini Pro is now free for students across India and I believe this is the best time to upskill, build and leverage it to the full potential. Thanks to the team at Google for facilitating this conversation at Google I/O Connect. #google #googledeepmind #gemini #coding

697

19 Comments

🚀 Latest AI Tech Headlines (Sept 19, 2025) AI is moving fast—here are 5 headlines from the past week that matter 👇 1️⃣ Anthropic Launches Claude’s First Brand Campaign Anthropic unveiled “Keep Thinking”, its first big brand push for Claude—positioning it as a thoughtful, safe AI partner rather than just another assistant. 🔗 https://lnkd.in/dbMHynQM Learning: Branding and trust are now as critical as raw model performance. 2️⃣ DeepSeek Claims Ultra-Low Training Costs China’s DeepSeek says it trained its R1 reasoning model for just $294,000—a fraction of typical frontier model costs. 🔗 https://lnkd.in/dC2g-muY Learning: Cost efficiency may redefine global AI competition and access. 3️⃣ US-China AI Race & Regulation Dominate DC Summit At the Axios AI+ Summit, leaders debated regulation, global rivalry, and job displacement. The US and China’s approaches are diverging fast. 🔗 https://lnkd.in/dyQyKSuB Learning: Geopolitics is now central to AI strategy—companies must track policy as closely as tech. 4️⃣ OpenAI's GPT-5 and DeepMind’s Gemini 2.5 Wins Gold in Programming Contest GPT-5 and Gemini 2.5 earned a gold medal at ICPC, beating human teams on complex optimization problems. GPT-5 managed to solve all problems correctly! 🔗 https://lnkd.in/dz8VBD9m Learning: AI reasoning is reaching new milestones—practical problem-solving is the next frontier. 5️⃣ Thinking Machines Lab Pushes for Determinism Mira Murati’s Thinking Machines Lab is exploring ways to make AI deterministic—ensuring the same input always yields the same output. 🔗 https://lnkd.in/dWdruJFt Learning: Reliability and predictability are becoming core priorities for enterprise AI. 📌 Quick Takeaways – Trust & branding will define adoption beyond performance. – Cost efficiency is a competitive weapon in the AI race. – Policy & geopolitics are shaping the global AI landscape. – Reasoning ability is the new benchmark. – Reliability & determinism are key for enterprise trust. #AI #ArtificialIntelligence #AIstrategy #AIinnovation #EnterpriseAI #FutureOfWork #TechLeadership #AItrends

6

Does the PR or SRK model predict the phase behavior of methane (C1) better 🤔? This is a question where I know that a lot of people have speculative opinions and hand-wavy answers to, but have you ever tried to use the models to predict actual data 👨‍🔬? This is what I have done in the figure below for two temperatures 👨‍💻. The first temperature (160K) crosses the vapor pressure (left plot), while the second temperature (190K) crosses the critical point of methane (right plot) 🌡️. The models used here are what I will refer to as "vanilla" PR and SRK models, where volume shift have not yet been introduced. Just to emphasize, I would *NEVER* recommend using a vanilla PR or SRK (i.e. without proper volume shifts - matched to data) for any engineering calculation ‼️ In fact, the reason I'm showing the vanilla model predictions is to show that they don't work 🤓. So, which vanilla model is better? Well, based on the prediction of the data, they are both "bad" depending on the temperature 💡. In the next post, we'll see what the effect introducing the volume shift to the EOS model calculations. Keep posted and happy learning 👨‍🏫!

78

5 Comments

Ever wondered what a Data Scientist actually does all day? Spoiler: it’s not just AI, ML, LLM and fancy models — the buzzwords don’t tell the full story. In our latest blog, my teammate Abhirath Anupam Joshi, Senior Data Scientist at Zepto, shares what his day really looks like: 🧩 Solving problems at the intersection of infrastructure and applied ML ⚗️ Experimenting, deploying, and monitoring models at Zepto-scale 🤝 Collaborating across teams to ensure solutions are not only technically robust but also adopted by the business 📈 Finding the balance between building accurate models and scaling them reliably to handle millions of requests His biggest learning? At scale, system design and data engineering are just as critical as modeling. The real thrill comes from watching systems run smoothly under massive traffic — chai helps too ☕. A great read for anyone curious about what it takes to thrive as a Data Scientist in a high-growth, high-scale environment. Check it out here 👉 Link in comments Bhavya Sharma, Shourya Roy, Nikhil Mittal

193

3 Comments

15 years back, when we were building voice bots for the US market - one of the key debates used to be whether to declare or not to declare its a bot. Even if not explicitly say things like "I can help you with X,Y, Z queries and you can say things like .... " indicating a limited understanding. Even if the users did not understand what a virtual assistant is - they understood this is not a human being speaking on the other side. European regulations strictly mandate the voice agents to declare before proceeding in a conversation. India is an upcoming market but getting into the space at different stage of maturity of conversational engines - Hence there is a lot of asks to cover up the underlying bot and not by any means show up its a machine. The experience should be "human like" but that does not mean we should hide the machine behind. For an end consumer, its important to know the species they are speaking to and I now see why some of the western paradigms firmly support the declarative approach. Its just more human.

100

3 Comments

If there’s one piece this year I’ve read multiple times over, it’s AI as Normal Technology by Arvind Narayanan and Sayash Kapoor. “The statement ‘AI is normal technology’ is three things: a description of current AI, a prediction about the foreseeable future of AI, and a prescription about how we should treat it.” I first heard Arvind unpack this at Open Data Science Conference (ODSC). It crystallized the gap I see between the high investment sentiment in AI and the reality of deployment. Here are three themes that stood out: 1. Value and risk live in deployment, not to be confused with development Most arguments about AI’s imminent impact on the economy (negative and positive) focus on the pace of model advancements. This does not account for the friction involved in deploying AI to production. Some of the most compelling opportunities lie in deploying AI in safety critical areas, like healthcare, financial services, education, and security. We've seen this pattern before. We've had deep learning models in production for a decade, yet most production ML use cases in healthcare, for example, rely on decades old statistical methods. 2. Our tools to measure the capabilities of AI models are necessary but not sufficient for generalization Benchmarks, used to measure the progress of AI models are "quickly becoming saturated and losing utility as we try to measure performance on more general tasks" "This is a problem of ‘construct validity,’ which refers to whether a test actually measures what it is intended to measure." Scoring well on the bar exam doesn’t mean a freshly minted lawyer can sit in a deposition or navigate ambiguity in a contract. The same is true of an LLM. "Intelligence is not the property at stake for analyzing AI’s impacts. Rather, what is at stake is power—the ability to modify one’s environment." And with that lies an opportunity to improve how we measure the utility of AI models on highly generalizable tasks. 3. The future of job growth will be in AI control AI is much more like a bionic arm than a piece of factory equipment on autopilot. It can give us superpowers, but still requires direction and control. "when models are more complex and less intelligible, it is hard to anticipate all possible deployment conditions in the testing and validation process." There will be an awkward period where AI can automate a lot of tasks but we won't have prepared fresh grads with the skills to critique and oversee AI systems effectively. More AI deployment will also mean more testing and validation. And more jobs in these areas, likely also using AI tools! The good news? There's a lot of upside to getting AI right. And there's a lot of work that still needs to be done. You may have heard me say boring AI is the best AI. This article lays out why. If you influence AI strategy, this article is a must-read. https://lnkd.in/gy7EGSKQ

37

3 Comments

Glad to join Sriram Ganapathy and Guido Zuccon as a panelist in the session on AI and social impact (moderated by Ashwani Sharma) during the Google.org internal APAC team summit. It's interesting to hear the different perspectives on AI for social impact across India, Singapore, and Australia. I've shared about AI Singapore's efforts through the AI in Health and AI in Education Grand Challenges, among others. We've also discussed whether universities should still be involved in #LLM research. I've shared about whether we can 'do more with less' by achieving state-of-the-art performance in AI using less resources (i.e., data, model parameters, compute, hardware, energy, developers’ time, etc.) and hence lower the costs and impact to the environment? Recent works (including our own) have shown evidence of such a possibility, which signals the beginning of the Resource-Efficient AI era (if not already). This is a research area focus in the National AI Strategy 2.0 that was announced during the Singapore Conference on AI in December 2023. Thanks to Ashwani and Marie Efstathiou for organising this panel!

76

When JAM and UPI went live, India quietly solved two of its hardest problems: identity at scale and money you can move in seconds. The next frontier is tougher. Can land records become part of India’s Digital Public Infrastructure the way Aadhaar and UPI did? In Business Standard, Arvind Gupta argues that tokenising land titles could do for property rights what UPI did for payments. Think of each parcel of land as a programmable, auditable token mapped to a verified registry entry— not a dusty file sitting in a tehsil office. Why this matters for DPI Aadhaar made people visible. UPI made money liquid. Tokenised land records can make ownership trustworthy and portable, with every transfer writing to a shared, tamper-evident ledger. This connects directly to the ease of justice problem. Nearly two-thirds of India’s civil cases involve land and property disputes, locking up wealth and clogging courts. Clean, digital titles on a common stack would: • Reduce litigation • Unlock credit • Improve state revenues from stamp duty and registration Some states are already moving. Telangana is experimenting with blockchain-linked land registries and tokenisation pilots, showing how reform can start bottom-up and plug into a national DPI layer over time. If JAM was India’s Web 2.0 unlock, land tokenisation is a serious Web 3.0 candidate, where law, code, and registries sit on the same rails. Strongly recommend reading the full piece by Dr. Arvind Gupta at Digital India Foundation in Business Standard: “Tokenising trust: How land reform can meet India’s digital ambition.” Full blog link attached in the comment.

363

9 Comments

A big bet on India. OpenAI has launched ChatGPT Go, its first-ever country-specific subscription plan, exclusively for India. ₹399 per month. Payable through UPI. This is OpenAI officially recognizing India as their second-largest market and adapting their entire strategy around us. Here's what ChatGPT Go offers: // 10x more message capacity than the free version // Daily image generations and file uploads // 2x memory length for personalized responses // Powered by GPT-5 with better Indic language support // Priced 5x cheaper than ChatGPT Plus (₹1,999/month) // For the first time, you can pay for ChatGPT with UPI. No credit cards needed. OpenAI literally rewrote its payment infrastructure for India. When a $157 billion company creates India-only products and accepts UPI payments, they're not just entering a market - they're betting their future on it. The message is clear: India is no longer just consuming AI. We're driving it. - Google has Gemini teams in Bangalore. - Microsoft runs Azure from Hyderabad. - AI at Meta's AI research happens in Gurgaon. And now OpenAI is customizing products specifically for Indian users. The opportunity window is massive, but it won't stay open forever. Every month you delay upskilling in AI is a month your competition gets ahead. The companies betting on India are the same ones hiring AI-skilled professionals at 3-5x salaries. Are you upskilling in AI, or still waiting for the "right time"? My comments are open. 👇 PS: Join Jan AI for daily prompts, upskilling opportunities, income streams + more (Link in comments)

33

3 Comments

At DocLens.ai, we’re building in a space where AI is still taking shape. In the context of complex P&C claims there is no playbook, only real problems waiting to be solved by those willing to understand and take initiative. This recent blog by Mrutunjay Kinagi, one of our senior engineers, is a great example of what that looks like in action. He didn’t just write about ClaimLens™, he built a working demo. He took the time to understand the workflow and the persona's Way of Work and Way of Thinking. Integrated it with the Guidewire environment template. Tested real-world scenarios. And then documented the entire journey. Proud of Mrutunjay and equally proud of the culture we’re building at DocLens.ai. A culture where individual curiosity, hands-on problem-solving, and bottom-up innovation aren’t just encouraged, but expected. Read the full piece here: https://lnkd.in/eZcSGhkp

61