Sudhir is a mechanical engineer from VJTI. He built Medimanage and ran it for about twenty years, then sat on the boards of small manufacturers in Nashik, and now sits on the governing council of the ASCENT Foundation. With his partner Rajendra Bagwe, who has run the Teaching Learning Community for 21 years, he turned that mentoring into a hundred business frameworks at HowFrameworks and put an AI in front of them.

For Sherpa, their AI-assisted advisory service, a three-hour discovery call goes through ElevenLabs and his own AI skills and comes out as a year's operating plan. He charges ₹3–4 lakh for it; he says BCG or McKinsey would want ₹50–60 lakh, and would never go to Kolhapur or Satara. Owners who spend their days on the road get 20-minute podcasts in Marathi that he makes with NotebookLM. And in 75 hours of interviews, not one of them brought up data privacy.

Highlights

• The LLM asks the questions, the framework gives the answer — ChatGPT gives him three different answers to the same question in one day. His product uses the model for conversation and context and takes every answer from a framework.

• Judgment is what AI can’t do yet — Two Kolhapur foundries at the same revenue can need different fixes: a dashboard for one and daily leadership reviews for the other. He expects AI to get there eventually and says it hasn't.

• Five stages of a business — founder-run, team-capable, systems-run, strategy-led, self-sustaining. The stage caps the growth target he will sign off on. An owner at ₹15 crore who wants ₹25 crore next year gets a target of ₹18 crore.

• Questionnaires lie, calls are necessary — SME owners answering questionnaires leave a lot to be desired. "Yes, we have SOPs" usually means habits nobody has written down, and "yes, we have a website" can mean a listing on a B2B portal. So, 3–4 hour calls are needed, and then an AI transcription followed by making sense of everything.

• The founder's story goes in the context — an ITI-trained owner running a ₹40 crore company still embarrassed about his English needs different advice from a second-generation US returnee, so the narrative history is recorded along with the numbers.

• The office, mostly — his one shop-floor example is a Nashik plant that points cameras at old machine counters because the machines have no API.

• A 30-tonne casting for Sweden — a Kolhapur foundry made the hub of a windmill, and still pitches abroad with weak presentations. Second-generation owners are fixing that with Apollo, LinkedIn automation and AI-drafted email.

• Marathi podcasts from an English document — a break-even explainer went into NotebookLM and came out in Marathi, Hindi and Konkani. He checks every number and listens to every episode first. In a survey last month, owners named these podcasts as where their insights came from.

• 25 clients instead of five — each monthly review is recorded, and minutes with action items go out within the hour, in Marathi where needed. Weekly reminders follow on email and WhatsApp.

• Father against son — first-generation owners have FOMO and barely use AI. Their children use it for the wrong jobs, like building the company website themselves, and read the dashboards as a case for firing a worker the father has trusted for fifteen years.

• Owner, graduate, AI assistant — he would rather an SME hire a young graduate with an AI subscription than make the owner learn it. The owner brings the context and the graduate does the finishing.

• Pay for it, and skip the courses — he has no time for free tools or for ₹1–5 lakh courses. He moved to Claude on 1 January after one conversation, learnt from Claude's own courses, and did last year's company accounts in it from his card and bank statements.

Notable quotes

I ask ChatGPT same question three times a day and get three different answers every day.

Basically, I would say one sentence: judgment. That judgment is something that cannot come to AI soon.

There is a lot of talk about AI taking away jobs and all. I have been working with these companies for last 18 to 24 months very closely. Nobody is bothered about it. Rather they are saying, I have so many jobs to do, can I get more people?

Anybody who can articulate well can delegate well.

Today I do work with 25 people. 5x is only because of that, that I have a layer where I can get things done very fast.

Not a single time anybody talked about data privacy. Yes, they are worried about employee taking my customer details.

You have to write and don't use AI as a crutch, use it as a research tool. (Amit Varma's advice to him, as he tells it.)

Links & resources

• HowFrameworks: Sudhir and Rajendra Bagwe's hundred frameworks, with the AI mentor in front of them.

• ASCENT Foundation: the peer-learning network for entrepreneurs where Sudhir sits on the governing council.

• The Seen and the Unseen with Sudhir Sarnobat: Amit Varma's long interview with him, for his life story. Amit is also the friend who told him to write his own LinkedIn posts.

• The tools he names: Claude, ElevenLabs for transcription, NotebookLM, since renamed Gemini Notebook, for the Marathi podcasts, and Apollo for export leads.

• KnOT (Knowledge on Tap): the Pune talk series Sudhir says he wants to come to. The link is its WhatsApp community.

Connect with Sudhir Sarnobat

• LinkedIn, where he says he is most active and posts two or three times a week about MSMEs

• Website howframeworks.com; he says enquiries sent through the site get a reply within 24 hours

Our guest, Binoy Oommen, works in electronic design automation at Siemens EDA. He began his semiconductor career at Mentor Graphics and has spent much of it in functional verification: checking whether a chip design behaves as intended.

We begin with cars, transistors and the extraordinary compounding behind modern computers. Then we ask why NVIDIA’s advantage extends beyond its chips, where India fits into the semiconductor industry, and whether AI could help Indian engineers build verification products of their own. By the end, we’re talking about a crowded engineering lab, the cost of discouraging children from asking why, and a choice between money for education and ten times as much for industrial policy.

Highlights

• A chip, explained through a car: Binoy separates design from manufacturing before taking us into the much smaller world of transistors.

• What decades of doubling actually mean: Moore’s law, compounding, and the gap between knowing the formula and grasping its consequences.

• Why faster computers can still feel slow: Navin explains how software absorbs the hardware improvements we have come to expect.

• The limits to shrinking: Heat, quantum tunnelling and the rising cost of fabrication complicate the next generation of chips.

• NVIDIA’s software advantage: Binoy argues that CUDA and the ecosystem around it help explain why a competing chip alone is not enough.

• Why AI needs so much hardware: The connection between parallel computation, GPUs and the workload behind AI systems.

• Talent is only part of the story: Binoy distinguishes having engineers in India from owning a product and making its strategic decisions.

• An opportunity in verification: His proposal to build on India’s existing expertise through AI-assisted test generation, bug localisation and coverage optimisation.

• The problems between specialities: Why Binoy sees opportunities where chip design meets manufacturing, and hardware meets software.

• The habit of asking why: All three speakers connect classroom habits with the questions people do—or do not—ask at work.

• AI in an engineering syllabus: Binoy asks what an AI course should help an electronics student actually do, and argues for closer work between industry and universities.

• Choices, Choices: Education or industrial policy; teachers or bureaucrats; tea or beer; and a career answer that only an old friend predicts.

Links & resources

• Electronic design automation, explained by Siemens

• NVIDIA’s CUDA FAQ

• TSMC’s 28nm technology, for context on the mature nodes discussed in the conversation

• ASML on the optics behind lithography

Connect with Binoy

Binoy Oommen on LinkedIn

Binoy welcomes conversations with students interested in semiconductors and is happy to help with mentoring.

He is a professor of biology at IISER Pune and chaired the committee that wrote the institute's AI policy, a document that inverts the default nearly everyone else chose. AI is permitted unless a faculty member forbids it in writing, and anyone who forbids it has to say up front how they intend to catch it.

He set 100 undergraduates a research grant proposal he could not have assigned before, then built an agentic system on Claude Opus to review the results, hit his session limits, bought more subscriptions and put four TAs on it. One student's proposal turned on eggs being blue-green; the AI caught it, and the eggs are brown. On the research side he expects AI peer review within a few years, and expects it to make journals think alike, which closes the door that path-breaking papers have always come through.

Highlights

• Permitted by default — IISER Pune allows AI unless a faculty member forbids it in writing. The policy came out of an all-institute survey, six weeks of comments and a senate vote.

• Forbid it and you must say how you will catch it — the clause that keeps the policy honest. No surprise accusations in week twelve.

• Detection does not work — false positive and false negative rates on plagiarism tools are high enough that students will run their work through the detector until it stops complaining.

• What he does instead — an unaided quiz after the assignment, with the mark penalty announced in advance.

• An assignment that needed AI to exist — 100 students writing full grant proposals, at a level he would not have set before.

• And needed AI to grade — 15 to 20 minutes of Claude Opus per report, session limits, extra subscriptions, four TAs. He says the reviews were good enough that he would have been glad to get them as a journal editor.

• The egg — one proposal built its experimental design on the wrong shell colour. He checked Wikipedia; the AI was right.

• "Am I going to do it again? Most probably not" — the effort was too high.

• Grade appeals, automated — students photograph their marked scripts, feed them to AI and come back with a case. "We are not fighting the students anymore, we are fighting Claude."

• AI peer review flattens science — a rejected paper wanders until it meets the one odd reviewer who says yes, and that is how a lot of good work got published. Run the same model at every journal and the wandering stops.

• What is left for PhD students — he does not think good labs used them as cheap hands anyway, so he expects less disruption in research than in industry.

• Motivation, not access, is the divide — the same tool speeds up the self-driven and lets everyone else coast. Second-years buy Claude Max to do number theory in Lean while classmates ask why they should attend a lecture.

• Khanmigo worked and still failed — measurable exam gains for the students who used it, and Khan Academy called it a failure because 90% did not.

• Cognitive friction — the movie-theatre argument, the maahaul a good teacher makes, and why he thinks learning from a screen alone does not stick.

• Cost worries him at the lab, not the desk — free tiers are generous and students rotate between them. Funded groups pulling ahead of unfunded ones is the inequality he expects.

Notable quotes

Use of AI is permitted by default until and unless somebody, somebody as in either a faculty or a PI or an administrator explicitly and in writing forbids the use of AI.

You use AI in whatever way you seem fit, but end of the day, the responsibility for the content is entirely the authors. […] You cannot simply say that I use AI, ChatGPT told me so. Doesn't work.

This is a AI native use case, so to speak. […] Was it beneficial to the students? I would definitely like to think so. Am I going to do it again? Most probably not.

So we are not fighting the students anymore, we are fighting Claude.

You fight with 300 students for one-one mark, you decide only one way.

If you learn from AI, it is not possible for you to fall in love with what you are learning.

That cognitive friction, if AI totally eases it out, then where is the joy of finding things out?

— Ashish Kulkarni: Slight paraphrase, but I hope we end up titling this episode, No Student is an Island.

Links & resources

• IISER Pune — where the policy was written, by a committee Sutirth chaired.

• Daniel F. Chambliss, The Mundanity of Excellence — the study of swimmers at every level that Sutirth calls brilliant, and How College Works, written with Christopher G. Takacs, where his argument about college being mostly social comes from.

• Isaac Asimov, The Fun They Had — the 1951 story he retells, about a child who finds out schools used to exist. He also paraphrases Asimov's "Such folly smacks of genius. A lesser mind would be incapable of it."

• Hollis Robbins — writes Anecdotal Value on AI and higher education; Ashish's peg for the learning-rate question.

• Khanmigo — Khan Academy's tutor, and Navin's example of a tool that works and goes unused.

• Refine.ink — the AI paper-review service Sutirth prices at about $40 a review, and the peg for the West-versus-India cost question.

• Claude, ChatGPT, Gemini, NotebookLM, and Lean — the tools named on air. NotebookLM, since renamed Gemini Notebook, for students working in a second language; Lean for the number theory.

Connect with Sutirth Dey

• Email s.dey@iiserpune.ac.in, which he says is the best way to reach him

• IISER Pune faculty page

• Population Biology Laboratory — his lab

Bhasker “Bosky” Kode builds internal products at OneCard by day. On evenings and weekends he runs AI Grants India, a non-profit that hands free model and infrastructure access to students, indie developers and idea-stage founders, which he co-founded with Vaibhav Domkundwar of Better Capital. He also runs Epoch, a four-weekend programme that takes engineers to a working proof of concept.

At a hackathon in February 2025 he stuck a note on his laptop reading free API keys, and 24 of the 25 people in the room came to his table. He funded the first six months on his own credit card; now whatever he puts in is gone in a week. He has no application form and no interview — a GitHub account, an email and a phone number, and the key is yours. And when founders come back asking for the newest, most expensive model, he has a standard answer ready.

Highlights

• Twenty-four of the twenty-five people in the room — the response when he first advertised free API keys, at Vaibhav Domkundwar’s February 2025 hackathon. Vaibhav became the founding sponsor; Bosky carried the first six months on his own credit card, and now whatever he puts in is gone inside a week. The first large donation landed two days before this recording, from the edtech company NxtWave.

• “Anthropic did not have Opus to build Opus” — His standing reply to founders who insist they need the latest and greatest model. Take a step back, work from foundations, and build with what you have.

• Deliberately no gatekeeping — GitHub, email and phone, once each, purely to stop abuse. No interview, no committee, no pitch: “In fact, there is no gatekeeping. I don’t check anything.”

• He never gives the full amount — A maximum of 50%, on principle: “now find other ways to do it. You should not get everything with a spoon.” He has funded train tickets to conferences on the same terms.

• The first question he asks anyone who hits a rate limit — Are you charging your customers? If yes, the problem solves itself.

• Building stopped being the hard part — Roughly 700 startups have signed up with their own domain and project. About 10% are doing genuinely cutting-edge work — brain-computer interfaces, noise cancellation, energy — and about 70% are SaaS-shaped. His advice is that the journey now starts where it used to end: deciding what to build, and finding the first ten customers.

• “Nine months in a cave” became nine minutes — The old failure was disappearing for nine months and emerging to no customers. Now you emerge in nine minutes to no customers. The reality check arrives sooner, which he counts as a saving.

• What makes something a product rather than a weekend build — Niche plus data. His example: a startup that mapped every road, bend and tunnel in the country to route wind-turbine blades, because a wrong turn costs crores. Two hours of code, but nobody else has the data.

• Two jewellery designers in Jaipur who had never written code — Now building a 3D Photoshop for jewellery design, prompt-driven. Neither has a tech background, let alone a STEM one.

• India is “do it for me”, not “do it yourself” — Even at an automated DigiYatra gate, someone is still standing there helping. His vision follows from it: a shop counter you walk into, say what you want, and someone prompts it into being while you wait — “almost like McD”. Navin points out MCCIA has already built the help-desk version of exactly this.

• The hacker-house economics — AI Grants funded 30 people living, cooking and building together in the Himalayan foothills for three months, because it was cheaper there. Inside Bangalore the same thing runs 6–10 lakhs of rent. In Pune he ran it out of a friend’s spare space on weekends instead — and points out that anyone can do this anywhere, by asking a local office for their empty room.

• Epoch is that hacker house, compressed into weekends — The Himalayan version needs real capital and three months of everyone’s life. What scales instead is four weekends in a friend’s spare office, because “weekends are when it’s free”: engineers arrive and leave with a working proof of concept. Hundreds applied from across the country; he kept it to Pune and Mumbai for practical reasons. One of its founders, Sriram Kintada, is still a student at IISER Pune.

• The age distribution is wider than you would guess — A nine-year-old building robotic droids, whose year of tooling Bosky funds. Seventeen-year-olds in Bangalore customising LLMs and raising money.

• Set aside a budget for your own AI use — Bosky’s framing is that AI turbocharges whatever you already are: designer, lawyer, estate agent. Navin offers a number for it — 2–3% of income, after the rent is covered.

• What he would rather have — Asked to choose between one company that becomes fundable and ten founders whose ambition rises with no outcome at all, he takes the ten. Navin ties it to Tyler Cowen’s line that raising the ambition of young people is among the most valuable things you can do.

Notable quotes

“Anthropic did not have Opus to build Opus. So why do you need Opus?”

“In fact, there is no gatekeeping. I don’t check anything.”

“You can go nine months into a cave and then come out saying that, yes, I’ve done it, I’ve made it, and it’s empty. But now you can come out in nine minutes and it will still be empty saying where is the customers.”

“One thing I do is that I always never give the full amount. So I gave maximum 50% because I’ll say, okay, now find other ways to do it. You should not get everything with a spoon.”

“India is a country of not do it yourself, it is do it for me.”

“My WhatsApp is full of thank yous all day long.”

Links & resources

• AI Grants India — free model, API and infrastructure access for Indian builders. Partnering is a matter of aigrants.in?ref=<your event>; Bosky says the default answer to anyone asking is yes.

• Better Capital — Vaibhav Domkundwar’s fund. He ran the February 2025 hackathon where AI Grants India began, and became its founding sponsor.

• NxtWave — the edtech company behind CCBP 4.0, and the source of AI Grants India’s first large donation, two days before this conversation was recorded.

• Zo House — the Bangalore hacker-house network AI Grants partners with, alongside a residency in Bangalore.

• deAsra Foundation — Anand Deshpande’s Pune non-profit for nano-entrepreneurs, raised by Navin when asking whether AI reaches gig workers and the self-employed.

• NVIDIA Inception — the partner behind free inference across a large model catalogue through a single API.

• MCCIA — the Pune chamber running the SME help desk Navin describes, and the subject of our first episode.

• Sanjay Phadke, FinTech Future: The Digital DNA of Finance — an Epoch participant and co-investor; Bosky cites him as someone who has written the book he hasn’t got around to.

• Tyler Cowen — the source of Navin’s closing line about raising the ambition of young people.

• AI Tinkerers Pune — Bosky’s meetup group; an upcoming edition is aimed at designers.

Connect with Bhasker “Bosky” Kode

• AI Grants India — aigrants.in

• X — @boskykode — the handle linked from aigrants.in. (He gives a different one on air; this is the one that works.)

• LinkedIn — bhaskerkode

• GitHub — bosky101

• Blog — bosky says

Pranay Kotasthane is deputy director of the Takshashila Institution and chairs its High-Tech Geopolitics Programme. Before he was a policy person he spent seven years designing chips at Texas Instruments, which is why he can tell you what an EDA licence costs and why that matters. He co-wrote When the Chips Are Down with Abhiram Manchi, the first book to look at semiconductor geopolitics from an Indian vantage point; he co-writes the newsletter Anticipating the Unintended with RSJ; and he co-hosts the Hindi-Urdu podcast Puliyabaazi. He also ships software. That last part is why he belongs on this show: most people who write about AI policy in India do not open a terminal, and most people who open a terminal do not write about policy.

Some of what you get in the hour: the argument that the absence of state capacity is a reason to use AI rather than a reason to despair of it, worked through the specific case of court scheduling. A correction to the founding myth of Indian software — the industry did not grow because the government stayed away, it grew because software services were filed under the Shops and Establishments Act instead of the industrial and labour acts. And a warning about buying GPUs at national scale that ends on the phrase “what will we do of that bricked chip?” Along the way, a tour of what a 40-person think tank actually builds when it decides its only durable advantage is disproportionate use of technology — including a bot that trawls the internal channels and produces an Eisenhower matrix every Monday.

Highlights

• The frame: success in AI is a societal problem, not a technological one — Pranay opens on a line from Michael Mazarr’s RAND work on the societal foundations of national competitiveness, which looks back at earlier technological revolutions and finds that the role of diffusion is consistently underplayed. He then runs the Indian question through Samaj, Sarkar, Bazaar rather than through chips and compute. “We don’t need an AI strategy, necessarily.”

• State capacity is the argument for AI, not against it — Ashish asks the obvious sceptic’s question: whatever you design for the centre will break at the local level. Pranay inverts it. Not having people is precisely why you reach for the tool. His example is court scheduling — judges spend enormous time maintaining rosters rather than deciding cases, which XKDR Forum has documented, and roster-building is an operations research problem. The alternative is recruiting a new cadre of court managers, which stalls exactly the way police vacancies stall.

• Produce, finance, regulate — governments do three things, so ask the question three times. Produce AI: no. Finance it: yes, and more usefully as a procurer than a funder — nobody builds AI-integrated autonomous weapon systems in India unless the government places volume orders up front. Regulate it: India, unusually, is getting this right. No AI-specific regulator, no ministry of AI, no per-model licensing, in deliberate contrast to the EU.

• The Shops and Establishments Act, and the myth it corrects — the best two minutes in the episode. Navin asks whether Indian software grew because the government stayed out, and whether it should stay out again. Pranay’s answer is that it never stayed out. It made a choice: software services were classified under shops and establishments rather than the industrial acts, with their hiring, firing and PF regimes. Regulation existed; it was just the lighter kind. The live question for AI is which side of that line it lands on.

• A Purvapaksha of his own co-author — Ashish asks Pranay to steel-man RSJ’s bearish case on Indian IT services: with AI and a solid engineering team, enterprises bring the work in-house, and demand for Indian services falls over three or four years. Pranay’s counter is that legacy companies cannot move to AI workflows unaided, and that the leaders of the services and SaaS eras need not be the leaders of the AI era. “Why should the TCS of the AI era be the TCS of the software services era? It’s good if there is churn.” What he does concede: the intake numbers from engineering colleges will not hold.

• Richard Baldwin’s asymmetry — manufacturing supply chains are shortening under geopolitical pressure, but services supply chains can lengthen because of AI. Navin pins down what “lengthen” means — literally distance in metres — and the worked example is surgery: today the doctor flies to the patient and requalifies in that jurisdiction; with good enough robotic and AI-assisted surgery, the doctor stays in India. Navin brings the week’s news to the same point: an Anthropic and Goldman Sachs enterprise AI tie-up read by some as the end of Infosys, and by others as the beginning of Infosys++, because enterprises need more hand-holding than a model vendor wants to provide.

• “There is nothing like sovereign AI” — the flattest rejection in the episode. You cannot be sovereign in semiconductors, so the idea that you can be sovereign in AI does not survive contact. What is sovereign, he argues, is not AI but particular applications: defence targeting, or a Ministry of External Affairs model reading ten-year risk, where you want something fine-tuned on your own data. Build those on open weights. Sarvam exists; the harder question is who adopts it and why they would.

• What ₹10,000 crore should have bought — a large share of the national AI mission went on NVIDIA and AMD chips. Pranay thinks the intent is right and the prioritisation wrong. He would run a challenge grant aimed at getting out of CUDA — Apache TVM, MLIR, LLVM, ONNX Runtime, the work of making the GPU a switchable layer underneath PyTorch. The reason is not techno-nationalism, it is the export-control ratchet: US diffusion rules, on-chip thresholds, and then “what will we do of that bricked chip?”

• And why the alternative needn’t be Indian — RISC-V started at UC Berkeley, and its foundation moved to Switzerland precisely to sit away from both Washington and Beijing; MeitY is now a sponsor. Pranay’s test is not who builds it. “If it is available to every Indian, then our goal is satisfied.” He adds the corrective he keeps having to make in semiconductor debates: we cannot become Atmanirbhar, and neither can Taiwan.

• Open versus closed, not China versus the United States — and the worry that actually keeps him up — on whether Indian firms should use Chinese open-weight models, Pranay’s answer is that open weights are open to everyone, and it merely happens that many of the leading open models are Chinese. Navin adds the technical half: a backdoor in a model is not a backdoor in a Huawei chip or in shipped software — you cannot steer it the same way, and fine-tuning can wash it out. Pranay draws his line at hardware in critical systems: not in telecom, probably fine in an electric vehicle with audit mechanisms. But the risk he takes seriously is not the bazaar’s, it is society’s — a stack fine-tuned to omit things produces users for whom Tiananmen Square did not happen, and the quotidian example matters less than the structural one: a slow drift toward believing that authoritarianism works, because that is what the information said.

• Substitution versus stimulation, and the skill that now matters — the distinction he drew in the newsletter and sharpens here: substitution is outsourcing your thinking, which for a think tank is fatal; stimulation is widening the range of problems you can attempt. Which leads to the practical claim — with execution cheap, asking the right question has become disproportionately valuable. In Takshashila’s weekly session he makes people start there: name the unique problem in your research domain.

• What a 40-person think tank actually ships — a weekly show-and-tell with two or three people rostered to demo, a Mattermost channel where experiments get posted, and roughly 20 of 40 colleagues who have now built their own sites and trackers. Non-engineers learned Git because they had to. The geospatial lead built a mapped analysis of the Iran war’s effect on India with Claude Code that would otherwise have taken two more people. There is a Claude Code plus ElevenLabs “fourth guest” with a persona that steel-mans a position through an episode; a tool that turns a PDF or EPUB into a 45-minute podcast, with Marathi output via Sarvam, on the theory that most non-fiction books are a paper with history stapled to it; an institutional brain that classifies decisions into decision, playbook, or principle so a joiner in three years can find the provenance nobody wrote down; and the Eisenhower-matrix bot. Navin’s request, on air, is that Pranay write the show-and-tell method up — he calls it homework.

• The diffusion balance sheet — Pranay then scores India against Mazarr’s societal factors, one at a time. National ambition and will: better than before, because Viksit Bharat 2047 is a defined and measurable target — a developed country means about $13,500 per capita. Unified national identity: hard here, though UPI shows adoption can outrun education. Shared opportunity: the real promise, which is why speech-to-text and text-to-speech matter so that this does not stay with “the English-speaking elites like us.” An active and engaged society: fine, median age under 30. Effective institutions: the acute weakness — and note that GST’s data architecture, the vaccination drive and direct benefit transfers worked because technology came with institutional redesign, with UIDAI built new rather than bolted onto an existing department. Intellectual climate: credentialism will matter less and our institutions are stuck on it. Diversity and pluralism: an Indian advantage.

• Choices Choices — coffee with Jensen Huang or Dario Amodei: both hosts guess Jensen, and Pranay agrees, on the grounds that it would be good to have a fight with him. Reborn in the early 1980s onto a TSMC fab floor, or exactly where he is now: he takes now, and corrects the premise — TSMC was founded in 1987, so he would have had to be born in the 1960s. A year inside the India AI Mission or inside an Indian semiconductor startup that ships: Ashish guesses the startup, Navin guesses the mission, and Pranay sides with Navin — “Navin knows me better than Ashish” — because getting government policy 1% right is a non-linear payoff, and there are far fewer people who can do that than there are good chip engineers. Finally, model monogamy for a year, Claude or Kimi: “My stated and revealed preferences are different.” He is experimenting with open models, still shipping on Claude and Codex, and asks to be asked again in a month.

Notable quotes

“I don’t think there is anything like sovereign AI. You cannot become sovereign in AI. I mean, you can’t become sovereign in semiconductors — how will you become sovereign in AI? No country can.”

“The fact that you don’t have state capacity is itself reason enough for you to then use AI.”

“I don’t want us to think it should be something Atmanirbhar, that only Indians will build. If it is available to every Indian, then our goal is satisfied.”

“If all our information is through that, that is the challenge. So how do we preserve cognitive autonomy in the AI age?”

“One year of second best is not as good as one year of best.” — Navin Kabra

“I’m more hopeful about India’s diffusion than India’s innovation.”

Links & resources

Everything below was named on air. Where a URL was not given, we have said so rather than guessed.

• When the Chips Are Down — Pranay’s book on semiconductor geopolitics, with Abhiram Manchi.

• Anticipating the Unintended — the newsletter he co-writes with RSJ.

• Puliyabaazi — his Hindi-Urdu podcast on policy and technology.

• Michael Mazarr, The Societal Foundations of National Competitiveness (RAND, 2022) — the source of the diffusion framework the episode is built on, and of the seven societal characteristics Pranay scores India against.

• Kai-Fu Lee, AI Superpowers — cited for its four-wave framework and the claim that enterprise AI is where adoption is slowest and the money is.

• Richard Baldwin — on manufacturing supply chains shortening while services supply chains lengthen.

• XKDR Forum — cited for work on judicial scheduling and case management.

• RISC-V and the RISC-V Foundation — started at UC Berkeley, now based in Switzerland, with MeitY among the sponsors.

• Apache TVM, MLIR, LLVM, ONNX Runtime, PyTorch — the open-source layer Pranay would fund to make GPUs switchable and break the CUDA dependency.

• Synopsys, Cadence, Mentor Graphics (now Siemens) — the three EDA incumbents, and the reason two Bangalore companies are building AI-based tool flows for mixed-signal design.

• Sarvam-105B — Pranay’s “the Sarvam 105”, cited as the existence proof that India does have a sovereign model, and again for the Marathi output in his book-to-podcast tool. The number is both the name and the parameter count.

• DeepSeek, Moonshot / Kimi, Mistral — the open-weight models under discussion. Model versions were not always clear on air.

• OpenCode — the multi-model router Pranay was pointed to after tweeting about Kimi’s performance.

• Mattermost — Takshashila’s open-source internal communication tool.

• Claude Code and ElevenLabs — the stack behind the geospatial mapping tool and the AI “fourth guest.”

• EkStep and Nandan Nilekani — cited on benefits reaching across the economic spectrum rather than trickling down, with farmer advisory tools in Maharashtra as the example.

• Andrej Karpathy’s LLM wiki — Navin’s recommendation, and his argument that every person, not just every company, should keep one.

• Should the US Sell Advanced GPUs to China? An Indian Perspective — Takshashila, 29 April 2026. The paper on more competition being net-net good that frames the first Choices Choices question.

• TheProfesseer — theprofesseer.com. The litigation-analytics engine collecting records from various courts so a litigant can judge an advocate’s track record, which matters because lawyers cannot advertise.

Connect with Pranay Kotasthane

• Takshashila Institution — takshashila.org.in

• Website — pranaykotas.com

• Twitter — @pranaykotas

• Anticipating the Unintended — the newsletter, with RSJ

• Puliyabaazi — the Hindi-Urdu podcast

Our guest is Rohit Srivastwa, one of India’s foremost cybersecurity experts. He founded ClubHack, India’s first hackers’ conference; built a company that was acquired by Quick Heal; exited; and is now building his next one. He has advised governments and their various arms — including work on the drafting of India’s privacy law and the early IT Act — along with large companies, small companies, and individuals, and he has written two books: one on privacy for individuals, one on security for businesses. His thesis in this conversation is symmetrical: AI is not a defender’s tool or an attacker’s tool. It is weaponizing both sides at once, and in India the deciding variable won’t be the technology — it will be whether people bother to understand a risk before they adopt the shiny thing that carries it.

Some of the specifics that stuck with us: a live video call on which Rohit can change his own face so you believe you’re talking to Virat Kohli, and why that breaks the video-KYC that now gates every SIM card and bank account; a scam WhatsApp voice call he received that was “90, 95% similar” to a friend’s actual voice, built from audio that is already public; and, at the other end of the spectrum, his non-technical wife sitting down in Google AI Studio and building a working app to track the kids’ tutors and fee payments — phone-only, because she flatly refused a version that also worked on a laptop. Attack and defence, professional and domestic, all moving on the same curve.

Highlights

• The phishing tell is dead — For years the standard advice was “look for the spelling and grammar mistakes.” AI has erased that signal: a phishing email now arrives in clean, fluent English, so defence has to be rebuilt around behavioural anomalies — is this even a humanly plausible email for this sender? — rather than typos.

• Deepfaked video KYC — India moved to video-KYC for SIM cards and bank accounts, and attackers are now defeating it in real time. Rohit can swap his on-screen face mid-call; on the defence side, a Pune startup he judged as an awards-jury member has built a tool that flags, live, the exact moment a deepfake switches on.

• Thirty seconds of your voice — That’s roughly all a modern voice-cloning tool needs; give it a longer sample and it also learns your pauses and your odd pronunciations. The people most exposed are precisely those whose voice is already all over the internet — podcast hosts included.

• What a SOC actually does — Every device, firewall, and server writes logs one line at a time; a SIEM funnels them into one place, and analysts (graded L0 through L3) hunt for anomalies around the clock. AI has now absorbed the L0/L1 tier — which is either a layoff story or a “far more companies can finally afford to be monitored” story, depending on where you’re standing.

• The cost curves diverge — Ashish tried the economist’s framing: attack and defence both getting cheaper. Rohit corrected him. The cost of attacking is falling exponentially — you only need five or six point-and-click tools — but the cost of defending is going up, because the defender has to understand and block every one of those tools.

• Security is a cost of doing business — Like GST, EPF, or insurance, you don’t get to opt out because you’re small. His CA’s email signature puts it best: if you think compliance is costly, try non-compliance once.

• Ransomware grew a second head — It no longer merely encrypts your files; it copies them out first, so the attacker can also sell your data if you don’t pay. The baseline answer is EDR (endpoint detection and response) — “antivirus plus plus,” with threat intelligence built in to contain the blast radius — but too many mid-market firms still believe a 2010-era firewall and antivirus will hold in 2026.

• 270 standards and counting — Rohit’s next company distils the world’s ~270 security standards (GDPR, India’s DPDP Act, California, Brazil, and on) into plain action items a junior can execute. Only a handful apply to any one firm, but the legal language is unreadable, and the “just press a button and the agent fixes everything” dream runs straight into a governance wall.

• Automation is not governance — An agent told to install software on “laptop_Navin_02” can just as easily reboot a machine in the middle of a client presentation. His cautionary tale is SOAR, a Gartner category that quietly died because buyers refused to let software auto-block and auto-remediate — the blast radius of one wrong automated action was simply too large to accept.

• Grey hair vs. black hair — Grey hair brings structure, governance, and no knee-jerk reactions; black hair wants it shipped now. Both are necessary, and the friction between them is the feature, not the bug — a live illustration of pace layers, the idea Navin traces to Stewart Brand.

• The Indian cheque as a security philosophy — In the US you buy a chequebook at a stationery shop; in India the bank prints it, and the teller can’t credit your account until the cheque actually clears. Slower and more bureaucratic — and it forecloses entire categories of fraud that still work in the West. Ashish reframed it in statistics: the US treats a cheque as innocent until proven guilty, India as guilty until proven innocent, and for AI security we’re probably better off starting from caution.

• ChatGPT is a yes-man — It never says no; you’re always right. That is the opposite of Stack Overflow, where the first five replies tell you not to do it. Wonderful when you know the domain and can steer; dangerous when you don’t and take a confident wrong answer at face value.

• The prompt-injection problem — Navin’s clean explanation: hand an agent your spreadsheets, it reads a web page to help with the analysis, and that page carries a hidden instruction that quietly exfiltrates your data — no malice required from the AI vendor at all. Rohit keeps a live injection sitting in his own LinkedIn “About” section as a running demonstration.

• The security expert’s own smart home — Navin’s caricature: the real expert’s house has exactly one internet-connected device, the printer, “and if it makes a funny noise I will shoot it.” Rohit’s confession is more instructive — his whole house is automated, but the command centre lives inside the house and he reaches it only over VPN with two-factor auth. AI can walk you through that setup, but only if you’re curious enough to ask for the protection layer, not just “download the ISO and run it.”

Notable quotes

• “In my domain, AI is weaponising both — the attackers and the defenders. It’s actually weaponising both.”

• “For years we taught everyone: if there’s a phishing email, look for the spelling mistake, look for the grammatical mistakes. Now that problem is already gone. AI has taken care of it. You don’t see it anymore.”

• “The cost of doing the attack is going down. The skill set required to do a successful attack is going down exponentially. But from a defender’s point of view, you need to understand all of those and start blocking all of them.”

• “My CA uses this line as his email signature: if you think compliance is costly, try non-compliance once.”

• “ChatGPT never says no. You’re always right — it will guide you towards your direction. But it’s not Stack Overflow. On Stack Overflow, the first five comments will be: don’t do it.”

• “The only thing connected to the internet in my house is the printer — and if it makes a funny noise I will shoot it. There’s a gun kept next to it.” — Navin, on the caricature of the paranoid security expert

Links & resources

• Quick Heal — the security company that acquired Rohit’s first venture.

• ElevenLabs — voice-cloning tool cited as needing only ~30-40 seconds of audio.

• EDR (Endpoint Detection and Response) — the “antivirus plus plus” category recommended against ransomware.

• SIEM (Security Incident and Event Management) — the central log-aggregation-and-analysis platform behind every SOC.

• SOAR (Security Orchestration, Automation and Response) — the Gartner-coined auto-remediation category Rohit cites as having failed in the market.

• GDPR (Europe) and India’s DPDP Act — the two named data-protection regimes; California and Brazil also referenced as having their own.

• Wireshark / Ethereal / tcpdump — packet-analysis tools, invoked for the “learn to read it raw first” point.

• Hugging Face — where people casually pull down models to run locally.

• Stack Overflow — as the pre-AI answer culture (“the first five comments say don’t do it”), and as the source of the copy-paste-a-hidden-malicious-command trick.

• Google AI Studio — where Rohit’s wife built her tutor-tracking app.

• Rob Joyce, “Disrupting Nation State Hackers” (USENIX Enigma 2016) — the former head of the NSA’s TAO (Tailored Access Operations) unit on knowing your own network better than the attacker does. (usenix.org/conference/enigma2016/conference-program/presentation/joyce)

• CISA (US Cybersecurity and Infrastructure Security Agency) — via a news story Rohit cites, about its chief uploading confidential data to ChatGPT.

• Pace layers — Stewart Brand’s fast-layer/slow-layer framework, which Navin invokes for the grey-hair/black-hair tension.

• Pune Knowledge on Tap — the “knowledge is free, beer you pay for” meetup Ashish and others started; Rohit notes the last-Thursday-of-the-month cadence.

Connect with Rohit Srivastwa

• Website: rohit11.com

• X/Twitter: @rohit11

• LinkedIn: rohit11

• As he says on-air: “first call is free, after that I’ll bill.”

Prashant Girbane, an IIM alum with a career spanning the UN and TCS, leads MCCIA — a 91-year-old regional chamber with 3,000+ member companies across all 36 districts of Maharashtra, mostly small and medium, mostly from the manufacturing sector. In this episode he lays out his core thesis: while the world races to build the largest models, India’s real opportunity is diffusion — the “cherry on the cake” he adds on top of Jensen Huang’s five-layer AI stack. Access creates opportunity, opportunity creates choice, and choice is freedom.

He walks through MCCIA’s three-stage Applied AI program, shares vivid case studies (an attar maker in Ratnagiri, a dairy farmer optimizing cattle feed, a language trainer who saved ₹2.5 lakh), and explains why an AI session held in a temple, led by a 23-year-old, might be the most telling image of how AI actually reaches India.

Highlights

• The five-layer AI “cake” — Jensen Huang’s framing (energy → semiconductor chip → cloud → large language models → applications) and how Union Minister Ashwini Vaishnaw maps India’s approach onto it.

• The sixth layer: diffusion — Prashant’s central argument for why access and adoption matter more for India than competing at the frontier. “We don’t want to run the race of building the largest model — we want to run the race of fastest diffusion.”

• Stage 1 — Going to the people: colleagues physically visited 13 districts, talked to 1,000 micro and small companies in groups of 20–50, including one session held in a temple in Ratnagiri.

• Stage 2 — One-on-one consultations: a target of 1,000 personalized sessions (430 completed), via a free helpline, identifying common patterns in how MSMEs want to use ChatGPT, Google AI Studio, Gemini, and Copilot.

• Stage 3 — Co-developing “applets”: building no-code tools each useful to 100,000+ companies (attendance management, visitor management, invoice/brochure scanning), co-created with users and shared freely, hosted on the user’s own cloud.

• Real case studies: the attar (scent) maker testing combinations virtually; the dairy farmer optimizing feed for milk yield; the language-training institute that avoided a ₹2.5 lakh marketing quote.

• Who pays for it: legacy and goodwill — member fees, supportive board members, and purpose-driven young engineers (and non-engineers) on the team.

• AI beyond MSMEs: the three big-impact areas — agriculture (crop disease diagnosis from images), healthcare (telemedicine, ABHA), and education (personalized, contextualized access).

• The UPI analogy and the trusted intermediary — if commerce diffusion rode on the local shopkeeper, education diffusion rides on the school teacher.

• The school as a Schelling point — using schools for low-cost AI experimentation and “pre-solving” the workforce problem five years out.

• Government’s role in scale — market failure, the cluster-based model (Auto Cluster, food processing, electronics), and why governments should fund small experiments and scale what works.

• Coping with the pace of change — tokens getting cheaper and models getting better; why “learning to learn” and WhatsApp-group “tribes” matter more than waiting for the next model.

• Advice for young people — don’t fear the technology, you don’t need to code, learn faster, and the difference between “jobs” and “work.”

• What’s next: a digital MCCIA — the AI Studio and helpline now help ~4,200 companies a year digitally (5× the ~700–800 helped in person), embodying Mashelkar and Borde’s “More from Less for More.”

Notable quotes

“Access is opportunities, opportunities are choices, choices are freedom — and there’s nothing more important than freedom.”

“We do not want to run the race of building the largest model. We want to run the race of fastest diffusion.”

“AI takes away tasks, not jobs — and those who use AI will be ahead of those who don’t.”

“In the future you may not have a lot of jobs, but you’ll have a lot of work.” (paraphrasing Dr. Anand Deshpande)

“Everybody doesn’t have to boil the entire ocean. Boil your cup of water, get your tea, feel refreshed — that’s good.”

“Every time we have listened, it has worked. And it has taken less effort.”

Links & resources

• MCCIA (Mahratta Chamber of Commerce, Industries & Agriculture)

• MCCIA MSME Helpline — free expert consultations

• MCCIA MSME Helpline — 100 Case Studies: Real Stories of Guidance and Transformation

• Jensen Huang / NVIDIA — the five-layer AI stack

• Ashwini Vaishnaw, Union Minister of Electronics & Information Technology (MeitY) — on India working across all layers of the five-layer cake

• Dr. R.A. Mashelkar & Sushil Borde — “More from Less for More” (affordable, inclusive innovation)

• Pune International Centre — the think tank Prashant is associated with

Connect with MCCIA

• 🌐 Website: mcciapune.com

• 💼 LinkedIn: linkedin.com/company/mcciapune

• 📸 Instagram: @mccia.pune

• 👤 Prashant Girbane on LinkedIn: prashant-girbane