and the intelligence layer. two and a half days to replace familiarity with understanding: what this thing is, what india does with it, and where it must never decide.
Anant Srivastavamentor · ai fundamentals & the intelligence layer
cold open
you've used AI every day for ten weeks. today: what have you actually been talking to?
75 years in nine moments · 1 of 3
the slow years
75 years in nine moments · 2 of 3
the machinery arrives
75 years in nine moments · 3 of 3
the fast years
pivot
it doesn't know anything. so how does it answer everything?
the core mental model
prediction, not knowledge
the honest slide
hallucination is structural
the machine's job is to produce the most plausible next word, and it does that job perfectly even when the plausible answer is false. a confident wrong answer isn't the system failing. it's the system working, on the wrong question.
so it's not a bug being patched out next quarter. it's a property being managed: grounding, retrieval, checking. the founders who understand this stop asking "is AI reliable?" and start asking "where can I afford this failure mode?"
three generations
narrow, generative, agentic
vocabulary · the six foundations
the six that will still be true in five years
LLMa large language model: the prediction machine from this morning, at planetary scale.
hallucinationa fluent, confident, false answer. structural, not accidental.
RAGretrieval-augmented generation: let the model look things up before it answers, so it predicts from your documents, not its memory.
agenta model given tools, memory and a goal, allowed to take steps on its own.
fine-tuningadditional training on your own examples, to specialise the general machine.
token & context windowthe pieces text is chopped into, and how many of them the model can hold in mind at once. its working memory, and its meter.
vocabulary · the live-2026 six · this list ages by the month
the six people are saying right now
currentMCPmodel context protocol: the plug standard that lets one AI talk to many tools. why agents suddenly work.
currentcontext rotquality decaying as a conversation grows: the model drowning in its own context.
currentvibe codingbuilding software by describing it and steering, not typing it. real products ship this way now.
currenttokenmaxxingcramming maximum useful context per token spent: the new frugality.
currentcontext engineeringthe craft formerly known as prompting, grown up: what the model sees, in what order, from where.
currentAI-native companya company designed after AI: workflows assume the machine, headcount assumes the leverage.
lab a · this afternoon · 75 minutes
the prompt tournament
formatsame task, every teamone real business task, three escalating rounds. same model, same time box: only the instructions differ.
judgingblind, on outputoutputs go up anonymised. the room ranks them before authors are revealed; anant breaks ties and names why the winner won.
the pointthe gap is the lessonidentical machine, wildly different results. by round three you'll know exactly which words did the work.
mission 10 · briefed today · submits with the PM mission by mon, oct 5
put AI to work for a real business
a real business you can reach: family, mission contact, or your own. find one workflow where AI genuinely helps, implement it with them, and document what changed: time, money, quality, or honesty about why it didn't.
submits jointly with the product management mission: the same business can serve both. choose it well this weekend.
day 2 · saturday
silicon valley built it. now watch what india does with it.
case · sarvam ai
the language moat
english-first AI serves perhaps 100 million indians well and 800 million badly. sarvam's bet: models built for indic languages from the ground up, priced for india.
indic tokensindian scripts chop into far fewer tokens on a model built for them: the same rupee buys multiples more hindi than on a western model. directional
~₹1/minutevoice AI priced like a phone call, because for most of india, voice IS the interface
800 millionpeople who don't work in english: not a segment of the market. the market
case pair · frugal ai
zoho vs salesforce
salesforceAI as a premiumintelligence sold as an add-on: per seat, per usage, on top of software already priced for enterprises. the machine is a revenue line.
zohoAI as an ingredientown models, own data centres, intelligence folded into the existing price. built frugal from chennai and tenkasi, because their customer is a business that counts every rupee.
same technology, opposite theories of who it's for. frugality here isn't cheapness: it's a different answer to "who deserves this?"
case · niramai · you named this void type last week
AI into the healthcare void
breast cancer screening in india needs machines, radiologists and hospitals that most districts don't have: a textbook institutional void. niramai's answer: a thermal scan read by AI, no radiation, no touch, no radiologist on site, cheap enough for a camp in a small town.
the AI isn't replacing an institution here. it's standing in for one that never existed. last week's question, "fill, bridge, or wait", just got a new answer: fill, with intelligence instead of infrastructure.
where AI deploys first
payments vs healthcare: the four preconditions
the leveller, made real
the pune solopreneur
one person in pune runs what used to be an agency: AI drafts the proposals, produces the first cut of every design, answers routine client mail, and prepares the books for the CA. clients pay for judgement and taste. the machine does everything that used to justify headcount.
ten weeks ago this room met the idea that AI is a leveller. this is what it looks like with a face: the smallest possible company, competing on the largest possible leverage.
pivot
so what does a founder actually do about it?
the decision
build, buy, or partner
buildwhen the intelligence IS the productsarvam builds models, because the model is the company. costliest, deepest moat.
buywhen it's a tool, not the businessthe pune solopreneur buys subscriptions. so does almost every business you'll run: the leverage without the lab.
partnerwhen someone owns what you lacka hospital chain partnering with niramai gets the screening layer without pretending to be an AI company.
the honest slide
when NOT to build
data you don't havea model is only as good as what it learns from. if the data lives with someone else, so does the moat.
problems that are featuressome friction is why customers trust you: the human call, the site visit, the handshake. automate it and you've deleted the product.
hype masquerading as need"we need an AI strategy" is not a customer problem. last week's test applies: does the pain survive normal conditions?
india lens & trends · discussion, not lecture
the room's turn: where does this land?
which business from your missions this term gets transformed first by what you've seen today, and which is safest?
where in kerala is the sarvam story waiting to happen: the void AI can fill in language, health, or credit?
india lens & trends · continued
and where should it not land?
name one workflow from a business you know where automating it would quietly destroy the trust the business runs on.
what's the "uncomfortable middle" job in your own family's work, and who in the family holds it?
day close
monday morning: the two questions that decide whether AI helps you or embarrasses you.
day 3 · monday · first half only
the last morning: honesty.
question one · what automates
the automation pyramid, taught honestly
question two · where AI never decides
firing someone. pricing that can ruin a customer. anything medical, legal, or irreversible. the machine drafts, ranks, flags, recommends.
who is accountable? the founder. always.
parked, deliberately
the governance challenge
the governance challenge returns when you have a product of your own to govern.
close and bridge · by 12:45 sharp
mission 10 submits oct 5, with your PM mission.
this afternoon: vaman. what to build, and what to refuse.
the school of future