Get live statistics and analysis of Natasha Malpani 👁's profile on X / Twitter

karpathy says we’re a decade away from AGI, because we don’t yet know how to make systems learn continuously. the deeper problem is that we’ve built this entire field on metaphors, not mechanics. we keep saying AI can think, reason, remember, create. but those are human verbs, not model capabilities. AI isn’t intelligent. it’s efficient. it doesn’t reason . it pattern-matches. it doesn’t remember. it reconstructs. it doesn’t reflect. it re-runs. we confuse language with understanding. just because a model can describe thought doesn’t mean it’s having one. real intelligence has intent. it knows why it’s thinking. AI predicts what comes next. and yet, even without intent, these systems are starting to functionally mimic cognition. they reason, recall, and reflect. not consciously, but effectively. that’s why both statements can be true. AI is a bubble. because capital, hype, and valuations have outpaced genuine capability. but it’s also here to stay. because the direction of progress is right. the crash will clear the noise. what remains will be systems that truly learn. memory that compounds, feedback that refines, intelligence that grows by living inside workflows. we’ll look back on this phase the way we look at the early web: messy and magical. the beginning of machines that finally learn, not just perform.

i’d say the opposite: the real white space is at the application layer. everyone wants to sell shovels, but the gold is in how people actually use them. the infra race is a knife fight between hyperscalers: openai, google, anthropic, meta, amazon. they’ll undercut each other on price, latency, context window, and token cost until margins collapse. developer tooling looks safer, but it’s crowding fast and every improvement gets absorbed upstream by the foundation models or downstream by open-source forks. meanwhile, applications are where behavioural moats form. data isn’t the only barrier. habits are: -users don’t live in apis or eval dashboards; they live in experiences. -context, workflow, brand, and trust compound fast -distribution and feedback loops create data advantages that scale locally even when models converge globally. you win if you own feedback surface to capture every edit, action, and intent/ build domain depth/ embed in daily workflows/ collect proprietary exhaust (behaviour and telemetry that the model providers will never see). some infra will break through- security, evals, low-latency edge, compliance- but the broader white space is still at the application layer, where people, agents, and systems actually interact. go deep enough that a foundation model can’t care, and sticky enough that users won’t leave even when it can.

more than asking why india didn’t build deepseek, we should be learning from: -cursor → AI-native coding assistant ($100M ARR in 21 months) -bolt → AI-powered command bar for enterprise apps ($20M ARR in 2 months) -lovable → AI-powered UI generator for devs ($10M ARR in 2 months) -elevenlabs → AI voice synthesis at scale ($100M ARR in 24 months) -glean → AI-powered enterprise search ($100M ARR in <36 months) -heygen → AI-driven video avatars ($35M ARR in 12 months) -harvey→ AI for legal and enterprise workflows ($50M ARR in 36 months) what worked? speed, distribution, and ai-native ux—not just better models. india's edge isn't in playing catch up—it’s in solving for: -vernacular AI: multilingual assistants for SMBs, education, and commerce -AI-powered enterprise search: india’s data fragmentation is an opportunity, not a barrier -AI-driven legal + fintech automation: streamlining compliance, contracts, and documentation -consumer AI-first tools: education, creator economy, and regional AI-first interfaces the AI playbook is clear—early, fast-moving startups are hitting $10M+ ARR in months, not years. the question isn't if we'll crack the code- it's who will.

shuttling between bangalore + bombay frequently has gotten me thinking: bangalore's startup/vc community is doing a great job of creating hype + then turning it into a self-fulfilling prophecy of growth. the city is an upcoming centre of power as paper wealth converts to real money- but it’s still early as the city transitions from town to thriving metropolis bombay, on the other hand, with its multigenerational wealth feels like it's past its prime, but its not going to go down without a fight, and is clawing its way back to its glory days. the cities reflect the ages & ideologies of their residents: idealism v/s the establishment which would you bet on?

the bangalore/ sf corridor is finally real: founders, capital, and ideas moving both ways. but the conversation needs way more common sense. silicon valley doesn’t fund tourism. it funds ambition, originality. the physics are completely different across the two markets. what actually works in the us: enterprise + vertical ai (trust + network driven) ai infra + devtooling (evals, model orchestration, security, compliance) foundational research + frontier labs: (stanford, berkeley, openai, anthropic: deep tech compounding) to win here, you need presence, references, domain fluency, enterprise relationships, patience for immigration + burn india is built for: -consumer ai (billion-user distribution, low CAC, mass-market behaviour) -smart hardware & robotics (cost curve, manufacturing, gov support, talent density) -space tech, defence, deep ops (no-bloat engineering + regulatory openness) -commerce (zepto, razorpay, meesho, groww: real scale, real money) india’s advantages are structural: cheapest feedback loops in the world, engineering ambition + hunger, cultural comfort with hybrid workflows, vernacular + multimodal by default, cost curves nobody else can match. copying what’s hot in SF won’t work. match the problem to the market. not the other way around. sf still sometimes looks down on india. and india still gets performative about sf. both are useless. credibility is earned, not borrowed. geography amplifies your work, it doesn’t replace it.l i meet founders on both sides every day. the outliers don’t chase vibes or locations, use the corridor strategically, build where their product deserves to live understand distribution > hype and pick markets, not aesthetics. the india/ us corridor is finally compounding. use both, but be honest about which game you’re actually playing.

agents look magical in a demo. they collapse in production. the reason is simple: they can’t remember, can’t ground, and can’t adapt when the workflow drifts. people are throwing everything at the problem: scratchpads, tree logs, bolted-on knowledge graphs. most of it is duct tape. but it’s now clear: memory and context are the moat. knowledge graphs aren’t new. they started in the late nineties, went mainstream when google branded its own in 2012, and have quietly powered search, ads, ecommerce, and fraud detection ever since. outside big tech, they never broke through given they were too expensive to build, too hard to keep fresh, too dependent on specialists. now the equation has changed. language models can extract entities, map relationships, and keep graphs updated in real time. suddenly, graphs are less academic overhead and more practical grounding. but graphs don’t replace search across embeddings. they complement it. what works in production is hybrid retrieval: -wide recall through vectors: “show me things like this” -precise reasoning through graphs: “show me exactly how these things connect” -fused by the agent to deliver both breadth and depth what we’re seeing live: -agents without persistent memory churn out after six weeks. users won’t re-train every morning. -buyers don’t care about clever prompts. they care about trust: the ability to evaluate, to roll back, to audit, and to enforce permissions before an agent takes an action. what’s working: vertical hybrids. a finance agent grounded in accounts, policies, and permissions. a healthcare agent grounded in medical codes, medications, and compliance rules. memory, context and trust aren’t features. they’re infrastructure.

bangalore is not the valley. the us shapes infrastructure markets. india shapes adoption markets. in the US, the AI action is down the stack: tooling, evals, orchestration, data pipelines. refining cognition infrastructure for cheaper inference, safer alignment, and better latency. in india, the action is up the stack: vertical apps in health, education, finance, commerce. founders here are operationalizing AI in chaotic markets with uneven data and deeply human workflows. that’s divergence, not weakness. bangalore loves to mimic america. but the valley’s playbook doesn’t work here: cheap capital, uniform markets, and perfect infrastructure don’t exist in india. AI is a clean slate. the challenge isn’t capability, it’s market design: pricing, trust, distribution, and localization. If we solve for our advantages: our chaos is our edge. if we design for that, we’ll build systems the world will need next. -abundant context, scarce structure: unstructured data (voice, vernacular, whatsapp, paper trails) is perfect for multimodal, low-resource models. -low-cost feedback: we can label and tune faster and cheaper. hybrid workflows: half-human, half-machine systems are already the norm. price elasticity: AI here must run on Rs 5 inference, not $0.05. constraint. -vernacular markets: local ai demand will 10x english-first use. -trust networks: adoption depends on social proof, not ux. ai must be earned, not downloaded. the us will keep build cognition infrastructure. 🇮🇳 will build and scale applied cognition: how AI behaves in messy, emotional environments. we don’t need to copy the valley. we’re not catching up. it’s a different race. where friction is the ultimate stress test for intelligence.

the difference between the tech ecosystem in the valley and 🇮🇳 is obvious even on twitter: we’re terrified of original thought + dissent. we follow scripts, jump on bandwagons, and stay in line. india’s obsession with hierarchy, age and ‘expertise’—paired with a copycat culture, acceptance of the status quo + a tendency to do the bare minimum—will be our undoing. on the other hand, the standards are so low that if you are willing to go the extra mile, it’s strangely easy to stand out.

most people don’t understand how vc works. that’s by design. it’s gatekept. built on FOMO, status, private networks and signaling. pre-seed and seed aren’t a science. they’re an art. taste, access, people, instinct. there’s almost no hard data. the edge is knowing about the ‘good’ (read: hot) deals first, and being trusted enough to get invited in. FOMO comes from the power law: missing an outlier hurts more than getting it wrong. this is social chess as much as finance. the best players don’t just pick. they’re in the room before there’s anything to pick from. in early stage, money follows access. access follows trust. trust follows time. taste decides what you do with it. if you’re building and want to raise (and very few should): don’t debate whether it’s right or wrong. learn the game. and create the FOMO.

the pattern is obvious in every conversation i have in SF with foundational model companies, stanford labs, scaled AI founders and top-tier investors: sf is now: -a magnet for foundational research -a dense cluster of world-class scientists + frontier labs -a concentration of enterprise buyers, infra thinkers -the only place where new architectures, multimodal systems, RL, safety and deep technical work compounds this is the valley’s last true moat. the billion-dollar seed rounds with <$10M revenue is essentially a research arms race. and here’s the uncomfortable truth everyone admits: the AI race will have very few concentrated winners. the hype is loud. the bloodletting will be quiet. meanwhile, india is compounding in the opposite direction: -consumer AI with real distribution -sovereign + vernacular data at billion-user scale -multimodal behaviour the west doesn’t even know how to model -hybrid workflows and human-in-loop culture by default sf is brilliant. and also expensive, lonely, narrative-driven, and brutally competitive. india is chaotic. and also the best real-world testbed on the planet.

recruiting in india today hurts everyone. managers scroll 300 CVs and still say “no talent.” candidates refresh inboxes and hear nothing back. recruiters juggle 200 chats with zero memory. “AI hiring” today is focused on faster rejection. but speed ≠ better hires. the real leak is that every warm intro, interview note, ghosted offer, and notice-period excuse lives in WhatsApp threads and dusty sheets, and then disappears. what an AI recruiter *should* do: -remember: “crushed May task, ghosted in March- 68 % risk” -surface founder-to-founder intros before cold CVs -flag offer drop-off before the letter goes out -ping candidates when intent spikes, not on a CRM timer -let devs apply with a 90-sec voice not, résumé optional the need is obvious: 60 M SMEs hire in WhatsApp + Excel.. 40 % of tech talent ghosts once. 50% of offers die on 60/90-day notice. LLMs + RAG can now turn hiring rubble into a live fit-&-risk graph. early sparks include Kula (outbound), BarRaiser (structured interviews), HireSure (offer risk), but it’s still early days in AI x recruitment. the breakout play is to price on joins closed, not seats sold. founders will pay for certainty per week. my bet is that AI that screens faster will become a commodity. AI that erases regret- by owning memory and closure- will win in the medium-term.

AI is schrödinger’s bubble. it’s both overheated and underpriced until the box opens. gpu valuations feel dot-com, but underneath the froth, trillion-dollar incumbents are the ones paying the bills: microsoft, nvidia, meta, google. bubbles burst when nobody can pay. here, the buyers print money. what’s inflated is the behavior, not the technology re: FOMO, valuations and hype / performance over substance. the crash will clear that layer, not kill the shift. real products will keep compounding. the smartest move right now is to build like the bubble has already popped.

shuttling between bangalore + bombay frequently has gotten me thinking: bangalore's startup/vc community is doing a great job of creating hype + then turning it into a self-fulfilling prophecy of growth. the city is an upcoming centre of power as paper wealth converts to real money- but it’s still early as the city transitions from town to thriving metropolis bombay, on the other hand, with its multigenerational wealth feels like it's past its prime, but its not going to go down without a fight, and is clawing its way back to its glory days. the cities reflect the ages & ideologies of their residents: idealism v/s the establishment which would you bet on?

karpathy says we’re a decade away from AGI, because we don’t yet know how to make systems learn continuously. the deeper problem is that we’ve built this entire field on metaphors, not mechanics. we keep saying AI can think, reason, remember, create. but those are human verbs, not model capabilities. AI isn’t intelligent. it’s efficient. it doesn’t reason . it pattern-matches. it doesn’t remember. it reconstructs. it doesn’t reflect. it re-runs. we confuse language with understanding. just because a model can describe thought doesn’t mean it’s having one. real intelligence has intent. it knows why it’s thinking. AI predicts what comes next. and yet, even without intent, these systems are starting to functionally mimic cognition. they reason, recall, and reflect. not consciously, but effectively. that’s why both statements can be true. AI is a bubble. because capital, hype, and valuations have outpaced genuine capability. but it’s also here to stay. because the direction of progress is right. the crash will clear the noise. what remains will be systems that truly learn. memory that compounds, feedback that refines, intelligence that grows by living inside workflows. we’ll look back on this phase the way we look at the early web: messy and magical. the beginning of machines that finally learn, not just perform.

i’d say the opposite: the real white space is at the application layer. everyone wants to sell shovels, but the gold is in how people actually use them. the infra race is a knife fight between hyperscalers: openai, google, anthropic, meta, amazon. they’ll undercut each other on price, latency, context window, and token cost until margins collapse. developer tooling looks safer, but it’s crowding fast and every improvement gets absorbed upstream by the foundation models or downstream by open-source forks. meanwhile, applications are where behavioural moats form. data isn’t the only barrier. habits are: -users don’t live in apis or eval dashboards; they live in experiences. -context, workflow, brand, and trust compound fast -distribution and feedback loops create data advantages that scale locally even when models converge globally. you win if you own feedback surface to capture every edit, action, and intent/ build domain depth/ embed in daily workflows/ collect proprietary exhaust (behaviour and telemetry that the model providers will never see). some infra will break through- security, evals, low-latency edge, compliance- but the broader white space is still at the application layer, where people, agents, and systems actually interact. go deep enough that a foundation model can’t care, and sticky enough that users won’t leave even when it can.

the difference between the tech ecosystem in the valley and 🇮🇳 is obvious even on twitter: we’re terrified of original thought + dissent. we follow scripts, jump on bandwagons, and stay in line. india’s obsession with hierarchy, age and ‘expertise’—paired with a copycat culture, acceptance of the status quo + a tendency to do the bare minimum—will be our undoing. on the other hand, the standards are so low that if you are willing to go the extra mile, it’s strangely easy to stand out.

more than asking why india didn’t build deepseek, we should be learning from: -cursor → AI-native coding assistant ($100M ARR in 21 months) -bolt → AI-powered command bar for enterprise apps ($20M ARR in 2 months) -lovable → AI-powered UI generator for devs ($10M ARR in 2 months) -elevenlabs → AI voice synthesis at scale ($100M ARR in 24 months) -glean → AI-powered enterprise search ($100M ARR in <36 months) -heygen → AI-driven video avatars ($35M ARR in 12 months) -harvey→ AI for legal and enterprise workflows ($50M ARR in 36 months) what worked? speed, distribution, and ai-native ux—not just better models. india's edge isn't in playing catch up—it’s in solving for: -vernacular AI: multilingual assistants for SMBs, education, and commerce -AI-powered enterprise search: india’s data fragmentation is an opportunity, not a barrier -AI-driven legal + fintech automation: streamlining compliance, contracts, and documentation -consumer AI-first tools: education, creator economy, and regional AI-first interfaces the AI playbook is clear—early, fast-moving startups are hitting $10M+ ARR in months, not years. the question isn't if we'll crack the code- it's who will.

bangalore is not the valley. the us shapes infrastructure markets. india shapes adoption markets. in the US, the AI action is down the stack: tooling, evals, orchestration, data pipelines. refining cognition infrastructure for cheaper inference, safer alignment, and better latency. in india, the action is up the stack: vertical apps in health, education, finance, commerce. founders here are operationalizing AI in chaotic markets with uneven data and deeply human workflows. that’s divergence, not weakness. bangalore loves to mimic america. but the valley’s playbook doesn’t work here: cheap capital, uniform markets, and perfect infrastructure don’t exist in india. AI is a clean slate. the challenge isn’t capability, it’s market design: pricing, trust, distribution, and localization. If we solve for our advantages: our chaos is our edge. if we design for that, we’ll build systems the world will need next. -abundant context, scarce structure: unstructured data (voice, vernacular, whatsapp, paper trails) is perfect for multimodal, low-resource models. -low-cost feedback: we can label and tune faster and cheaper. hybrid workflows: half-human, half-machine systems are already the norm. price elasticity: AI here must run on Rs 5 inference, not $0.05. constraint. -vernacular markets: local ai demand will 10x english-first use. -trust networks: adoption depends on social proof, not ux. ai must be earned, not downloaded. the us will keep build cognition infrastructure. 🇮🇳 will build and scale applied cognition: how AI behaves in messy, emotional environments. we don’t need to copy the valley. we’re not catching up. it’s a different race. where friction is the ultimate stress test for intelligence.

recruiting in india today hurts everyone. managers scroll 300 CVs and still say “no talent.” candidates refresh inboxes and hear nothing back. recruiters juggle 200 chats with zero memory. “AI hiring” today is focused on faster rejection. but speed ≠ better hires. the real leak is that every warm intro, interview note, ghosted offer, and notice-period excuse lives in WhatsApp threads and dusty sheets, and then disappears. what an AI recruiter *should* do: -remember: “crushed May task, ghosted in March- 68 % risk” -surface founder-to-founder intros before cold CVs -flag offer drop-off before the letter goes out -ping candidates when intent spikes, not on a CRM timer -let devs apply with a 90-sec voice not, résumé optional the need is obvious: 60 M SMEs hire in WhatsApp + Excel.. 40 % of tech talent ghosts once. 50% of offers die on 60/90-day notice. LLMs + RAG can now turn hiring rubble into a live fit-&-risk graph. early sparks include Kula (outbound), BarRaiser (structured interviews), HireSure (offer risk), but it’s still early days in AI x recruitment. the breakout play is to price on joins closed, not seats sold. founders will pay for certainty per week. my bet is that AI that screens faster will become a commodity. AI that erases regret- by owning memory and closure- will win in the medium-term.

agents look magical in a demo. they collapse in production. the reason is simple: they can’t remember, can’t ground, and can’t adapt when the workflow drifts. people are throwing everything at the problem: scratchpads, tree logs, bolted-on knowledge graphs. most of it is duct tape. but it’s now clear: memory and context are the moat. knowledge graphs aren’t new. they started in the late nineties, went mainstream when google branded its own in 2012, and have quietly powered search, ads, ecommerce, and fraud detection ever since. outside big tech, they never broke through given they were too expensive to build, too hard to keep fresh, too dependent on specialists. now the equation has changed. language models can extract entities, map relationships, and keep graphs updated in real time. suddenly, graphs are less academic overhead and more practical grounding. but graphs don’t replace search across embeddings. they complement it. what works in production is hybrid retrieval: -wide recall through vectors: “show me things like this” -precise reasoning through graphs: “show me exactly how these things connect” -fused by the agent to deliver both breadth and depth what we’re seeing live: -agents without persistent memory churn out after six weeks. users won’t re-train every morning. -buyers don’t care about clever prompts. they care about trust: the ability to evaluate, to roll back, to audit, and to enforce permissions before an agent takes an action. what’s working: vertical hybrids. a finance agent grounded in accounts, policies, and permissions. a healthcare agent grounded in medical codes, medications, and compliance rules. memory, context and trust aren’t features. they’re infrastructure.

building in voice AI already looks crowded from the outside. but the real flip will happen in sales. support was the easy wedge: faqs, ticket routing, refunds. useful, but incremental. the bigger shift will be in sales roles that are mostly script, compliance, and repetition: pharma: reps spend hours repeating approved detail. an agent can deliver it word-for-word, capture objections, schedule follow-ups. humans focus on trust, not repetition. real estate: top-of-funnel is budget, location, timeline on repeat. an agent can qualify hundreds of leads overnight, push everything into crm, and hand off when negotiation starts. insurance + banking : loan pre-quals, kyc calls, policy explainer scripts - the backbone of sprawling call centers. agents compress cycles, reduce cost, and log every step. why now? speech recognition finally keeps callers on the line. synthetic voices no longer sound uncanny. crm hooks turn every call into analytics humans never captured. what’s still missing: persistent memory (so buyers don’t re-explain themselves), trust rails (so pharma stays on-script and banks don’t mislead) and seamless handoff when the call gets complex. support was the demo. sales is where the economics snap: the agent runs the script, the human earns the yes.

the bangalore/ sf corridor is finally real: founders, capital, and ideas moving both ways. but the conversation needs way more common sense. silicon valley doesn’t fund tourism. it funds ambition, originality. the physics are completely different across the two markets. what actually works in the us: enterprise + vertical ai (trust + network driven) ai infra + devtooling (evals, model orchestration, security, compliance) foundational research + frontier labs: (stanford, berkeley, openai, anthropic: deep tech compounding) to win here, you need presence, references, domain fluency, enterprise relationships, patience for immigration + burn india is built for: -consumer ai (billion-user distribution, low CAC, mass-market behaviour) -smart hardware & robotics (cost curve, manufacturing, gov support, talent density) -space tech, defence, deep ops (no-bloat engineering + regulatory openness) -commerce (zepto, razorpay, meesho, groww: real scale, real money) india’s advantages are structural: cheapest feedback loops in the world, engineering ambition + hunger, cultural comfort with hybrid workflows, vernacular + multimodal by default, cost curves nobody else can match. copying what’s hot in SF won’t work. match the problem to the market. not the other way around. sf still sometimes looks down on india. and india still gets performative about sf. both are useless. credibility is earned, not borrowed. geography amplifies your work, it doesn’t replace it.l i meet founders on both sides every day. the outliers don’t chase vibes or locations, use the corridor strategically, build where their product deserves to live understand distribution > hype and pick markets, not aesthetics. the india/ us corridor is finally compounding. use both, but be honest about which game you’re actually playing.

the pattern is obvious in every conversation i have in SF with foundational model companies, stanford labs, scaled AI founders and top-tier investors: sf is now: -a magnet for foundational research -a dense cluster of world-class scientists + frontier labs -a concentration of enterprise buyers, infra thinkers -the only place where new architectures, multimodal systems, RL, safety and deep technical work compounds this is the valley’s last true moat. the billion-dollar seed rounds with <$10M revenue is essentially a research arms race. and here’s the uncomfortable truth everyone admits: the AI race will have very few concentrated winners. the hype is loud. the bloodletting will be quiet. meanwhile, india is compounding in the opposite direction: -consumer AI with real distribution -sovereign + vernacular data at billion-user scale -multimodal behaviour the west doesn’t even know how to model -hybrid workflows and human-in-loop culture by default sf is brilliant. and also expensive, lonely, narrative-driven, and brutally competitive. india is chaotic. and also the best real-world testbed on the planet.

group chat was built for talk. work needs movement. adding AI doesn’t fix it. it just summarizes the noise. Instead of adding ‘AI in chat’, how do we rethink coordination itself: -humans and agents shouldn’t message each other, they should act from a real-time source of truth -conversations should branch, not scroll -context should compound, not disappear. -communication should only happen when alignment breaks. the future of might look more like a living kanban board.

AI is schrödinger’s bubble. it’s both overheated and underpriced until the box opens. gpu valuations feel dot-com, but underneath the froth, trillion-dollar incumbents are the ones paying the bills: microsoft, nvidia, meta, google. bubbles burst when nobody can pay. here, the buyers print money. what’s inflated is the behavior, not the technology re: FOMO, valuations and hype / performance over substance. the crash will clear that layer, not kill the shift. real products will keep compounding. the smartest move right now is to build like the bubble has already popped.