The AI Assistant is Finally Here. Notes from the Start of the Cycle

The AI Assistant is Finally Here. Notes from the Start of the Cycle

The consumer AI-assistant cycle has finally started — because the technology finally works. After a year spent meeting founders building assistants, Harsha Kumar argues the no-brainer product was always the obvious one: not a chatbot that answers, but an assistant that acts. The market is behaving like she’s right. What follows are her notes on where the moats are and are not, which wedges work for India, and why the family — not the individual — may be the ideal customer.

The market is behaving like a cycle has begun

Instinct repriced from a $50M round in April to a $2.5B valuation by August — without a public launch or even a price. Town is reportedly raising $1B. In India, M raised ₹102 Cr ($12M) pre-product, and Equal AI reports a million monthly users on call screening alone. Every lab is in the race: a ChatGPT lifestyle app store, Gemini reading your Gmail and Photos, Meta AI inside WhatsApp, SpaceXAI’s Grok Bot, a new Siri on the way. The money is not debating whether assistants matter; it is fighting over who becomes the default.

Why 2015–17 failed — and why this time differs

The last cycle — Magic, GoButler, Fin, Facebook M — ran on a single bet: humans do the work while the software learns and takes over. It was a good story and a bad match for the technology. That era’s machine learning learned one narrow task from thousands of examples, while consumer requests were thousands of tasks with a handful of examples each. Automation covered the repetitive slice; humans kept the rest. So cost grew with every user, and the price exceeded what users would pay. This time the models can actually do the long tail of tasks, which flips the unit economics from “human-supported software” to “software-supported human” — or just software.

The two takeaways that reframe the product

Takeaway 1: human-assisted AI survives only at the ultra-premium end — executives who won’t spend time giving context, exclusive services like exotic destinations, impossible restaurant tables, front-row seats. For everyone else it’s AI over WhatsApp or a call with no human intervention. Human-in-the-loop becomes a premium feature, not a crutch. Takeaway 2: the personal-vs-professional assistant distinction collapses. One entry point, with data silos behind it — work vs personal, sales vs financial data. The assistant is singular; the contexts are many.

The economics are brutal and weird

The parts users love cost the most. Remembering everything means indexing every email and message as it arrives; being proactive means inference running in the background. Quality tracks model spend: the same product on a cheaper model is visibly worse. Founders are subsidising acquisition today, which puts viable consumer economics 1–2 years out. Instinct’s bet is explicit: raise more than anyone, run the best model on everything, let brand and habit set in before the bill comes due.

There is no data moat — except one

Your messages live in WhatsApp; a competitor can pull the same history in minutes. Raw data is not a moat. The only data moat is derived context — things the assistant learned on its own platform: that you hate aisle seats, how you write to your boss versus your spouse. The real moat is brand and habit, with a thin head start from ingesting everything within the first weeks.

The labs own the default — for now

Nine in ten consumers try their general assistant first; fewer than one in ten ChatGPT users ever visit another provider. The labs also set the price: ChatGPT costs $20 while a startup paying the same labs for hard tasks must charge $50. The window is that OpenAI’s agent features haven’t landed and it still doesn’t ask for the email/calendar access that makes an assistant useful — but, as Kumar notes, the window closes daily.

Speed matters more here than anywhere

Without moats, the assistant is a commodity, so the market is already behaving like one: valuations sprinting without a public launch, Poke crossing 100M messages between March and June and getting acquired in July. Churn’s quiet cause: people don’t know what to delegate — “one app per task” muscle memory runs deep. Onboarding that teaches delegation retains; proactive assistants that scan email and messages and prompt new tasks win. “A task not done is worse than a task not done well.” And a telling detail on engineering priorities: at least one leading assistant quietly put humans behind its purchase flows, because reliable browser automation was simply more expensive to run. Most of the engineering budget goes to keeping up.

Every channel is a trade-off

Messaging (iMessage/WhatsApp) is free reach but capped: no task list, no rich UI, and the platform writes the rules — Meta banned general-purpose AI chatbots on WhatsApp last October. Gmail is the only channel with a real API, but it selects for executives. Slack is work-only. Your own app has no ceiling — tasks, suggestions, relationship — but install friction is real, and asking for email/calendar access at signup loses about a third of users. And the surface everyone wants, the OS, belongs to Apple and Google.

The wedge ladder — and why India starts with a phone call

Wedge choice decides everything. A restaurant-booking concierge is acquisition without retention — too low-frequency. An email/calendar wedge has frequency but selects only executives: a homemaker, student, shopkeeper, or nurse has little inbound email; their coordination runs on WhatsApp, calls, and errands. India’s special wedge is call screening: absurd spam-call volumes (numbers handed out at storefronts, UPI running over mobile numbers), unusually high NPS and frequency, selects everyone with a phone — not just inbox people — and lives in an app the platforms can’t switch off. From there: call-making → offline vendor discovery → staff delegation → the family as a unit. A family butler has natural lock-in and network effects, solves the delegation gap (a household generates tasks daily), and holds a dataset no digital platform has — the one place the data-moat argument reverses. One caveat: Bharat needs a far more guided, hand-held product — possibly a different interface paradigm, not a translated, stripped-down app. A real opportunity, but possibly a short-lived one.

• The cycle started because the tech finally works on the long tail of tasks — 2015’s human-in-the-loop bet was a technology mismatch, not a go-to-market failure.

• Consumer AI economics run on subsidy today; viable unit economics are 1–2 years out, and fundraising is itself a short-term moat.

• There is no raw-data moat. Derived context plus brand and habit is the whole game — and the labs own the default entry point.

• Wedge > vision: low-frequency concierge dies, exec-only wedges cap out; in India, call screening is the rare wedge with frequency, NPS, and universal selection.

Why AI Is Booming, but Productivity Isn’t

In June Chamath declared vibe coding dead and said ROI-driven analysis of AI was about to go from nice-to-have to necessity. This is the follow-through: a full framework for what AI ROI actually is, and why it’s invisible in GDP numbers so far. His answer, in short — the tools are fine; the humans are the bottleneck.

The arithmetic that reframes the debate

On the cost side, Ramp’s AI Index — tracking what 70,000+ US businesses spend on AI — finds the median company spends $12.50 per employee per month. Against an average employee cost of ~$8,500 a month, AI only has to make that person ~0.15% more productive to pay for itself: about three extra productive minutes a week. The token bill is close to a rounding error. The real costs are intangible: change management, re-training, the rest of the corporate machinery.

Value is harder — so use productivity as a stand-in

At P&G, one person using AI produced product proposals as good as a two-person team without it. MIT economists showed automation pays when AI steps are grouped so a person checks the work once, not after every step. But here’s the catch: most of a corporate employee’s week goes to coordination, meetings, alignment, and process sign-offs — not producing work. What if humans are the bottleneck to AI ROI? If the saved minutes flow into another meeting or another week waiting on legal, the return is zero. AI may multiply output rather than add to it, so the best firms pull further away.

The experiment that settles it

A 2026 experiment with 515 startups gave every firm identical AI tools, training, and support — but a random half also saw how other firms reorganised work around AI. For the typical startup, revenue barely moved in either group. But almost all the gains came from the top 10% of firms, and the treatment group pulled furthest ahead — 1.9× the control group’s revenue. The tools were identical; what separated the leaders was how they rebuilt their work around them. AI lowers the cost of doing things, but does nothing about whether they were worth doing in the first place.

Chamath lines up the canon: Gates’ magnification rule, Musk’s five-step process (question every requirement → delete → simplify → speed up → only then automate — Musk said his biggest Tesla mistake was doing those steps backward), and Solow’s 1987 paradox: “You can see the computer age everywhere but in the productivity statistics.” Productivity grew ~1.5% a year from 1973–1995, then ~3.3% a year from 1995–2003 — once work was reorganised around the machines. We are in the 1987–1995 interregnum again, waiting on the reorganisation.

• AI’s break-even bar is absurdly low (0.15% productivity) — the spend was never the problem.

• The gains concentrate: in a 515-startup experiment, nearly all ROI went to the top 10% — the firms that rebuilt workflows, not the ones with better tools.

• Sequence matters: delete, simplify, speed up first; automate last. Musk learned it backward at Tesla.

• History rhymes: the Solow paradox resolved only after a decade of reorganisation. We’re in that decade now.

Lenskart Got Faster Without Anyone Getting Faster

A $12B public company — Lenskart — got measurably faster across ten departments without a single employee getting faster. Ashwin’s team built a world model of the company from its own records: not faster individuals, but faster handoffs. The arithmetic is the point — the gains show up in the P&L, without the theatrics of token-maxxed leaderboards. It’s Chamath’s thesis in the wild: the company got more productive because the system changed, not the people.

The Amplification Gap

Nikunj walked out of a meeting with an early-stage life-sciences founder with deep impostor syndrome — “was he talking about me?” Every early-stage founder tells him the same story. The essay is about the middle: people who are AI-curious but haven’t “fully gone off the deep end.” His argument: AI alone doesn’t amplify you anymore, because someone with AI can theoretically out-amplify you. The edge is agency — the “no fear” approach of constantly spinning up and not asking — combined with leaning hard on what you’re “out of distribution” at, and becoming 10× better at it using AI. Good product instincts still matter for generative ideas; the tools just raise the floor. “I feel like I need to rewire my brain, even though I am trying to do that actively.”

What Counts as a Moat in the Age of AI?

A friend running a nine-figure AI services business has “no conviction” about his moat. So Diamandis went asking — Astro Teller, Steve Jurvetson, a training-data founder, Dave Blundin and Ramez Naam on the Moonshots podcast — and came back with an obituary and a shortlist. Buffett popularised the moat in the 1990s; AI collapses the time and effort that made moats work.

The most valuable moat on earth is leaking

CUDA — NVIDIA’s moat of moats. Dave Blundin on the Aug 28 Moonshots episode: “Every AI researcher…” is working around it. Ramez Naam on Aug 4: “The CUDA moat is broken this year and next.” Jensen Huang saw it coming. Diamandis’ rule: your moat is only good until the next moat. Chips are the easy case — the same erosion hits P&G-style brand fortresses (Gillette’s playbook, copied by frontier models overnight) and SaaS (cloned by the models themselves). He cites the Coasean Singularity paper from MIT/Harvard: when transaction and coordination costs collapse, the firm itself gets unbundled.

The macro caveat

1999, the internet: the first owners rarely won — the second owners did. Diamandis’ closing drill is practical: entrepreneur — pitch your rate of reinvention, not your features; executive — audit your moat list and cross out what AI copies in ninety days; investor — underwrite management speed; student — learn faster than the tools change; parent — teach reinvention. “If AI could copy everything you’ve built in ninety days, what would you have left? Whatever your answer is, that’s your actual moat. Go widen it.”

• Patents, code, features, and distribution are off the moat list — AI copies all four in ~90 days.

• The three survivors: speed of reinvention, capital velocity, brand.

• Second owners won the internet; position for the second act, not the first hype cycle.

The Application Layer Strikes Back

“There’s a popular argument that foundation models will capture all the value in AI, leaving little for the application layer. Higgsfield is proving the opposite thesis.” The numbers: Higgsfield crossed $1B in annualised revenue and grew 20× since September 2025. The mechanism is the counter-thesis itself — better models expand what enterprises can achieve and thereby increase the value of applications embedded in their workflows. The foundation layer doesn’t eat the app layer; it fertilises it.

To Build a Cool Product Now, You Need Two People

“One dude with exceptional vision & taste who knows what should exist & can provide the right feedback loops for iteration. & one dude who is absolutely relentless about making it real no matter what. that’s it.” (He clarifies it wasn’t meant to be gendered.) The thread’s replies sharpen it: his first hire was a bad idea with lasting negative impact — the two-person rule is as much about who you don’t add. In the age of coding agents, headcount is a liability and taste is the scarcest input.

A Snack Catalog for Kids, Built with Claude in the Loop

Obsessed with self-driving companies for over a year, tried a lot of products, none met the bar — so he built one himself with Claude in the loop. First experiment: nosugarforkids.com, a catalog of only healthy snacks for kids (“surprisingly this didn’t exist”), with chat for recommendations and a full-blown MCP/WebMCP so you can point your agent at it. His life philosophy is “everything in moderation” — but what kids intake matters supremely. The pattern is the story: taste identifies the gap, the agent builds the thing, MCP makes it agent-addressable from day one.

Jev, and the Memory Problem in AI

Dhravya wrote up how TypesafeAI’s Jev can improve AI memory — walking through all the approaches “in quite detail, with a bunch of benchmarks and cool things in the middle.” It’s the engineering companion to this edition’s lead feature: Harsha Kumar says the only data moat is derived context — things the assistant learned on its own platform. Jev is the machinery for building exactly that: memory as a system, not a prompt trick.

The 90th-Percentile Educational YouTuber Is Fully Automated

Deedy generated an entire 3Blue1Brown-style video — eight minutes summarising the paper “Regularized Recursive Self Improvement of Agent Harnesses” — with Opus 5.5. The stack, which he didn’t even specify: manim for the visuals, Kokoro-82M for the voice, ffmpeg to compile. The model figured out the pipeline and generated it automatically. (He concedes the obvious: the real 3Blue1Brown is still better — but video is “far more approachable to any reader than actually reading.”)

One Shot: Bacterium to Quantum Fields

“Claude 5.5 is crazy good with animations.” One prompt — a self-contained, aesthetically pleasing, scientifically accurate animation starting at a single bacterial cell, zooming through DNA, atoms, ending at quantum fields — produced a 1:47 video. It even added an original score. His thread shares the full prompt. Pair with Deedy’s piece above: the model is now a competent motion-graphics studio.

Notes from a Clipper Who Wanted You to Watch

“You’ve watched 100+ clips today. Here’s notes from a clipper who wanted you to.” Clipping in one line: taking long-form material and turning it into portable short-form stories that travel. The essay covers what changed (an organised distribution system, not just reposts), what’s eligible, what not to clip, the mechanism (feed behaviour), how brands ruin it, 7 questions to evaluate a clipper, Editor vs Network vs Agency, and how to start: one source asset, a licensed pilot, track retention/saves/shares. She runs ‘The Distribution Cafe’, an offline session IP.

The Bitter Lesson of Outbound: Fix TOFU, Not Conversion

“Improving TOFU is always better than improving conversion” — h/t Sam Blond on the Uncapped pod. More top-of-funnel beats a better-converting funnel. In the replies he adds the calibration note: pre-PMF GTM has to be more measured than post-PMF — the venture equivalent of “if you don’t have high loss rates, you’re not taking enough risk.”

How to X-Ray a Company in 15 Minutes

A carousel walkthrough for reading a company fast — the kind of post that earns 1,272 bookmarks because it’s a checklist you’ll reuse. The discipline: separate narrative from numbers, find the unit economics before the story, and check where the cash actually goes. Fifteen minutes, honestly applied, beats fifteen hours of vibes.

• Read the cash flow before the narrative — stories are written, cash isn’t.

• Find the unit economics: what does one more customer actually cost and return?

• Check who gets paid first when things go wrong — the capital stack tells the truth.

• Fifteen honest minutes > fifteen hours of vibes.

The Signs You’re a Generational Outlier

“Outliers are secret weirdos” — they live a life most people don’t understand. Denning, fascinated with outliers since Gladwell’s book, admits the uncomfortable thing: “I have probably become an outlier.” Six signs follow, and they’re worth reading straight — but the spine is the Three Extremes Rule (from Zach Pogrob, Aura’s founder): extreme focus (one thing), extreme timeline (one decade), extreme iteration (1% better every day). Trauma as an amplifier. A chip on the shoulder as fuel — “quiet rage” with a target. Obsession at the risk of looking like a maniac. Eccentricity as a feature, not a bug: non-linear careers, miniature trains, the strange stuff you do in the shadows.

How to Become Disgustingly Self-Disciplined

The most-bookmarked piece in this edition, by a mile. Koe’s system starts with a provocation: become so disgusted with where you are that you have no option but to win. The rocket analogy carries it — 85–90% of a rocket’s mass is fuel for launch. Launch fuel is motivation, pain, willpower, tactical stress. Cultivate disgust for the life you don’t want; spend a week observing society and let it work on you.

Identity, not discipline

“You aren’t disciplined because you aren’t the person who doesn’t view it as discipline.” Identity change first (his examples: Elon, Arnold), then James Clear’s line: be more concerned with your current trajectory than your current results. Then the input/output equation — Input = everything you’re exposed to; Output = your decisions, actions, works — with four input rules: information diet, people, interrogate beliefs, active over passive consumption.

India Is Inevitable

Every country that has manufactured consumer electronics at scale has eventually produced its own global champions — Japan did, Korea did, China did (Pei watched it over eight years living there). India is next: the world’s second-largest smartphone manufacturer, the largest youth population, one of the deepest software talent pools — but no global consumer-electronics brand of its own. Yet.

R&D is the whole game

The how is R&D: the ability to design, engineer, and keep improving a product in-house — what separates a company that lasts from one that rides a moment. The cautionary tale is 2015: Indian brands sold close to half the smartphones in the country; one was briefly India’s best-selling brand, ahead of Samsung. Within two years foreign brands arrived with what the domestic ones lacked — real engineering. The locals were licensing off-the-shelf designs, changing the back cover, adding a logo. By 2025 they held under 1% of the market they once led. “The hard part comes later, when the window closes and only the product is left — the better-engineered phone wins.”

Pei makes R&D checkable — five questions: who engineers the structure (materials, thermals, durability); who selects camera sensors and builds the imaging pipeline; who writes and maintains the OS and commits to years of updates; who re-tunes and certifies antennas when the design changes; who works directly with suppliers to co-engineer displays, chipsets, camera modules. If the honest answer is “our manufacturing partner,” the brand is picking a phone, not engineering one — a ceiling.

Ecosystems form around hard questions

Pei moved from Stockholm to Shenzhen in 2012; by 2020 Chinese suppliers were among the world’s biggest in OLED panels, camera modules, in-display fingerprint sensors. Policy alone didn’t do that — Chinese phone makers built engineering teams that refused the catalogue, and suppliers stopped being vendors and became engineering partners. Manufacturing was step one: 99% of phones sold in India are now made in India. Step two is R&D. India has the base, a market absorbing 150M+ smartphones a year, and a government aiming for a global export hub. What it lacked was the brand to drive engineering demand. The bet: 100M phones a year is the line between a brand and a platform — above it, you’re the reason the supply chain exists. “India will produce global consumer electronics companies. The only question is when.”

• Manufacturing without R&D is a ceiling — 2015’s Indian brands prove the window closes fast.

• R&D is five checkable things, not a vibe: structure, imaging pipeline, OS, antennas, supplier co-engineering.

• Ecosystems don’t follow policy; they follow engineering demand — suppliers become partners when brands refuse the catalogue.