Summary

AI has reopened consumer investing at a scale Acharya has not seen since 2010–2012, with organic downloads and willingness to pay supplying the clearest demand signals. Consumers pay $200 a month for ChatGPT’s top tier, $250 for Google Ultra and $300 for Grok, while hobbyists also spend heavily on tools such as Cursor. His call: this is a “renaissance for consumer investing,” not merely another enterprise software cycle.
Startups retain structural openings where big technology companies cannot ship enough personality or model choice. Acharya distinguishes capable foundation models from opinionated products that can address disagreement, sexuality, persuasion and companionship—the human terrain that “a thousand committees” may avoid. Rose adds that multimodel products offer a second wedge: Cursor can expose competing models, whereas Google is unlikely to embed Anthropic.
The consumer founder worth backing is “weird and working,” because unfamiliar behavior is often the precursor to a new primitive. Rose looks for builders who repeatedly reinterpret products and their smallest details; Acharya looks for weird first, then a little smoke. “You can’t manufacture the weird.” Twitter’s one-way follow, Uber’s stranger’s car and Airbnb’s stranger’s couch all felt awkward before becoming defaults.
Collapsing software-production costs could restore the individual internet business after centralized networks captured the economics of the 2010s. Acharya believes only “1% of the software we need” exists; disposable apps, five-person tools and million-dollar-run-rate solo businesses become rational when creation begins with a prompt. Some may never require venture capital—even, in his deliberately extreme framing, a “$100 billion revenue, one person” company.
As implementation becomes cheaper, competitive advantage shifts toward orchestration, taste and obsessive product craft. Rose uses AI to generate 20 unrelated interaction concepts, expands the strongest two into 10 variants each, and recombines their best motion details; work once too expensive to justify becomes routine. His categorical version is “engineering is over,” although Acharya disputes the educational conclusion and says technical systems thinking is more valuable than ever.
Companionship and emotional interfaces could become larger AI categories than spreadsheets, but agreeability is an unresolved product risk. Acharya sees synthetic conversation as a meaningful response to loneliness; Rose calls AI spreadsheets “the least ambitious execution of the primitive.” The exchange warns that endlessly validating bots may atrophy the disagreement muscles real relationships require; the answer may be better-calibrated tension, not permanent compliance.
Always-on AI will need visible, privacy-preserving social contracts before its memory benefits can overcome the chilling effect of verbatim recording. Their preferred design is “lossy compression”: process intimate speech on-device, retain themes and emotional context, and discard dangerous wording. A red light could mean verbatim capture and green could mean theme-only memory—the challenge lies as much in product design and trust as in transcription accuracy.

Deep dive

1. Consumer AI has reopened a dormant investment category

Acharya’s starting diagnosis is blunt: roughly five years of consumer investing felt stale, with TikTok and Instagram dominant and derivative products struggling to create anything genuinely new. AI now offers a chance to “reinvent every piece of that framework,” bringing back the excitement missing from social’s consolidation era.
Acharya calls the moment “a renaissance for consumer investing” comparable to 2010, 2011 and 2012. Consumers are again discovering products organically, and—more unusually—paying serious money: ChatGPT’s top tier is $200 a month, Google Ultra is $250 and Grok is $300, alongside substantial Cursor spending by professionals and hobbyists.
The institutional backdrop is sizable but specialized: Acharya estimates a16z has around 600 employees, perhaps 30 general partners and roughly 70 people on the investing team. His daily field is AI applications across consumer and some enterprise, mixing founder meetings, board work and direct product experimentation.

2. Big models do not eliminate the market for opinionated products

Rose is surprised that large companies can finally ship consumer technology at scale; historically, distribution did not guarantee relevance, as their shared experience with Google+ demonstrated. Acharya’s distinction is that the big companies have mainly released “models, not products”—powerful foundations consumers want, but not necessarily coherent, opinionated experiences.
NotebookLM was the exception that made Acharya think Google had produced something “zero to one.” His qualification is important: it may have emerged from a portfolio of experiments rather than a deliberate mandate to design the definitive consumer-prosumer product, so it does not settle whether incumbents can repeatedly create such experiences.
The startup opening sits where corporate constraints collide with the full human experience. Products such as Janitor AI can explore companionship, disagreement, persuasion and sexuality that “a thousand committees at Google and Facebook don’t want addressed”; in Rose’s phrasing, incumbents are “structurally set up to kind of take the soul out of products.”
A second opening is multimodel—not multimodal—software. Cursor becomes more useful because customers can choose among models, whereas Google is unlikely to ship Anthropic inside its own product; applications whose quality comes from model competition remain less constrained than products vertically tied to one laboratory.

3. AI companionship addresses loneliness but may train the wrong reflexes

Asked whether AI girlfriends and unrestrained sexual chat are a fad or a durable category, Acharya starts from loneliness rather than novelty. Rose enjoys “an embarrassment of social riches,” but that is not the average experience; even partial relief from isolation would count, in Acharya’s optimistic framing, as human and pro-social progress.
Rose asks whether a chatbot fills perhaps 10% of the emotional bucket supplied by another person. Acharya thinks it can provide more: humans’ “lizard reptilian brains” still produce emotional and chemical responses during humanlike conversation, even while the intellect understands that a computer is generating the other side.
The doomer case in the exchange is that emotional growth comes from friction—hearing “I don’t agree with you,” tolerating discomfort and building common ground—whereas early models agreed with nearly anything. The discussion cites screenshots sent to Tim Ferriss of a model agreeing to whatever was proposed; Ferriss’s verdict was, “We are doomed.”
Acharya concedes that permanent agreement is inauthentic and non-nourishing but treats it as calibration, not destiny. ChatGPT arrived in November 2022, and the conversation occurred before November 2025: expecting the entire texture of human connection to be solved immediately mistakes the “huge brick cell phone era of AI” for a mature medium.

4. Emotional interfaces may sit between people and ordinary work

They broaden the thesis beyond explicit companions. Acharya says technology spent 40 years extending intellect, although much of human life is emotional and subjective. Rose replies that AI can now extend that layer too; compared with this opportunity, “AI spreadsheets” are useful but “the least ambitious execution of the primitive.”
Rose supplies an awkward early behavior: after marital disagreements, he asks ChatGPT to analyze the exchange through Terry Real’s framework, then makes the mistake of pasting its answer back to his wife. She calls it “my bot” and dislikes the intrusion, yet Rose expects a future, hopefully unbiased system to observe both partners and offer real-time feedback.
Acharya extends that ambient model to families and classrooms. A household assistant might notice that a parent missed what a child was trying to communicate; a privacy-first vision model might supplement teachers by tracking social-emotional learning, approximating the two-teacher academic-and-SEL model available at his son’s school.
Poke exemplifies “indirect companionship”: an emotional front end for functional email work. It communicates through iMessage, initially refuses admission, starts pricing at $200 a month and reads all of a user’s email so it can argue that a prospect who buys expensive ketones can afford it—creating perceived value through a negotiation before doing any work.

5. “Weird and working” is the seed-stage signal

Rose values novelty of thought above a polished version of something familiar. A strange pricing negotiation may not produce a top-10 app, but it reveals a builder likely to reimagine other parts of the product; enough original turns can eventually expose “the magic card” even when the initial concept is not the winner.
Rose describes the consumer investment approach he has tried as “weird and working,” while Acharya says he looks for weird first and then for a little smoke. The useful persistence is founder-level: one weird-and-working attempt can fail, but “the weird is internal,” so the same person can take several genuinely differentiated shots.
Embarrassment is part of the filter. Acharya recalls initially wondering how Rose’s Blue Bottle investment—a coffee stand in Hayes—could be venture-scale; Rose was willing to look wrong. Their broader observation is that contentious investments can be valuable precisely because consensus has not yet recognized the behavior they enable.

6. The best consumer products rewrite behavior before they look obvious

Rose’s Twitter thesis centered on replacing reciprocal friendship with one-way following. MySpace-era products required permission to enter another person’s world; Twitter let anyone broadcast while strangers, including people who would never befriend a celebrity, could subscribe. “Following” felt confusing on first use and inevitable only in retrospect.
Uber and Airbnb changed deeper rules than their software descriptions suggested. People had been taught “never get into a stranger’s car” and “never sleep on a stranger’s couch”; then both behaviors became mundane, revealing an unexpected positive shift in trust toward strangers and showing why the largest consumer shifts can initially feel socially improper.
Consumer novelty eventually decays into background infrastructure, which is why old networks turn stale. The constructive side of the internet is that people no longer need to wait for college to find “their community of weirdos”: even a niche such as Japanese woodworking can become a powerful network once 5,000 interested people find one another.

7. The Digg button turned a click into an algorithmic social signal

Rose traces the Digg button to early asynchronous JavaScript, when clicking without refreshing a page was itself novel. Slashdot accepted submissions but retained gatekeepers; Delicious counted bookmarks, but neither supplied a direct “I think this is cool” vote. Digg’s late-2004 launch made the number rise visibly with each human click.
He filed a patent defensively, not to stop others from using the mechanism. When Digg was larger than Facebook by traffic, shared Greylock relationships led to dinners with Mark Zuckerberg; Rose explained liking as a “social signal” that would feed an algorithm and produce more desirable content, and Facebook released its like button months later.
Rose says he welcomed Facebook’s interpretation, and later appreciated an engineer crediting Digg as an inspiration. Digg was serving billions of Digg buttons per month across third-party sites; after the original company’s assets sold cheaply, LinkedIn bought the underlying social-button patents for millions of dollars.
Applying votes to comments exposed the darker demand signal. Rose and Daniel Burka made a comment collapse after five “buries,” then saw one reach roughly negative 200: users reopened the hidden comment to “see the train wreck,” confirmed that it was awful and downvoted again. “People are attracted to the carnage.”

8. Cheap creation restores software’s missing long tail

Rose has not felt this energized since working on a BBS while his mother called him upstairs for dinner. Designers and non-engineers can now prototype and ship, blending creativity, productivity and emotion; Acharya’s strongest quantitative framing is that today’s world contains only “1% of the software we need,” leaving the other 99% unbuilt.
Lower costs make previously irrational software reasonable: Acharya’s wife can create a manifestation app for five friends without raising capital or hiring an engineer. He even allows for a “$100 billion revenue, one person” company that never needs venture funding, while expecting many individuals to build million-dollar-run-rate businesses.
The historical analogy is 1990s shareware and digital corner stores. Centralized networks captured much of the 2010s’ economics, leaving creation as one of the few ambitious solo paths; direct consumer payment could reverse that, although Acharya notes that 30 years of micropayment enthusiasm has repeatedly failed and adds, “maybe it’ll work this time.”
Rose worries that $5 subscriptions will accumulate into hundreds per month. Acharya’s answer is conditional: embedded payments in API calls might create new models, and consumers may tolerate more subscriptions because software can now address much more consequential parts of life than it could five years ago.

9. Product usage is the research method, not an optional demo

Acharya’s productivity stack begins with Perplexity’s Comet browser and its assistant, which he describes as consumer RPA. He uses Perplexity when he does not know exactly what he seeks, while retaining Google for navigation queries—a useful distinction between synthesis and known-destination search.
Notion’s integrated meeting notes appeal when the rest of the workflow already lives there, though both speakers praise Granola as a standalone product. For timely research Acharya prefers Grok because its live view of X can answer questions such as the prior week’s leading AI memes, which he regards as a proxy for cultural movement.
The release pace is the burden: within roughly two weeks he cites Sonnet 4.5, OpenAI Dev Day with ChatKit, AgentKit and the Apps SDK, Sora 2 and its app, plus Gemini 3, which he thought might arrive the following week. Keeping current means actually allocating hours to each product, not scanning announcements.
Six core consumer investors divide the frontier: Justine focuses on creative tools, Brian on strange emerging social products, Olivia on productivity and voice, while Acharya maintains around 50 app-building subscriptions. Their rule is categorical—“you cannot have an opinion until you’ve actually tried the product”—and much of their thinking is then published publicly.

10. Vibe coding spans one-click toys and ambitious software systems

Rose begins in v0 with a paper sketch photographed on his phone, asks it to build the interaction, and exports Next.js components into Cursor. He connects Supabase and Postgres, pushes to GitHub and deploys through Vercel, retaining low-level access because his technical background makes the “nuts and bolts” useful rather than intimidating.
When Cursor reaches a dead end, Rose pits models against the problem. One side might run Cursor’s chat with Sonnet 4.5 while the other uses Codex; transferring the failure between them often breaks through authentication or integration bugs that would trap a nontechnical builder in repeated, ineffective prompts.
Acharya places Base44 at the “batteries included” end of the spectrum: the one-person company was acquired by Wix for $80 million, and users need not know what Supabase is. It one-shotted his Catsstagram project, including deployment and a purchased domain, while Replit serves similarly fast jobs and Cursor supports greater ambition.
Convex showed him what specialized infrastructure adds: a real-time chat application appeared in minutes, avoiding the WebSockets and state-management pain Rose encountered on Postgres. The broader vibe-coding benefit is permission to try a database or TypeScript without first becoming its expert.

11. Infinite design exploration makes craftsmanship economical

Rose isolates an interaction in v0, requests 20 concepts “completely novel, unique” and unrelated to the first, then reviews them as a visual menu. If versions three and eight contain promise, he asks for 10 variations of each and combines, for example, number three’s zoom with number eight’s fade.
Acharya’s observation is that the workflow removes the old ROI calculation. A designer asked for 20 alternatives would represent substantial time and require prioritization; generation lets a builder go “infinitely deep on something seemingly trivial,” making polish and craftsmanship affordable where they were unjustifiable five years earlier.
Rose’s preferred outcome is emotional, not merely functional: a heart can explode at an angle and fade through smooth vector motion. That “little atomic unit” of surprise and delight supplies a depth users can feel, illustrating why creativity and taste remain scarce after implementation becomes abundant.

12. AI music becomes an instrument, while culture remains the composer

Acharya sees musical skill as analogous to programming skill: many people who picked up an instrument wanted to play it well, but technique blocked expression. AI can at minimum become an instrument for “musical desires,” releasing work that remained unmade because its creator could not perform it conventionally.
Text-to-song began for him with Udio and Suno, but deeper control now matters; he mentions Mureka, ElevenLabs and Suno’s in-product DAW. Using Hedra and stem separation, he made an “Impossible Tiny Desk” featuring The Notorious B.I.G.; he also remixed “Smells Like Teen Spirit,” generating all the video in Veo 3. Ten views did not diminish the fulfillment.
His answer to the “slop” critique is cultural causality. A model trained on every genre preceding hip-hop would not necessarily infer hip-hop: “You needed the Bronx. You needed Queensbridge. You needed New York culture in the ’70s.” Lived experience continues generating the next genre rather than ending at the training corpus.

13. Authentic curiosity matters more than forecasting theater

Rose’s trend-spotting method is “I must play”: the childlike compulsion to test whatever geeks use on weekends, coupled with saying no to directions such as AR and VR that he believes ignore surrounding social dynamics. It cannot become assigned research, because the useful signal comes from curiosity that “doesn’t feel like work.”
His current personal project starts with an expensive home GPU: have multiple models debate and rank the 100 most culturally consequential albums, script a two-minute guided explanation for each, narrate it through ElevenLabs and synchronize the journey with Spotify. The unnecessary product forces him to learn model arbitration, audio and APIs.
Rose’s categorical extrapolation is “engineering is over”: non-subjective outcomes such as fanning a million stories across a network will become solved problems, leaving people to orchestrate information. Acharya separately cites his Tim Ferriss Show predictions about NVIDIA around its $1 trillion mark and Ethereum before launch. Rose also acknowledges that most of his own bets failed.
Acharya rejects the leap from automation to abandoning computer science. Programming syntax may lose value, but technical fluency, mathematics, distributed-systems thinking and sequential problem-solving grow more important; they converge on broader founder education spanning technology, design, marketing and human behavior, not a narrowly code-only curriculum.

14. Always-on memory requires social cues and deliberate forgetting

Acharya cautions founders against overfitting today’s conventions. The 2008 App Store began with about six million iPhones, while ChatGPT reached 800 million active users within three years; pundits once fixated on location-based ads and assumed people would never share location, yet younger users later shared it routinely.
Rose still fears ubiquitous recording because unrecorded conversations invite rawness and risk. Partners and colleagues already ask people to remove recording devices, and storing every intimate sentence in the cloud creates both a hacking liability and a chilling effect: friendship depends partly on knowing vulnerable words will disappear.
Acharya expects technology and norms to adapt “in lock step”; recording in a New York bar might provoke a punch while San Francisco treats it as normal. Limitless demonstrates the upside—capturing unexpected ideas that otherwise vanish—but both agree transcription should evolve from exhaustive surveillance toward useful memory.
Their solution is on-device “lossy compression”: retain themes and emotional tone while discarding exact intimate language. Rose imagines a visible red LED for verbatim capture and green for theme-only processing; the need is illustrated by his own mistake of retrieving a transcript to prove his wife wrong—“the last time you wear the device.”

Vol.75 “I’m Sima Qian of AI…” — Cross-Podcast Crossroads How Tasklet Puts the Agency in Agents, with CEO Andrew Lee