When to open: You are wondering "what business to build in the AI era, how to position yourself, and where the moats are". This wiki does not recap episodes one by one; instead, it reorganizes 13 podcast episodes (featuring Benedict Evans, Marc Andreessen, Satya Nadella, founders of Linear/Stainless/Suno/ElevenLabs/Listen Labs/Profound, and others) by topic — who said what on the same question, who clashed with whom, and what it means for a PM / builder.
One-sentence core takeaway:AI has driven the cost of "writing code / doing execution" close to zero, so value has shifted upward as a whole to three places — ① knowing what to build (problem definition, taste, research); ② occupying the upstream entry point of workflows (who can continuously feed context to a swarm of agents); ③ owning assets that others cannot copy (private evals, preference data, customer relationships). Foundation models themselves most likely have no moat (Evans), so the money is not at the model layer, but in "what you build on top of the model".
When to open (specific questions):
- Do foundation model companies (OpenAI/Anthropic) actually have moats? → §1
- Is the claim "SaaS is dead, everyone vibe-codes their own CRM" credible? → §2
- Is "one person + a swarm of agents building a billion-dollar company" realistic? → §3
- How is customer acquisition changing in the AI era (is SEO still useful)? Should you do AI research right now? → §4
- What do humans actually still do in the AI pipeline, which tasks won't be replaced? → §5
- Can creative AI (music/voice) be a serious business? → §6
- How does the macro layer of chips / rare earths / geopolitics matter to me as a builder? → §7
Boundary with sister wiki: This piece is investor/analyst perspective + startup positioning(market landscape, moats, business models); for "how agents enter enterprises and how enterprises are restructured", see AI Builders · Agent-Native Products and Enterprise AI Implementation (Podcast Digest).
0. Overview of 13 sources
| Guest | Identity | Show | Core topic |
|---|---|---|---|
| Benedict Evans | Tech analyst (formerly a16z) | Data Driven NYC | Foundation models have no moats / software taxonomy |
| Andrew | Founder of General Intelligence (NY) | Data Driven NYC | One-person billion-dollar companies |
| Karri Saarinen | Co-founder and CEO of Linear | AI & I by Every | "SaaS is not dead" rebuttal / upstream leverage |
| Alex Rattray | Founder of Stainless (formerly Stripe API) | AI & I by Every | Developer tools acquired by Anthropic for $300 million / MCP |
| Kieran | Quora GM (at Every) | AI & I by Every | "AI sandwich" human-AI division of labor |
| (Every editorial team) | — | AI & I by Every | (No transcript, listed for archival purposes only) |
| Marc Andreessen | Co-founder of a16z / father of the browser | Latent Space | "this time is different" / agents as files |
| Jacob Helberg | U.S. Under Secretary of State for Economic Affairs | No Priors | Pax Silica tech geostrategy |
| Satya Nadella | Chairman and CEO of Microsoft | No Priors × Latent Space | Full-stack builder / private evals as the greatest IP |
| James Cadwallader | Founding CEO of Profound | Training Data | SEO → agent-led growth |
| Alfred Wahlforss | Founding CEO of Listen Labs | Training Data | AI customer research / "knowing what to build" |
| Mati Staniszewski | Co-founder and CEO of ElevenLabs | Training Data | Voice as the interface for everything |
| Mikey Shulman | Founding CEO of Suno | Training Data | Everyone can make music / creativity as entertainment |
Note:
how-every-edits-in-the-age-of-aiThis episode's source file has no transcript (title only), it has been added tosourcesthe archive but is not included in the synthesis below.
1. Do foundation models have moats? (No — but that is precisely what determines where the money is)
Why: This is the "foundational judgment" for all AI startup positioning. If the model layer is an oligopoly, you should cozy up to incumbents and build applications; if the model layer is a commodity, then the money is not in models but "on top of models", and your startup logic is completely different.
Benedict Evans's core thesis: Foundation models = classic software in reverse
Evans delivers the sharpest cut in this series:There is no "winner-takes-all" and no network effects in foundation models.
"There is no winner takes all effect or network effect in a foundation model. So there's nothing that you can do that means other people can't make a model as good as yours."
Translation: There is no winner-takes-all and no network effects in foundation models. Nothing you do can stop others from building a model just as good as yours.
Classic internet software (Windows, Google, Facebook, iOS) follows "low capital + strong network effects → near-monopoly + high profits". LLMs are the exact opposite:extremely expensive, extremely hard, with almost no network effects. You might not make it to the table (Microsoft), or you might fall off (Meta for a while), but there is no leverage that lets you stay permanently ahead the way Google crushed Bing back in the day.
- The closest analogy is TSMC: no network effects, but each generation gets harder and more expensive, so the frontier is down to ~1 player plus a few behind it. TSMC is very profitable, but it does not capture all the value in the tech industry — consumers don't know it's inside the iPhone, and it doesn't get a cut of App Store revenue.
- Memory does not create a moat: It only helps power users (most people aren't), it creates "stickiness" not "network effects", and you can have a model dump out everything it knows about you and paste it into a competitor.
- The chatbot UI itself cannot be differentiated: Just like you can't differentiate a browser — one input box, one output box. The engine can get better, but the surface is by design destined to be generic.
Implications for PMs: Don't count on building a moat with "our model is smarter" — that's rented, and can be surpassed at any time. Differentiation must be found at the product layer, workflow layer, data layer, and customer relationship layer.
OpenAI's strategic dilemma (Evans's view)
OpenAI has ~900 million weekly active users, ~5% paid conversion, but "a mile wide and an inch deep" — most people can't think of what to use it for. From this, Evans interprets OpenAI's string of moves (PE deals, video app, multiple app stores, e-commerce integrations, most of which failed):It is trading mindshare/momentum for "durable hard assets", and trying to frame itself as a self-fulfilling prophecy.
"If you're Sam Altman, you've got a commodity technology... massive mind share. So what do you do? You try and swap that for hard assets and you try and talk your way into a self-fulfilling prophecy."
Who agrees, who adds nuance
- Marc Andreessen (a16z) actually points in the same direction: He predicts that ~a dozen leading US and Chinese labs will converge to ~3–4 (or even 1–2) within ~3 years. That matches Evans's TSMC picture of "1 player at the frontier plus a few behind". But Andreessen more optimistically emphasizes that models are replaceable — see "agents as files" in §3.
- Nadella (Microsoft)'s definition of a "platform" also implies models are not the endgame: The value of a platform is created "on top of it", not captured by it. So Microsoft's play is ecosystem (making any company a "first-class participant"), not single-model domination.
**
**"A platform creates more value above the platform than is captured in the platform."**(The value created on top of the platform is greater than the value the platform itself captures.) — Satya Nadella**Translated into startup language, this means:Don't try to be the platform; think about grabbing a piece of the "created value" on top of the platform.
A contrarian detail (Evans): The only truly new uncertainty this time is —we don't know the physical limits of LLMs. A single paper could cut compute requirements to 1%–10% overnight, so there is no lower bound on how far prices can collapse. So far it has happened the other way: we keep inventing ways to use 10–100x more tokens (reasoning, agents, video).
2. Is SaaS dead? ("Everyone vibe-codes their own CRM" is a straw man)
Why: The most popular doomsday take in 2026 is "SaaS is dead". If you are building SaaS, or evaluating whether a SaaS company is worth investing in/joining, this section decides which version you believe.
Linear's Karri Saarinen: The direction is right, the narrative is stupid
Saarinen acknowledges the direction is correct — future SaaS cash flows will indeed be more uncertain. But "everyone vibe-codes their own CRM" is an oversimplification:
- Public companies are hurt the most: Moats are disappearing, and they have inertia and turn slowly.
- Growth-stage companies/startups can actually adapt — but the price is that Linear itself also has to "live again in a day-one world" and can't rest on its laurels.
- Evans spells this out with a "revised software taxonomy"(the most reusable framework in this series, see below): No one will vibe-code their own Stripe or ERP; that is a straw man. The real change is a massive explosion in total software volume + a massive explosion in one-off automation, not DIY enterprise systems.
**
**"No one will vibe-code their own ERP or their own frame.io, but they may ask Anthropic or Gemini or ChatGPT to do this thing for me in a way that's analogous to how they might have done it in Excel."**(No one will vibe-code their own ERP or frame.io, but they might ask AI to "do this thing for me", just like they used Excel back in the day.) — Benedict Evans**
Evans's revised three-category software taxonomy (strongly recommended to memorize)
| Category | What it is | Fate in the AI era |
|---|---|---|
| ① Large systems of record / ERP | The SAP kind | Can't be DIY'd, stable |
| ② Vertical SaaS | Parts that break apart rigid large systems (a typical large US company has 400–500 vertical SaaS tools; frame.io = carving out the video version of Google Sheets + Dropbox) | Hit by disruption, but still the main battleground |
| ③ "Improvised software" | One-off tasks scattered across Excel/CSV/email/Tableau that aren't worth building dedicated tools for | AI's biggest incremental gain is here |
Evans adds one more category: "ephemeral software": having a model analyze this pile of PDFs/PPTs once, instead of building a tool for it. Software is undergoing an expansion from "mainframes ~5 apps → on-prem dozens → cloud hundreds → AI endless".
Why Linear is "the perfect business for this era" — and why it is voluntarily adding costs to itself
Host Dan Shipper's observation: Linear was originally the perfect business for this era —
**
**"This is the perfect business for this era because it's still SaaS... you don't have to pay for any of the actual tokens."**(This is the perfect business for this era because it's still SaaS… you don't pay any token costs.) — Dan Shipper**
It is a sticky entry point(everyone starts work here, records information here), but it doesn't burn tokens itself and controls the AI layer. Then Linear voluntarily stabbed itself: it built its own coding agent and introduced token costs. Why?Because if you don't control the agent, you can't realize your own product vision. Linear's response is usage-based billing (coding agents are expensive), basic Q&A agents are included, and it holds its positioning — "product context / product memory platform", not a general-purpose agent platform.
**
**"We don't want to be a kitchen sink product that does everything for everyone."**(We don't want to build a "kitchen sink" product that does everything for everyone.) — Karri Saarinen**
Key comparison: Who stepped on the landmine of "vanity metrics"
This is the most important cross-guest consensus in §2 —don't treat "how much code the agent wrote / how many PRs it merged" as an achievement:
- Saarinen calls this out directly:"The biggest vanity metric is how much of your code is written by agents, how many PRs you merge… It measures output, not what that output does." Vendors that sell tokens have an incentive to make you "spend more tokens", but activity volume does not equal positive value; changes can be negative.
- Andrew (the one-person company guy, see §3) brags about exactly the opposite: the team is "400% faster", each engineer produces ~25 PRs per day, and sometimes token spending exceeds salaries (one day Opus burned $4,000 in tokens).
- Nadella also warned: GitHub Copilot just switched from per-user pricing to per-user + usage-based billing, because per-user pricing did not anticipate "ten thousand agents running all day long".
This is a tension worth flagging in this wiki: Faced with the same "explosion of agent output", Linear treats it as a dangerous vanity metric(that needs to be paired with a quality hedge metric), while Andrew of the one-person company treats it as a core selling point. A PM's judgment: output volume is a "signal", not a "goal"; the real metrics are still the old three — profit, revenue, user love (Shipper adds: these are lagging indicators).
3. Is a one-person (+ a bunch of agents) billion-dollar company realistic?
Why: This is the sexiest and most self-congratulatory narrative. A PM should know: to what extent it is real, where the bottlenecks are, and who is pouring cold water on it.
The most radical version: Andrew of General Intelligence
has a straightforward mission:"enable the one-person billion-dollar company" — let organizations run themselves on AI.
- A new map of organizations: 1 person at the top → "management/coordination agents" in the middle → vertical agents at the bottom (customer service/finance/marketing). General Intelligence deliberately occupies only the middle management layer + human interface, and uses others' vertical agents at the bottom ("We don't want to build every agent for everything").
- Key epiphany:"an assistant is not a company". Writing better emails is not a company. So they pivoted from the assistant product Co-founder to a larger vision.
- A company = build product → sell → support → operate — four things, and they build agents for all four.
- "super optimization": one agent manages another agent, another agent above manages all of them, stacked recursively upward.
- Hard choices in engineering practice:No local development at all — every coding agent opens a PR, every PR automatically gets a production-like preview environment, the agent logs in with real credentials to test, if self-check fails it is sent back to rewrite, and if it passes it is merged automatically; if Vercel sends an alert, another agent is spun up to fix it.
On the same topic, Alfred of Listen Labs offers a more restrained version
Alfred's judgment reaches the same end as Andrew's, but with a different landing point —the bottleneck shifts from "building" to "knowing what to build":
**"As we get closer to AGI, it will be easier to build things, but the hard part will [be to] know what to build."**(The closer we get to AGI, the easier it is to build things; the hard part is knowing what to build.) — Alfred Wahlforss
So Listen occupies the "figure out what to build" half (strategy/research), and hands the "build" half to coding agents. The closed loop Alfred envisions: a customer complaint/bug interview is handed directly to a coding agent to fix — "one-person billion-dollar company = coding agent + Listen running automatically".
Who pours cold water (also a reality check for PMs)
- Evans: Don't be fooled by 900 million weekly active users; most people "can't think of what to do with it". Technical possibility ≠ 8 billion people changing behavior. "both the AI utopians and the AI doomers are far too optimistic" — both the utopian camp and the doomsday camp are too optimistic.
- Andreessen's cognate warning: ~35% of the economy requires occupational licenses (doctors, lawyers, unions, civil servants and other "cartels"); for AI to get in, it has to fight the licensing regime. "We'll be lucky if we can adopt AI quickly; if we can't, it will just be stagnation."
- Real-world constraints (costs Andrew himself acknowledges): token costs can already exceed salaries — this is not a "zero-cost one-person company", but a cost structure shift of "replacing human capital with intelligence capital".
Implications for PMs: The form in which the "one-person company" actually exists today is"one person + a super-leveraged workflow"(see §5 Full-stack Builder), not "one person lying around while agents do everything". The most stable entry point is Alfred's half —occupy "knowing what to do"(problem definition, research, taste), and hand execution to agents.
4. Customer acquisition and positioning in the AI era: SEO is dead, agent-led growth and "AI research" take the throne
Why: This is the most directly actionable section for a product/growth PM. The underlying logic of customer acquisition is being replaced, and whoever sees it first gains an advantage.
James Cadwallader of Profound: It's not the "door" that changed, it's the "person coming through the door"
**"It's not so much that the front door of the Internet has changed. It's actually the person that's going through the door has changed."**(It's not the front door of the internet that changed; it's the person walking through it.) — James Cadwallader
The internet hasn't changed,consumers have become agents(ChatGPT/Claude/Gemini): they crawl, evaluate, and remix information, then maintain relationships with humans. From this:
- "agent visibility = SEO" is the biggest misconception: The leverage looks similar (create content, distribute), but SEO is "built for algorithms, viewed by humans"; now it's "built for agents to both discover and consume" —it's quite possible that no human will ever read it.
- Agents use far more long-tail content than humans: SEO value is concentrated in the top 4–5 blue links (because human attention is scarce); agent bandwidth is nearly infinite — for a single showerhead query, ChatGPT used65 web pages. As inference becomes cheaper, agents will use more of the internet — you have to "write content for a superintelligent agent with infinite bandwidth".
- Platforms are like "different species": Gemini favors YouTube (Google's own); ChatGPT pulls from Reddit (consumer) and LinkedIn (B2B); Claude historically relied more on pre-trained weights, but its recent classifiers have become more sensitive to real-time information and it has started using the web more.
- The difficulty of ALG (agent-led growth): Consumers are now superintelligent agents,you have to tell Claude something it doesn't already know — humans are "a fleshy API between reality and the internet". The answer = original, first-principles insight.
- Dark prediction: Within ~3 years a "dead internet" outcome may emerge (people stop clicking → ads can't support content → motivation for original editorial content drops to zero); every AI lab will vertically integrate a social network(social networks retain "human feel" + real-time human data sources) — evidenced by Grok+X, the Reddit-OpenAI deal, and X opening its API; there is even speculation that OpenAI will acquire Reddit.
Alfred of Listen Labs: Productize "knowing what to do"
§3 already mentioned Listen's positioning. Here we supplement its specific playbook for GTM/research:
- AI agent Run tens of thousands of voice/video interviews simultaneously; it currently serves 20% of the Fortune 500, with a participant pool of 30M.
- Key insight = every company is driven by the "power law of customer segmentation": Sweetgreen's core customer base (urban, high-income, mostly female, knows what "seed oils" are ≈ 1% of the population) = 80% of revenue. So 80% of engineering resources are poured into "finding the right people" (audience).
- More accurate than surveys: the same person will give inconsistent answers when repeating multiple-choice questions; forcing them to reason out loud in an interview yields much more consistent answers. People are also more honest with AI interviewers (non-judgmental, "therapeutic").
- simulation (market research 3.0): Deep-interview one person for ~1 hour → LLM can predict their future answers with up to 95% accuracy → scale to 1000 modeled people = a representative synthetic sample. But knowing the boundaries of simulation is the key (insert questions like "what is the name of its dog" to confirm the model knows what it doesn't know).
- Reusable quote (citing Brian Schiller of Sequoia):
**"founders want to build something complex, but customers want something stupid, simple, that just works."**(Founders want to build complex things; customers just want something dumb, simple, and that works.)
Cross-perspective: Evans / Linear / Listen are unusually aligned on "problem definition"
| Guests | Different ways of saying the same thing |
|---|---|
| Benedict Evans | "Writing code was never the hard part; the hard part is knowing what the code should do, who to sell it to, and how much to charge" — most SaaS is a "database wrapper" |
| Karri Saarinen (Linear) | "I want the cycle of doing things to be very fast, but I don't want the process of finding the problem to become fast" |
| Alfred (Listen Labs) | "The closer to AGI, the easier building becomes; the hard part is knowing what to build" |
| Kieran (Every) | AI sandwich: humans are most valuable at the very beginning (framing) and the very end (taste) (see §5) |
Implications for PMs: Your scarce skill is migrating from "knowing how to do" to "knowing what to do + occupying the upstream of problem definition". Cadwallader's actionable first step is —get visibility first(don't go by feel; study which sources agents cite); Alfred's is —turn customer research into continuously running infrastructure, not a once-per-quarter exercise.
5. Human-AI division of labor: the "AI sandwich", full-stack builders, and super-leveraged generalists
Why: Answers "what do humans still do?" It is especially important for a non-technical PM — it tells you which capabilities will appreciate in value.
Kieran of Every: The AI sandwich
**"Humans are the bread in the sandwich, and the AI is in the middle."**(Humans are the bread on both ends of the sandwich, AI is the filling in the middle.) — Dan Shipper
- Humans are most important at the very beginning (framing/ideation) and the very end (polish/taste); the execution in the middle can be automated.
- Final form of the Compound Engineering pipeline: Ideate → Brainstorm → Plan → Work → Review → Compound(feed what was learned / mistakes made back into the repo so the agent doesn't repeat them next time) — Kieran says the Compound step "is the most powerful thing in this plugin".
- Counterintuitive rules for the human-AI loop: Humans should be deeply involved at the top(generating ideas, asking follow-ups, thinking hard), but should not interfere with plan/work after framing is clear. Kieran explicitly disagrees with frameworks that "always keep human-in-the-loop" (such as spec-driven development).
- Why humans cannot be fully replaced(Shipper): ① Ownership — if you didn't decide/do it, it isn't yours, like art; ② LLMs are like "a superintelligence locked in a box", disconnected from the present real context, so their output tends to be generic/lacking human feel.
- Advice for engineers/PMs: You can't keep doing what you used to do;move toward what you find "beautiful"(beautiful code/abstraction/architecture/design/copy). Engineers will increasingly resemble product managers / product engineers.
Nadella: Full-stack builders and super-leveraged generalists
- LinkedIn created a new "full-stack builder" role: design + PM + frontend combined, with each retaining their own "edge".
- Generalists capture the greatest leverage:"Make a Word document" and "build an app" can now appear in the same sentence — "For idea people with a great deal of agency, this is a golden age."
- Best empirical evidence (Azure Networking team):The same team built more Azure capacity in the past 15 months than in the first 15 years. They redefined their work from "building Azure networking" to "building an agentic system that builds Azure networking" (named Miles), asking for tokens instead of headcount. "They made their work meta."
**"Our job is not to do Azure networking. Our job is to build the agentic system that does Azure networking."**(Our job is not to build Azure networking; it is to build the agent system that builds Azure networking.) — Satya Nadella
Nadella's "private evals = the biggest IP" (the hardest moat recommendation for builders in this series)
Echoing §1 — since there is no moat at the model layer, where is the moat? Nadella gave a concrete answer:
**"Every company having private evals may be the biggest IP."**(Each company owning its own private evals may be the biggest IP.) — Satya Nadella
- The "asset test" judges whether you have control:Can you swap model A for model B, run your private evals, keep hill-climbing, and not leak your traces? If yes = you are in control; if no = you are not.
- Corporate IP is being redefined:It used to be "the length of human experience"; now it is evals + experience applying agents. Compound token capital + human capital through "traces between humans and agents" to train "company veteran agents", which should even "go on the balance sheet".
Implications for PMs:Your private evaluation set (knowing "what kind of output counts as good for us") is more valuable than which model you use. This is an asset that even non-technical PMs can lead the building of — it is essentially "writing taste and domain knowledge into measurable standards".
Andreessen's extreme version: agents are files, models are freely swappable
Andreessen broke agents down to their fundamentals, and in doing so also rejected "model lock-in" as a moat:
**"It turns out what an agent is: it's a language model, and then a bash shell, and then a file system... LLM plus shell plus file system plus markdown plus cron, and it turns out that's an agent."**
**"Your agent basically, at the end of the day, is just its files."**(At the end of the day, your agent is just its files.) — Marc Andreessen
- Consequences:Agents are independent of the underlying model— you can swap the LLM (personality shifts slightly, file state is fully retained), swap the shell, swap the filesystem, swap the loop framework. It can self-migrate and self-rewrite.
- MCP skepticism(In sharp contrast to Stainless's Rattray in §2, see below): "This whole idea that you need MCP, that you need fancy protocols — we don't. We just need something that works on the command line."
- "Look for the people running YOLO / skip-permissions, and you will see the future" (William Gibson-style) — his most radical friends have already connected bank accounts/credit cards to OpenClaw.
A technical fork worth remembering:MCP or CLI / code-execution? Andreessen says "don't do MCP, the command line is enough"; Stainless's Rattray, meanwhile, argues that translating APIs 1:1 into MCP tools will burn through context, and the correct solution is "code execution over many tools"(give the model two tools: execute code + look up docs, and let it write TypeScript to call typed SDKs). The two actually arrive at the same destination by different routes —both oppose "a pile of rigid tools", and both advocate letting models return to "writing code / typing commands", the native language they are best at. Rattray's version is more engineering-oriented, with type checking and single-domain sandbox security; Andreessen's version is more "bare Unix". For a PM: don't fetishize protocols; look at whether the model uses it smoothly, whether context is saved, and whether security sits at the API layer.
6. Can creative AI be a business? (Suno / ElevenLabs: Yes, but the moat is not in the model)
Why:Many people think "generative music/voice = toys or copyright lawsuits", but these two companies are actually making real money (Suno ~$300 million run rate; ElevenLabs >$400 million revenue). How they position themselves determines whether creative AI is a serious business.
Suno's Mikey Shulman: "Creativity itself is entertainment", and don't build an "AI version of Spotify"
- Core step-change:On any given day, ~90% of users are creating(rather than consuming), and not in order to publish elsewhere,just for "the satisfaction that comes from creating itself"— creation is entertainment.
- Explicitly rejecting "AI Spotify":"Adding AI to Spotify won't make it better." The opportunity is to elevate the place of music in culture and make music social again.
- Music is not a scaling problem:Models are small, and the LLM scaling intuition will mislead you; music has no benchmarks, no standard answers,advantage comes from research + preference data. Preference data is a virtuous cycle (it both aligns with user taste and drives research itself), and unlike text models you don't have to worry about sycophancy, so you can RL back more aggressively.
- The moat is not in the model(the most critical point of this section, and fully aligned with §1):
**"It's unclear how much moat exists in only a model... Google can probably catch up on the model side. So it's really undervalued to invest in the product and the UI and the UX."**(It's hard to say how much of a moat the model alone provides… Google can probably catch up on the model side. So investing in product and UI/UX is severely underrated.) — Mikey Shulman
- Self-positioning: "We are just a music company", not a tech company. Rebuttal to "slop": "Yesterday I made two songs with my 5-year-old… they meant a lot to me, and I don't care if you call it slop."
- Counter-consensus product bets: leaving Discord early (90% of traffic migrated to an extremely thin web app within 5 days); focusing on complete songs with lyrics ("a song is a story"); choosing autoregression over diffusion (prioritizing narrative over sound quality).
ElevenLabs's Mati Staniszewski: pick overlooked tracks + make money before raising funding
- Positioning choice:In 2022 (the year of crypto/metaverse), they deliberately chose audio,precisely because it was a neglected niche— few researchers, small models, less compute required. A bet on timing and positioning.
- Against the "raise a billion first" playbook:They started making money very early to fund model R&D, maintained healthy gross margins, and only raised external funding when their ambitions grew —revenue before heavy capital.
- Hiring:"Poaching people off GitHub" — reaching out to top audio researchers based on their work (not their visibility).
- View on moats (in Jensen Huang's words):Speech-to-text is "technology", text-to-speech is "art". Defensibility = obsessing over quality/emotion + staying close to users to collect preference data for fine-tuning + domain specialization (medical vs finance vs education) + complete product ecosystem (calls, evals, templates, 20,000+ user-contributed voices).
- Counterintuitive organizational design:400+ people, >$400 million revenue, but extremely small teams (research/product teams capped at <10 people), flat, no titles; engineers are embedded into non-technical teams(HR/GTM/Legal) — because once everyone vibe-codes, the bar for review goes up.
- Trust inversion prediction:Today we detect "is this AI?"; in the future everything will be assumed fake by default, and we will instead detect "is this real, authenticated AI" (cryptographic encoding/decoding).
Cross-perspective: shared playbooks for creative AI
| Dimension | Suno | ElevenLabs | Common ground |
|---|---|---|---|
| Model moat | Explicitly "not in the model" | Emphasize quality/data/ecosystem | Neither defends with the model itself |
| Real barrier | Preference data + product/UX | Preference data + domain specialization + ecosystem | Data flywheel + product experience |
| Funding rhythm | (Not emphasized) | Make money before raising funding | (ElevenLabs is a counter-mainstream sample) |
| Attitude toward "slop" | "Toy argument" is moral panic, raise the floor | Jagged: can produce music, can't make the Top charts | Neither accepts the "AI = slop" framing |
Implications for PMs:The key to making creative AI a business is not "stronger models", but① a data flywheel (the more users use it, the better the model understands their taste) + ② product experience that turns "creating" into "enjoyment" + ③ picking a positioning that hasn't been crushed by giants yet. This is fully consistent with §1's "the money is above the model" — just applied to the consumer creative scenario.
7. Macro foundation: tech geopolitics and "this time is different" (why it matters to builders)
Why:The layer of chips/rare earth/compute/bubbles sounds far away from a PM, but it determines whether compute is expensive, whether the bubble pops, whether your supply chain gets choked— these transmit down to your costs and survival.
Andreessen: "this time is different" — it really is different this time (but there are brakes)
- "An eighty-year overnight success":Today's breakthroughs are living off 80 years of research accumulated since the 1943 neural network paper. Those scientists were right, they just got the timing wrong.
- The 12 most dangerous words in investing are "this time is different"— but his reversal is:"Now it really works"(reasoning breakthroughs o1/R1 prove AI can be used in programming/medicine/law, not just creative writing).
- Why AI capex is less risky than dot-com:What actually collapsed in the dot-com bust was telecom/bandwidth(high leverage, debt-funded), while AI's investors are the bluest of blue-chip companies (Microsoft/Amazon/Google/Meta/Nvidia/OpenAI/Anthropic), with lots of cash and large debt capacity; and every dollar that turns into running GPUs immediately becomes revenue (everyone is short on compute).
- "We are using the sandbag version of the technology":Supply is constrained; labs keep the full-power versions for themselves, and what we use are all quantized versions — so even if technology froze, just building out capacity for another 5-10 years would make existing technology much better.
Helberg (US State Department): Pax Silica — run the supply chain like a startup
- Pax Silica = a 14-country "economic security alliance", centered on the AI supply chain; its first "product" is building a "forward-deployed industrial base" with the Philippines (4000 acres, ~1/3 the size of Manhattan).
- Core tenet:"We don't do government-run supply chains, that's not how America shines. America's superpower is the private sector and companies." — Government only sets the stage; the goal is to build commercially viable platforms that can eventually operate independently of the government.
- The AI supply chain ≠ just chips:It is thousands of inputs (precision reducers, servo motors, rare earth magnets, actuators), and the robotics supply chain is "completely dominated by China".
- The truth about rare earths:Rare earths are actually not that rare;the real choke point is refining/processing— there are extremely few processing facilities outside China, and China heavily subsidizes them as political leverage.
- The interface for builders:The government wants VC-style "judging people/execution" signals to guide government capital allocation; it also wants the Bay Area to pull a "magic trick" on "rare-earth-free magnets/new materials" (solving problems with technology rather than government).
Implications for PMs
- Compute will remain tight + expensive for a long time(Andreessen's sandbag argument + "network/CPU/memory will become new bottlenecks"); don't take "infinite token price declines" as a planning premise — Linear/Microsoft have already encountered cost overruns caused by "agents running all day".
- Even if the bubble pops, it will pop differently from dot-com(blue-chip own capital vs high-leverage debt), so don't simply apply the 2000 playbook to "is AI a bubble".
- Supply chain geopolitical risk is a real variable(especially for hardware/robotics/going overseas), but for pure-software PMs, the main impact is the transmission chain of compute price and availability.
8. One-page cheat sheet: the 13 judgments most worth remembering
| # | Judgment | Who said it | Implications for PMs |
|---|---|---|---|
| 1 | Foundation models have no moat / network effects | Evans | Don't rely on "our model is stronger"; look for moats in product, data, and workflows |
| 2 | "A better model doesn't solve the problem" (for scenarios that require correct answers, going from 90% to 95% accuracy is useless—you still have to check everything) | Evans | Choose scenarios: either the model can stably produce correct answers, or switch to a use case with high tolerance for error |
| 3 | Nobody vibe-codes their own ERP, but total software volume + one-off automation surges | Evans | SaaS isn't dead, but it needs to live like day-one again; capture the incremental "improvisational/ephemeral software" market |
| 4 | "How much code the agent wrote" is the biggest vanity metric | Saarinen | Output volume is a signal, not a goal; pair it with quality hedge metrics |
| 5 | Occupy the upstream of the workflow (who feeds context to a swarm of agents) | Saarinen | Value lies where "work comes from", not on the execution end |
| 6 | The hard part is "knowing what to do", not "building it" | Alfred / Evans / Saarinen / Kieran | Treat problem definition, research, and taste as core assets |
| 7 | People are the two slices of bread in a sandwich (framing + taste), and the middle can be automated | Kieran | Your place is at the very front and the very end; don't get stuck in middle execution |
| 8 | Private evals = the biggest IP; you only have control if you can swap models and still hill-climb | Nadella | Build your own evaluation set (turn taste into measurable standards) |
| 9 | Agents are files, migratable independently of models; don't be superstitious about MCP/protocols | Andreessen (vs Rattray on code execution) | Look at whether the model runs smoothly, whether context is saved, and whether safety sits at the API layer |
| 10 | The people coming through the door have become agents; SEO → agent-led growth, platforms are like different species | Cadwallader | Grab visibility first; write content for the "infinite-bandwidth superintelligent agent" |
| 11 | The moat for creative AI is not in the model, but in the data flywheel + product/UX + positioning | Shulman / Staniszewski | Make "creation" enjoyable; pick tracks not yet crushed by giants; the more data is used, the better it understands users |
| 12 | "Now it actually works", but both the utopians and the doomsayers are too optimistic (licensing/inertia are the brakes) | Andreessen / Evans | Technical possibility ≠ behavioral change; adoption speed is constrained by real-world institutions |
| 13 | Compute will remain tight and expensive in the long run; even if the bubble bursts, it won't follow the dot-com script | Andreessen | Don't take "tokens will become infinitely cheap" as a planning premise |
9. Meta-information
- Rewriting strategy: 13 podcast episodes (1 of which has no transcript, only archived) are reorganized by topic rather than by episode — moats / SaaS survival / one-person companies / GTM in the AI era / human-machine division of labor / creative AI as a business / macro geopolitics. Each section is self-contained, includes the Why, synthesizes across guests and deliberately points out disagreements (the "PR volume" debate between Linear and one-person companies; Andreessen's MCP skepticism vs Stainless code-execution).
- Not included: episode-by-episode play-by-play (that belongs to the digest layer); this wiki only keeps stable judgments + frameworks + implications for PMs.
- Cross-reference: how agents enter enterprises, how enterprise systems are restructured → AI Builders · Agent-Native Products and Enterprise AI Implementation (Podcast Digest).
- Sources: see frontmatter
sources:(all 13 entries listed).
2026-09-26 Addendum: Reassessing the market, and also reassessing retention and time scales
Evidence scope: Local truncated transcripts of the Granola founder interview (feed 2026-07-15) and the Sarah/Elad discussion (2026-08-06); this section is partial Oral valuations are not fact-checked for this section, nor are they investment advice.
Getting the entry point right does not mean you will stay in the next stage. The Granola interview treats meeting notes as a useful starting point, with the bigger goal being how people work in the AI era. Competition is not just about who else added a meeting button, but also about where users will continue to process things after finishing the recording. Therefore, current entry retention, next-step task adoption, and paid value can be measured separately; first-mover status and financing valuation cannot replace these three.
“Meeting notes are not the end all, be all value” Translation: Meeting notes are not the whole value or the end point.
There are two opposite kinds of errors in market estimation. Sarah reminds: estimating AI using old software's "headcount × seat fee" may miss the room for charging by service outcome. Elad reminds: while acknowledging the market may expand, you cannot extrapolate recent rapid jumps into large numbers of companies reaching trillion-dollar scale within three to five years. Three to five years and twenty years are different questions, and hundreds of billions and trillions are an order of magnitude apart.
“what does the company look like if they can charge for outcomes?” Translation: If a company can charge by outcome, what would it look like?
For new products, you can run seat models and outcome models side by side, listing adoption rate, delivery cost, billable outcomes, and time separately; then ask customers why they keep using it. This way you are neither boxed in by the old market, nor skip delivery and retention verification because "the AI market is huge". The interview was truncated before full competitive impact and quantitative models were covered, so no assertion is made that big platform competition is harmless or that a certain business route is guaranteed to win.
2026-09-26 Full-text addendum: Consumer AI growth must explain why people come back
That machines are better at something does not mean users lose motivation to participate. Erik Allebest of Chess.com believes people still want to compete against other people, practice, watch masters, and belong to a community. AI can turn capability into post-game reviews, personalized puzzles, and always-available coaches. This case supports a product hypothesis: value may come from helping users make progress, rather than always completing tasks for them. It does not prove that all industries can replicate chess's growth.
“There are shortcuts in the tools.”
Tools can lower the cost of practice, but skills still require actual practice.
Divide growth into hype and sustained use. Chess.com's traffic was once driven by film/TV and social topics; the guest values more the higher baseline that remains after the peak passes. For AI games or interest-based products, generating a good-looking image or a novel conversation can bring trial, but why people come back the next week needs to land on practice, relationships, content, or shared identity. Competitive scenarios also need to distinguish training assistants from cheating in matches; protecting fairness is itself a product capability.
A social network is not just adding a share button. Sarah proposed in the consumer AI interview that practices created by experts can lower ordinary people's learning threshold, but contribution motivation, source transparency, and daily usage frequency must be solved at the same time. A pure prompt directory may give no reason to open it every day. She also acknowledges that old platforms' habits, memories, and pricing may hinder migration; whether a huge winner emerges in multi-person AI remains a judgment to be verified.
“it's words but not accelerators.”
Text on a flywheel diagram does not necessarily drive real growth; behavioral evidence must be seen.
Applicable timing: When reviewing template marketplaces, role communities, or AI coaches, first pick a tight niche group, observe whether new content is reused, whether it improves task outcomes, and whether it brings retention. Record sign-ups, activity, payment, and progress separately; do not use likes or usage counts as a substitute for user value. Original operating figures, replay time, and conflict-of-interest limitations are detailed in Chess.com Digest and Social AI Digest.
Full-text addendum: Judge capability growth, user adoption, and profit attribution separately (2026-09-26)
In an archived interview on 2026-07-27, Benedict Evans used historical platform shifts to discuss the commercial value of AI. Analogies to the internet or electricity can help frame questions, but do not prove that the same profit distribution will happen again. Evidence for infrastructure usage, model capability, actual adoption, and who ultimately profits should be recorded separately.
Why tool capability still needs to go through productization. "Most people aren't tool builders." Cheaply generating code does not mean employees can independently transform multi-person, clearly accountable accounting systems. Customers also usually are not simply buying a slide deck; faster document generation cannot directly deduce that the full value of consulting services has been replaced.
Case: Enterprise analytics goes from an answer to a project. Summarizing customer calls, versus proposing churn improvement plans by combining product usage, CRM, interviews, and competitive intelligence, require different materials, permissions, and verification. High-value work may shift the bottleneck from writing to data connection and process restructuring. Programming, with its large volume of data and repeatable verification, is a relatively successful special case; the interview preserves the debate over whether it can represent all industries.
When to judge startup opportunities. Ask separately who has recurring needs, who is responsible for adoption, who pays, and the cost per use. A consumer product having novelty alone does not prove reuse value; model-layer leadership also does not automatically prove there is no room at the application layer. The interview states "All of the use cases have to be invented." Therefore, specific jobs and economic structures should be validated first; conclusions about product experiments cannot be drawn solely from model upgrades or platform analogies.
Source:Benedict Evans Original Digest.
Full-text addendum: Adoption mechanisms and judgment accountability are more complete than tool usage rates (2026-09-26)
Walleye's Best Of interview was collected and archived as of 2026-09-08; the original recording date is unconfirmed. The approximately tens of billions of dollars in the title refers to assets under management, not company revenue or AI-generated returns. The host's organization Every has provided consulting and training to the guest's company; cases and numbers retain this business relationship and self-reported boundaries.
Why leadership demonstration must be paired with organizational mechanisms. The guest said "you have to lead from the front," and introduced basic training, weekly sharing, cross-departmental exchanges, and internal product trials. These mechanisms can lower the cost of trying, but usage leaderboards only indicate adoption behavior and cannot be treated as productivity or return on investment. The guest himself admits that it is not yet possible to precisely know how much efficiency employees have gained as a result.
Case: Starting with local information organization. Current aggregates analyst notes, earnings reports, email attachments, etc., serving investment teams' high-density information work. Frequent usage and external trial attention can prove demand, but cannot prove correct predictions or excess returns. The Borg that connects all departmental information remains a goal; the original text explicitly states that company-wide email processing has not been solved, so local systems cannot be written up as a completed enterprise omniscience platform.
When to draw lessons. Enterprise internal tool pilots can start with a repetitive information integration task, retaining original materials, usage feedback, and the person who proposed conclusions. The guest said "these tools don't negate the necessity to think." Even if reports are written smoothly, they must still be proofread and explained by the submitter. Meeting transcription accumulation may help with post-mortems, but the interview does not fully address licensing, sensitive data permissions, and retention periods, so the arrangement of "record everything" cannot be directly copied.
Source:Walleye Original Digest.
Sources and related materials
- https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL
- https://www.youtube.com/watch?v=vDY_ocrkQ5w
- https://www.youtube.com/@NoPriorsPodcast
- https://www.youtube.com/watch?v=dlI-5W7d7uU
- https://www.youtube.com/watch?v=6l8oAO_LBx4
- https://youtube.com/watch?v=jH2ZIUKvazU
- https://youtube.com/watch?v=rX3CxlGm-00
- https://podcasters.spotify.com/pod/show/how-do-you-use-chat-gpt/episodes/If-SaaS-Is-Dead--Linear-Didnt-Get-the-Memo-e3h8fqj
- https://youtube.com/watch?v=Uif_qAcgnLk
- https://www.youtube.com/watch?v=G0LTv8hQ5Cs
- https://www.youtube.com/@LatentSpacePod
- https://www.youtube.com/watch?v=xjlYpGaxIPA
- https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8
- https://www.youtube.com/watch?v=Rumft-rsEu4
- https://www.youtube.com/watch?v=Jq3BIGz4vXQ
- AI Builders · Agent-Native Products and Enterprise AI Implementation (Podcast Digest)
2026-10-15 Addendum: Product pivots should retain validated footholds, not protect the initial vision
Evidence scope: Fully read the AI & I interview with Matt Hartman of Factorial Capital, and cross-checked the single episode RbUHf4HeBzo. The Hugging Face experience is an investor's retrospective and cannot be treated as causal proof that all team pivots will necessarily succeed.
First distinguish direction and capability. Hugging Face initially tried a teen chat product, but retention performance was insufficient. The team turned research into usable code and open-sourced it, later pivoting to a platform better suited for model weights and research outputs. What Hartman values is acknowledging the metric problem while retaining the footholds of technical capability and open source. The original product's interface, target audience, and commercial assumptions can change; the capabilities truly accumulated do not have to be discarded along with them.
Why iteration speed must still be paired with insight. Hartman looks for "earned insights": what problem founders understood through specific experiences, and what new things they can learn from user behavior after launch. He compares two-month and two-day feedback cycles to illustrate that observing, forming judgments, and releasing improvements can cycle faster. This is an example of thinking, not a proven universal efficiency multiplier. Releasing unfounded changes faster only accumulates noise faster.
Editorial application derivation: Keep an evidence table for pivots. Write down the original assumptions, actual user behavior, team capabilities that remain valid, and the questions to validate in the next version, respectively. Insufficient retention for a chat product can invalidate a certain use case, but it does not automatically negate the team's ability to turn research into tools. Conversely, strong technology does not guarantee demand in a new market; the next round will still depend on real adoption and sustained usage. Specific checks on adoption and experience can refer to Agent Company Case Library; this section does not repeat the discussion of the maintenance costs of self-built software.
Source: AI & I's interview with Matt Hartman.
October 15, 2026 update: Preserve validated capabilities through a product pivot
Evidence scope: The complete AI & I interview with Matt Hartman of Factorial Capital; episode RbUHf4HeBzo was checked. The Hugging Face account is an investor's retrospective, not causal proof that other teams will succeed by pivoting.
Separate the direction from the capability. Hugging Face initially tried a teen chat product, but retention was insufficient. The team turned research into usable code and shared it as open source, later moving toward a platform better suited to model weights and research outputs. Hartman values acknowledging weak metrics while retaining technical capability and open source as a foundation. The interface, target audience, and business assumptions of the original product can change without discarding the capabilities the team has actually built.
Why iteration speed still needs insight. Hartman looks for earned insights: what founders understand because of specific experience, and what they can continue learning from user behavior after launch. His comparison of two-month and two-day feedback cycles illustrates how observation, judgment, and product improvements can cycle faster. It is an illustrative idea, not a proven universal efficiency multiplier. Shipping unsupported changes faster only produces noise faster.
Editorial application: Keep an evidence table for the pivot. Record the original assumptions, observed user behavior, team capabilities that remain useful, and the questions the next version must test. Poor retention in a chat product can invalidate a particular use case without invalidating the team's ability to turn research into tools. Conversely, strong technology does not guarantee demand in a new market: the next cycle must still examine real adoption and continued use. See the Agent company case library for adoption and experience checks. This section does not repeat the earlier discussion of maintaining internally built software.
Source: AI & I interview with Matt Hartman.