Week 42 · 2025-10-13 → 2025-10-19 · 19 newsletters

AI Hype Meets Measurement

ai-in-production · builder-economics · consumerism-and-belief

The week of October 13-19, 2025. Twenty emails across seven days, mostly Substack writers, mostly thinking out loud. No single news story dominated. What did emerge, across writers who do not normally talk to each other, was the same set of questions about AI moving from demo to production, the unforgiving math of building a software company, and a recurring critique of how belief gets manufactured and sold. A short week. A coherent one.

AI in Production: From "Look What It Can Do" to "Show Me the Numbers"

The largest thread of the week was AI maturity, but the angle was not capability. The angle was measurement, deployment, and whether the adoption curve is actually bending.

The slowdown question. Kyle Poyar at Growth Unhinged led the week with the most pointed data: the Ramp AI Index showed paid AI adoption declined in September, the second monthly spending decline of 2025. Tech leads at 73% paid penetration and is plateauing. Finance sits at 58%. Retail, construction, and accommodation all sit below 35%. The good news inside the bad news: annualized retention of AI products is on track to exceed 80% in 2025, up from 60% in 2023. Stickier when it lands, harder to land.

Practical usage advice. Ethan Mollick at One Useful Thing wrote the week's most useful operator-level piece, an opinionated guide to picking among the nine cutting-edge models (Claude, Gemini, ChatGPT, Grok, plus the open-weights families Deepseek, Kimi, Z, Qwen, and Mistral). His framing that "any other AI service you use is powered by one or more of these nine" is the right mental model for a market where most consumer AI products are wrappers. He annotated OpenAI's chart of actual ChatGPT use cases (more information-seeking than expected, less casual chat than expected) with model recommendations.

The frontier model rumor. Mark Humphries at Generative History tested an unidentified model that appeared in Google AI Studio's A/B testing surface (speculated to be Gemini 3) on handwritten historical document transcription. His verdict: not only near-perfect transcription at expert-human levels, but unprompted reasoning about the documents that he describes as "the most amazing thing I have seen an LLM do." A historian's verdict on visual precision plus reasoning, written before the model is even named, is worth saving.

Sora 2 and the social layer. Marily Nika at AI Product Academy spent a week inside Sora 2 and came out describing it as "a new medium for human expression and confusion." The vignettes she chose (someone making a cameo of her avatar getting pushed on day one, someone else taking her avatar on a spa date on day two) capture the uncanny problem better than any product review. Creation friction collapses to a prompt; consent and identity become the open questions.

Agents in real workflows. Carilu Dietrich at Hypergrowth Leadership wrote up Tebra's 11-agent battlecard orchestration, where Kevin Marasco's team daisy-chained eleven narrow agents (research, overview, messaging hierarchy, feature table, FAQ, email sequence) to produce competitive intelligence inside Slack in five to ten minutes. The operational insight: agents are more effective with specific tasks than broad ones, and the win is in chaining them, not in asking one agent to do everything.

The take: the AI conversation this week was already past the "is it real" phase and into "what's actually shipping, what's actually sticking, who's actually paying." Poyar's adoption-plateau data and Mollick's "pick a model" guide are both written for the same reader: someone who has stopped asking whether to use AI and started asking which one, for what, and at what cost. That reader is the market now.

Builder Economics: The Unforgiving Math

A separate thread ran underneath the AI thread: writers staring directly at the numbers of building a software company. No one was selling the dream this week.

Kyle Poyar's first piece of the week, built on ChartMogul's dataset of 6,525 software companies with a decade of history, made the line of the week: "It's easier to get into Harvard or Stanford than it is to reach $25 million ARR within 10 years." The piece is the cold-shower companion to every "20-person team, $10M ARR in 18 months" headline.

Matt Brown ran the structural version of the same argument with "Your data model is your destiny," making the case that the database choice a founder makes early ("what deserves to be the center of gravity") cascades into UI, pricing, and GTM, and is "nearly impossible to change" by the time architecture solidifies. His point that most companies should not innovate on their data model, but the breakouts almost always do, is the operator framing.

Nesrine Changuel at Product Delight Tips read Will Guidara's Unreasonable Hospitality and pulled out nine principles that map fine dining to product, with the central frame being that "functionality creates trust, but it doesn't create love." The piece pairs naturally with Taylor Majewski's General Partnership podcast episode with Chris Best (Substack) and Andrew Mayne (OpenAI), where Best repeated his "do everything but the hard part" product philosophy and argued that "all economics are downstream of culture." Two writers, two formats, same point: the moat is in the human layer, not the feature list.

Christopher Dowd at AIR opened applications for Cohort 2, with a thesis statement worth quoting: they want founders who see opportunity in "broken systems, platform shifts, overlooked niches, or emerging secular trends." The questions they want founders to answer ("how will being surrounded by a handful of the best AI operators in the world change the trajectory of your business forever?") are the right ones for the moment. Upen at MicroSaaS Idea provided the counterweight on the small-bet end of the spectrum, with a roundup of solo builders hitting $1K to $40K MRR on narrow tools (Brainrot at $26K in 30 days, My AskAI at $40K MRR over two years, Savewise on cashback comparison). Javier López at VC Hiring ran the week's job board, ten VC roles plus YC S25 startups plus a16z portfolio.

The take: the building-a-company conversation has clearly bifurcated. One side is the Poyar-Brown axis arguing that the math is brutal and the structural decisions matter more than the surface ones. The other side is the AIR-MicroSaaS axis arguing that the platform shift is real and the wedge is narrow. Both can be true. The writers who admitted both were true wrote the better pieces.

Consumerism and Belief: Three Writers, One Critique

A surprisingly cohesive cluster: three writers, across very different beats, all writing about manufactured belief and the systems that monetize it.

Brianna Zuniga at Circular Architect opened the week with the strongest piece of it: "The Mimetic and Memetic Economy," a personal essay about a stranger at church sending her a pre-filled $300 shopping cart of supplements after overhearing her cough, with an MLM ambassador pitch attached. Her observation that "every era has its pyramid; the products change, but the architecture of belief doesn't" is the line that organizes the week. She traces the lineage from Mary Kay and Tupperware through online coaching academies and manifestation masterminds to the current wellness-collective form.

George Milton at Gross to Net ran the macro version with "We'll Never Be Free," arguing that consumerism is not a feature of capitalism but a feature of any society that develops surplus production, mass communication, and human status anxiety. The historical sweep (Roman senators, medieval sumptuary laws, Soviet jeans black markets, modern Scandinavia) is the right move; the conclusion that "you've always been trapped" is the move that earns the title.

grace at The Friday Brief ran the lighter applied version, cataloging the week's "wait, what?" marketing moves: IKEA's phone beds, MrBeast's banking play, YouTube pushing podcasts to look like shows, LinkedIn paywalling competitor analytics at $99 a month. The pattern she names without naming it is that the platforms keep finding new surfaces to monetize attention, and the brands keep finding new ways to manufacture want.

The take: Zuniga, Milton, and grace are working three sides of the same critique. The MLM in the church pew, the iPhone bought to avoid green-bubble shame, and the LinkedIn analytics paywall are the same machine at three resolutions. The writers who saw the connection wrote the week's most durable pieces.

Outside Interests

Mark Humphries at Generative History doubled as the historian's note alongside his Gemini-3 test. Lingthusiasm hit nine years with an episode on how the nose shapes language (nasals appear in 98% of the world's languages), which is the kind of grace note a sparse week needs. Xinran Ma at Design with AI ran a comprehensive Lovable crash course comparing it to V0 from a product designer's perspective. Alec McNayr accidentally sent a Westside Story Club recap to his personal list and left it up as a standing monument to moving too fast. PSFK Weekly ran four themes driving travel experience, with the central frame that "the competitive edge in travel is no longer scale, it is sensitivity." Citrini Research ran the week's most serious investment thesis with "Power Struggle: Natural Gas," arguing that AI data center demand plus LNG export ramp will move US gas markets from supply-push to demand-pull, with structurally higher prices ahead.

Three Takeaways from the Week

The AI conversation in mid-October 2025 had already moved past the "what can it do" phase. Poyar's Ramp adoption data, Mollick's pick-a-model guide, Humphries' historian's verdict, and Dietrich's 11-agent battlecard writeup are four different angles on the same shift: the question is no longer whether AI works, but where it sticks, what it costs, and which model fits which task. That is the regime change worth carrying forward.

The single thread connecting the most durable writing of the week was a critique of manufactured belief. Zuniga on MLMs in church pews, Milton on consumerism as a human-systems problem rather than a capitalism-specific one, and grace on platforms finding new surfaces to monetize attention are not adjacent topics. They are the same topic. The writers who saw the connection produced the week's best work.

If you only revisit three pieces from the week, I would suggest Brianna Zuniga's "The Mimetic and Memetic Economy" for the cleanest essay of the week, Kyle Poyar's "Is AI adoption slowing down?" for the operator-level read on where the AI market actually sits, and Mark Humphries on the unnamed Gemini model for the first credible historian's verdict on AI plus visual reasoning. A short week. The pieces that earned the slot earned it by being specific.