Gianfranco's September 2025 Reading List

What’s worth reading this month—and why does it matter?

Each day, I devote three hours to reading. Occasionally an author writes a piece that feels like a gift—crafted with intention for the reader. It tends to be a piece that encourages a reader to pause, then invites them to feel how a strong idea, delivered with empathy, can resonate.

Few writers achieve this.

Each selection did one of three things for me: it sharpened a thesis, challenged a prior, or opened a new line of inquiry. That’s the bar for inclusion.

In this curation practice, I believe the hero of this story isn’t the author of any single essay, nor me as curator, but you—the reader.

If these pieces don’t move you, I haven’t honored the attention you’ve given me. Hold me to that.

Use this as a tasting menu: no set order, categories for context, and a brief note on what each means for builders and capital allocators.

AI Systems & Interfaces

• Peak SaaS - Doug O’Laughlin — Fabricated Knowledge
  • Doug O’Laughlin argues that coding agents such as Cursor and Claude Code are slashing the marginal cost of writing software in the same way YouTube slashed the cost of producing video, flooding supply and eroding incumbents’ economics. The analogy: traditional cable bundles (high-margin, limited creators) peaked ~2010 as cheap, abundant YouTube content exploded; likewise, high-gross-margin B2B SaaS—propped up by scarce dev talent and heavy S&M “moats”—faces an oncoming wave of near-zero-cost software that will fragment markets, commoditize point solutions and shift value back to hardware.
  • Read More

• Context Engineering: Building Production-Grade AI - Akash Bajwa
  • Context engineering is emerging as the critical layer that dictates production AI performance: by orchestrating KV-cache hits, external file stores, and compression models, builders turn raw model gains into faster, cheaper, more reliable agents. Growing evidence—from Chroma’s context-rot studies to Manus and Cognition case studies—suggests the winner in AI isn’t the biggest model, but the system that assembles the most relevant context on the fly. That reframes cost curves and defensibility for every AI startup.
  • Read More

• GPT-5: “It Just Does Stuff” - Ethan Mollick — One Useful Thing
  • Ethan Mollick’s hands-on review of early-access GPT-5 says the leap isn’t only raw IQ; it’s autonomy. GPT-5 is a meta-switch that routes each query to the right sub-model (fast vs. deep “Reasoner”) and self-decides how much “think time” to spend, so ordinary users see top-tier performance without tweaking settings. Humans remain in-the-loop for judgment, but the burden of prompting and model selection fades: gesture vaguely, the AI delivers.
  • Read More

• Content and Community (Stratechery) - Stratechery (Ben Thompson)
  • AI finishes publishing’s unbundling by commoditizing substantiation (turning ideas into artifacts), so search-referral economics collapse as LLMs summarize instead of sending traffic; early court rulings lean “transformative” on training/outputs, making copyright a weak shield. Cloudflare’s default AI-crawler blocks + “pay-per-crawl/402” aims to birth a machine-to-machine content market, but—like past platform shifts—AI-native token creators will likely dominate it, not legacy publishers. The durable strategy for media is community, not pageviews: own the relationship, publish consistent “totems” (essays/pods/videos) that create shared meaning and membership, monetize via subscriptions/rituals, and use AI tactically without surrendering editorial taste.
  • Read More

Operator Playbooks

• The benefits of writing code two days every week without AI, and why the percentage of code written by AI is a vanity metric. - Sarah Tavel (featuring Rekki CTO Borislav Nikolov)
  • Rekki CTO Borislav Nikolov codes “AI-free” two days a week to keep taste and a “nonsense filter” sharp, resisting the slow-boil toward bloated, alien LLM output; the % of code written by AI is a vanity metric if you don’t prune and refactor to your standard. Reviewing AI code lacks a human theory-of-mind, which tempts product-owner detachment; he argues for owning the code you accept, conserving complexity, and prompting to avoid premature token “collapse” (let models plan/score, not just answer). The risk ahead is a deeper black-box stack—models generating layers humans can’t reason about—so preserve agency with regular AI-free reps (even if ~70% of your code uses AI), and build a mental model of how LLMs work (e.g., Karpathy).
  • Read More

• Writing is thinking - Adam Singer — Hot Takes
  • Adam Singer argues that writing is a thinking practice; outsourcing it to AI steadily erodes agency, craft, and originality. Citing evidence (Nature on handwriting’s cognitive benefits and studies showing AI dulls creativity), he urges using tech only after you’ve mastered the basics—be the pilot, not the autopilot—and keeping messy human practice as the antidote to “AI nihilism.”
  • Read More

• How Might We Learn? - Andy Matuschak
  • AI tutors prove useful only when embedded in real projects, not isolated chat windows. Matuschak argues that merging authentic “doing” with cognitive scaffolding allows learning to stick while momentum stays high. He envisions an AI layer that watches every file, click, or coffee chat, then injects dynamic media and spaced prompts, turning work itself into curriculum. This reframes ed-tech opportunity from better courses to operating-system level companions that reshape how expertise is formed.
  • Read More

Market Structure & Moats

• Cloudflare’s Content Independence Day (Stratechery) - Stratechery (Ben Thompson)
  • Cloudflare is using its edge footprint to flip the web’s default from opt-out to opt-in for AI scraping—blocking AI crawlers by default—and, in practice, fortifying Google: publishers must allow Googlebot for search indexing, and Google can reuse that data for Search AI (e.g., AI Overviews). In parallel, Cloudflare is seeding a new “pay-per-crawl” market (HTTP 402, programmatic headers, Cloudflare as merchant of record) that could let agents buy access to high-value sources and reward content that “fills holes in the AI swiss cheese.” Translation: ad-funded discovery is giving way to metered ingestion, and actually changing the web’s economics requires a player with enough power and willingness to enforce new defaults.
  • Read More

• Cursor’s Conundrum - MBI Deep Dives
  • Cursor’s AI IDE sits in a supplier squeeze: users demand frontier coding performance while Anthropic/OpenAI control capability and token economics—and now sell rival tools (e.g., Claude Code). Priority-tier pricing and caching fees forced Cursor to cap usage and raise to $200/mo, risking churn if it downgrades models and margin collapse if it doesn’t. Cursor’s counter: an orchestration layer plus fast, code-specific custom models trained on rich editor telemetry (input selection, plan-to-diff expansion), aiming for a data-distribution flywheel akin to search. The question is whether that specialization can outrun platform risk as labs double-down and the frontier consolidates to ~3–6 players.
  • Read More

• Building for crypto-native vs. building for crypto-adjacent - Shower Thoughts (Substack)
  • Regulatory clarity and institutional demand have pushed crypto across the chasm. Value is migrating to “crypto-adjacent” startups that hide blockchain inside fintech, robotics, or retail workflows. Meanwhile, pure on-chain platforms remain small arenas ruled by a tiny, mercenary elite, forcing constant product reinvention to keep them. The divergence reshapes what counts in talent, defensibility, and distribution.
  • Read More

Capital & Macro

• Honey, AI Capex Keeps Eating … Everything - Paul Kedrosky
  • AI’s hardware binge is warping every scale it touches. Nearly 40 % of Q2 U.S. GDP growth came from server spending, half of it concentrated in four firms and Loudoun County’s data-center sprawl. Meta’s 30 % capex-to-sales ratio now mirrors utilities, signaling that balance sheets, real estate, and power grids are becoming the battleground, not algorithms.
  • Read More

• Where are Vacation Homes Located in the US? - Construction Physics
  • Where are Vacation Homes Located in the US?
- Vacation homes cluster where fixed natural amenities—coastlines, lakes, ski mountains—align with long-standing recreation patterns, and those concentrations have barely shifted despite rising incomes; Florida’s air-conditioned rise is the lone outlier. Census-tract data show over half of second houses sit in just 4 percent of tracts, highlighting land-use scarcity, not demand growth, as the binding constraint. The implication: geographic popularity endures while supply lags economic capacity.
  • Read More

• Fund Size is Strategy — Extreme Power Law Will Make Portfolio Company Conflicts a Relic of the Past - Charles Hudson — The Generalist VC (Substack)
  • Growing fund sizes turn the old “no competing bets” pact into a liability, because mega-funds cannot hit return targets without owning outliers. Vintage guardrails—board exposure, trust signals, easier fund math—worked when stakes were smaller. Now firms weigh seed scattershot, later-stage premiums, or secondary buys, even if it unnerves founders holding minimal bargaining power. A few with real soft power may stall the shift, but capital volume hints the norm’s expiry date is near.
  • Read More

• Apollo’s Web (Net Interest) - Marc Rubinstein — Net Interest
  • Apollo has rebuilt life insurance into a credit-and-liability machine: Athene/Athora vacuum up long-dated annuity and pension-risk-transfer liabilities, then feed premiums into Apollo’s origination + structured-credit stack (private assets, CLO tranching, affiliated platforms), earning a spread at the insurer and fees at the manager while using capital/regulatory arbitrage (e.g., Bermuda reinsurance, risk-based-capital relief) to boost ROE. Scale and simplicity of liabilities are strengths, but new risks loom: beneficiaries losing ERISA/PBGC protections post-transfer, related-party exposures (insurer buying parent-linked deals), opacity/cascade leverage across sidecars/JVs, and tightening scrutiny (IMF/BIS/BoE/NAIC). The model works now—more like a deconstructed bank than a classic insurer—but its complexity is the thing to watch.
  • Read More

• The Calculus of Value - Howard Marks — Oaktree Capital
  • Howard Marks formalizes value as earning power and price as psychology-driven consensus, arguing that long-run returns come from changes in intrinsic value and in the price-to-value gap—not short-term noise. Today’s setup looks “worrisome”: U.S. equities trade at elevated P/Es, record sales multiples, a peak Buffett indicator, tight credit spreads, and rising meme-style risk appetite, even as tariffs, inflation risk, and fiscal strains dim fundamentals. His playbook: dial to INVESTCON 5—trim aggressive equities, add defensives (especially credit with contractual cash flows), and avoid timing extremes while acknowledging AI may justify premium multiples for a select few.
  • Read More

• How Does the US Use Water? - Construction Physics
  • The U.S. withdraws ~322 B gal/day (2015): ~41% for thermoelectric cooling (≈97% non-consumptive), ~37% for irrigation (majority consumptive), ~12% public supply (homes ≈82 gallons/person/day), and ~5% industry (largely returned). Irrigation dominates in the arid West (Idaho’s irrigation ≳20% of state precipitation) and leans heavily on aquifers that many regions are depleting; thermoelectric withdrawals concentrate in the East/Texas. Data centers directly consume ~66 M gal/day (2023)—tiny vs. agriculture (and even golf) but likely to rise 2–4× by 2028; per-gallon economic value is orders of magnitude higher than many crops. Big picture: total withdrawals peaked around 1980 and have fallen (thermoelectric −37%, irrigation −21%, industrial −43%+), groundwater’s share has grown as surface use fell, and smart policy hinges on distinguishing consumptive vs non-consumptive use within the hard limits of precipitation and aquifer recharge.
  • Read More

People & Purpose

• 50 things I know - Cate Hall — Useful Fictions (Substack)
  • Cate Hall strings fifty crisp heuristics that pivot on variance: ideas thrive on it, people with it warrant caution. The list subverts metrics-driven hustle by tagging dopamine as endless and productivity as nonlinear, hinting that tempo, not brute force, dictates startup endurance. Service, framed as the only reliable balm, surfaces as the covert culture moat investors routinely overlook.
  • Read More

• Anhedonia (Mason J. Wang) - Mason J. Wang
  • A compact, first-person sketch of anhedonia—wanting/liking blunted, days washed out; badges and sleep data can’t touch felt emptiness; removing friction kills joy (Stardew cheat), while basketball’s fleeting flow and small sensory anchors (curbside watermelon, onions sizzling) hint at meaning; “live all the days” becomes a daily dare.
  • Read More