Tuesday, September 30, 2025 · 2 newsletters

The Sparse Month Earned Its Slot

ai-as-craft-question · founder-identity-reckoning · leaner-teams-harder-skills · ai-as-normal-technology · decision-proxy-slope · consumer-to-enterprise-bridge · going-to-see-the-physical-thing

Roughly 90 emails across five publishing weeks. The thinnest month of the year by volume, and the inbox knew it. There was no breaking news cluster, no dominant industry dispatch, no macro shock past the September Fed cut everyone already had priced in. What did show up, sorted across the month, was a slow rotation away from "what can AI do" and toward "what is the human still supposed to do," paired with a parallel rotation by operators publicly reconsidering their own roles. September was the month where the writers who had nothing news-shaped to say did the better work.

The Month in One Sentence

This was the month the AI conversation moved from capability to craft, and a quiet cluster of founders and operators started reckoning with what their job actually is in a world where the model does the first draft.

Arc: AI From Capability to Craft

The month opened with the cleanest version of the operator question and closed with the first honest measurement post anyone had read in months. The evolution week to week is the cleanest tell of where the conversation actually moved.

Week one set up the human-side question. Brianna Zuniga at Circular Architect ran the load-bearing aphorism of the month: taste is attention made visible. Her companion line, "let AI take the churn, let us keep the care," became the soft version of the same argument operators were running on the engineering side. Sahar Mor at AI Tidbits ran the workflow version with his breakdown of the move from desktop pair-programming agents (Cursor, Windsurf, Claude Code) to asynchronous cloud agents (Devin, Codex, Jules, Factory, Cursor Background Agents). The bottleneck, he argued, was about to shift from "can the agent code" to "can I review enough PRs in parallel."

Week two sharpened the frame. Arvind Narayanan relaunched his newsletter as AI as Normal Technology, a rebrand away from AI Snake Oil. The clarification mattered: "normal" does not mean predictable, it means AI's effects will diffuse the way other powerful technologies have, in years not months. Ethan Mollick at One Useful Thing, in "On Working with Wizards," argued that the co-intelligence framing from his book is starting to give way to something different: we are moving from partners to audience, from collaboration to conjuring. He fed his book plus 140 posts into NotebookLM, got a video back good enough to need fact-checking from scratch, and asked who is doing the verifying. Narayanan and Mollick were running the same play from different sides.

Week three returned to the craft question with two writers in direct tension. Amanda Natividad at The Menu wrote "AI can save you time, but don't let it steal your reps," the cleanest version of the month. Her theory: AI reduces friction, and friction is where craft is forged. She came up on Twitter writing thousands of often-bad posts that sharpened her instincts; she is honest that if LLMs had been what they are today she might have skipped that ramp and ended up "indistinguishable, passable." Sean Ellis ran the inverse argument in "AI as the New Mentor," that for new marketing grads AI may be the best thing that ever happened to them. Both are right in different domains. Natividad's frame is that AI removes the productive struggle that builds taste. Ellis's frame is that AI removes the gatekeeping that prevented juniors from operating at senior leverage.

Week five gave the conversation its first honest measurement post. Ethan Mollick at One Useful Thing unpacked OpenAI's GDPval test: experts averaging 14 years of experience in finance, law, and retail designed realistic four-to-seven-hour tasks, then humans and AI completed them blind. Human experts won, but barely, and the main reason AI lost was not hallucination but formatting and instruction-following failures, which Mollick noted are areas of rapid improvement. His read: this is not job replacement, it is task replacement, and the gap between the two is where the next two years of operator decisions live. Carilu Dietrich at Hypergrowth Leadership ran the case-study companion on Atlan's "AI-First" to "AI-Native" journey: a 400-person company built 152 agents that made 4,000 runs in five weeks. Sankar's line that anchors the piece: AI-First means AI is the primary solution; AI-Native means reimagining workflows from first principles. The hype cycle had moved a half-step toward instrumentation by month-end.

Arc: Founder Identity, Four Writers, One Reckoning

The strongest through-line of the month was not in any single week. It was a slow accumulation of founders and operators publicly working out what their role is becoming. The pattern did not surface as a cluster until the last week, but by then it had been building for thirty days.

Week one ran the high-altitude version. George Mack at High Agency updated his list of how to spot high agency people. The signals worth carrying: the golden question (who do you call when stuck in a 3rd world prison cell?), weird teenage hobbies as evidence of going against social pressure when it is hardest, treadmill energy, unguessable opinions, immigrant mentality, sending niche content without checking engagement first, being mean to your face and nice behind your back, and quitting something of prestige. Signull at Signull vs. Noise ran the negative-space version in "the myth of being well rounded," arguing that schools and big companies trained a generation to sand themselves down smooth. Ben James at Ben by Fax ran the social version in "Aliveness and where to find it," built around a four-quadrant frame where quadrant four is "creating together" and the other three are easier defaults.

Week three made the bet explicit. Signull at Signull vs Noise ran the most provocative version of the month: every startup is a bet on mispriced risk, and the founder's job is to identify the risks they personally can take that others can't. Low burn, cultural fluency, taste risk. The line worth sitting with: "you are not a founder, you are a portfolio manager of asymmetric, illiquid, non consensus" bets. Ami Vora at The Hard Parts of Growth ran the operator-side companion on the limbo of considering a job change without actually moving, prescribing an 8-week focused window and naming the half-in, half-out state as the most insidious failure mode.

Week five turned the bet into a chorus. Four writers, three different angles, the same instinct. George Milton at Gross to Net wrote "I Just Hired My Replacement" about stepping down as CEO of Yellowbird Foods after thirteen years; the honesty about the vain part and the prestige problem and the question of which pieces are him versus which exist independently is what made it land. Kyle Poyar at Growth Unhinged announced leaving his VC operating partner role to go full-time solopreneur, with the stated aspiration of becoming a modern Gartner or Forrester for B2B startups. Brianna Zuniga at Circular Architect wrote the inverse trajectory in "I don't want to belong to any club that would accept me as a member," about finally landing at 776 after eighteen months of applications, and the moment of getting in being the moment you start questioning whether being in was the point. Alec McNayr at Alec McNayr revisited his 2019 "Don't Lose the Courage to Experiment" essay six years later and admitted he was the one who needed to hear it now, after twenty-two open mics this year. None of these were coordinated, which is what makes them a signal rather than a trend piece.

Arc: Leaner Teams, Harder Skills

The management writing in September converged on a single picture across two weeks. Teams will be smaller, roles will be flatter, and the work AI cannot do (taste, judgment, knowing which experiment to run next) is the work that gets harder to outsource.

Week two had the senior-IC version. Julie Zhuo at The Looking Glass used the relaunch of her book to make the case directly: pure middle-manager roles will decline because AI reduces the coordination work that justified them, but the underlying skill of management becomes more valuable as teams get leaner. The Avengers-versus-Captain-Marvel framing is the line that will get quoted. Ami Vora at The Hard Parts of Growth ran the most useful tactical piece of the week on writing a parental leave plan that builds team capacity, with the reframe being to treat the leave as a chance to grow your peers and reports into ownership of workstreams you used to hold. Kyle Poyar at Growth Unhinged did the cold-water version with "Do you need a GTM engineer?" Poyar expected GTM engineering to be the hottest job of 2025 and instead found 45 job posts in the past month, 128 in the past three, mostly at OpenAI, Ramp, and Webflow. There are more LinkedIn posts about the role than actual companies hiring for it.

Week five extended the operator playbook. Chandra Narayanan at Opinionated Intelligence ran the sharpest operations piece of the month with "Death by 10000 dashboards," diagnosing how the self-service BI promise turned into conflicting truths, endless scavenger hunts, and dashboard bloat that lingers forever. The framing of BI as a treadmill measured by output rather than clarity is the line operators should screenshot. Paul Stansik at Hello Operator closed the week with "What It Takes To Win," anchored on Chuck Noll's line that "champions are champions not because they do anything extraordinary but because they do the ordinary things better than anyone else." George Mack at High Agency ran the playful complement, "9 fun ways to increase your agency with zero grinding required," with the giant home whiteboard, the luck razor, and swapping "problem" for "puzzle." Read in that order, the three compose into a single argument about why most teams know what to do and still do not do it.

Arc: The Decision-Proxy Slope and the Consumer-to-Enterprise Bridge

The industry-shaped AI conversation finally surfaced two real patterns in the last two weeks of the month. Both were quiet, both were the kind of post operators can act on rather than admire.

Week four opened the decision-proxy thread. Alex Furmansky at Magnetic Growth wrote the confession piece about letting ChatGPT talk him out of flying across the country for a third date. The model's advice was, in hindsight, correct. The unsettling part of the essay was the next beat: if AI was right about the date, why wouldn't you let it write your strategic recommendation, your difficult email, your boss-facing analysis? The boundary between "tool that helps me think" and "proxy that thinks for me" turns out to be a slope, not a line. Carilu Dietrich at Hypergrowth Leadership ran the marketing-side version of the same anxiety: SEO is tanking, Answer Engine Optimization is the replacement discipline, prompts replace keywords, Wikipedia and Reddit now rank in LLM responses far more than they ever did in Google, and new pages can get indexed by an LLM in 48 hours. The Furmansky date and the Dietrich AEO piece are the same story at different altitudes.

The same week named the consumer-to-enterprise pattern. Kyle Poyar at Growth Unhinged framed it as "the rising consumer to enterprise AI playbook," anchored on Fyxer, the Cursor-for-email startup that went from $1M to $17M ARR in eight months and closed a $30M Series B led by Madrona. The old consumer software model was built for churn and tiny ARPU; AI is breaking both, with breakout consumer AI companies reaching $100M ARR in under two years. Nikhil Basu Trivedi at next big thing ran the vertical AI version with his Waldo Series A piece: the race is on to be the AI leader in every vertical, and the winners will combine technical depth with domain expertise. Harvey in legal, Abridge in healthcare, OpenEvidence in medicine, Confido in CPG, Elicit in scientific research, WindBorne in weather, Waldo for agencies and brands.

Week five extended the consumer thread without naming it as such. Basu Trivedi ran two Footwork portfolio announcements: Anything, the AI app builder that hit $2M revenue run-rate in two weeks by focusing on production-ready output, and Benable, the recommendation platform with 500,000 users across 150 countries sharing 10 million recommendations, growing 50% month-over-month almost entirely from word-of-mouth. Christopher Dowd at AI Residency ran the consumer-brand companion piece on building generational community-to-commerce brands, with five strategies (start with an identity wedge, co-create with community, layer network effects on community resonance, position incumbents on the outside looking in, anchor growth to IRL moments) and a map of examples that read like where consumer venture is actually going. Trivedi and Dowd are making the same bet: the next decade of consumer growth comes from products that bring offline trust online.

Arc: Going to See the Physical Thing

The piece that lingered longest from the month was the one that did not fit any cluster. In week four, Citrini Research sent a drone to Abilene, Texas, to look at Stargate. The framing was an argument against pure screen-based investing in the AI data center buildout: chips don't float, fiber doesn't spool itself, power doesn't magically appear at the edge of the rack. They partnered with Hunterbrook Media and went to physically witness the megawatts meeting the mud. The companion read came from dynomight in his "Shoes, Algernon, Pangea, and Sea Peoples" essay, which opened with a riff about how we are in the "waning days of the People Read Blog Posts About Random Well-Understood Topics Instead of Asking Their Automatons Era." Both pieces argued the same thing from opposite ends: as more of the world becomes mediated by models, the value of direct contact with the actual thing (the field, the blog post, the megawatts) goes up, not down.

The Citrini AI-infrastructure thesis ran parallel through the month and was the only real macro thread. Week three had his single-stock long thesis tracing the buildout from servers (Nvidia, Supermicro, Celestica) to data center interconnect (Credo, Ciena) to high-voltage power (American Superconductor, Siemens Energy) and now to the new data centers themselves. The trillion-dollar capex story has at least three more chapters in it. Week one had his macro memo (cyclical slowdown, September Fed cut, equity volatility through Q3) which graded itself in advance. The Stargate field trip was the embodied version of the same discipline: go look.

The Story of the Month

The story of the month is the slow public unraveling of the founder identity, treated as a single arc across five weeks. Mack and Signull and Ben James in week one were writing the analytic version, arguing that the institutions of 2025 reward the smooth and the passive, and that the people building the future are the ones willing to stay weird and take social risk. Vora and Signull in week three were writing the calibration version, arguing that the half-in half-out state is the failure mode and that your edge is structural, the risks you can take that others cannot. By week five, four writers with real track records (Milton thirteen years at Yellowbird, Poyar four and a half years at his VC role, Zuniga eighteen months of applications, McNayr six years past his own essay) were publicly admitting they were rebuilding the version of themselves that built the last thing.

This matters more than the AI-coding cluster or the consumer-to-enterprise pattern because it is a leading indicator of what 2026 looks like inside companies. The agent-economy noise is a story about tooling. The founder-identity story is a story about supply. The next round of company formation is going to come from people who have already built one thing and are explicitly unsatisfied with the version of themselves that built it. That is the kind of signal you only catch when the inbox is too thin for the trend pieces to drown it out. The sparse month surfaced it because nothing else demanded the slot.

In Retrospect

The "GTM engineering is the hottest job of 2025" prediction aged badly. Poyar's own October 2024 prediction that GTM engineering would be the breakout role of the year ran into a brick wall of 45 actual job postings in a month. The corrective came from the same writer, which is the discipline worth noting. The role exists, the discourse exists, the LinkedIn posts exist; the hiring volume does not.

The cloud coding agent thesis from week one looked like the dominant frame, and it was not. Sahar Mor's breakdown of the move to async cloud agents was the sharpest workflow piece of week one, and by week five the actual conversation had moved past it to the measurement question. The cloud-agent shift is real and will compound, but the operator question that ended up mattering in September was Mollick's GDPval read on whether the agent's output is good enough to ship, not whether it can be spawned in parallel.

The Reid Hoffman-Alpha Schools interview looked like an inflection in week one, and the rebuttal landed harder. Falik's response to the Hoffman interview was the sharpest piece of the first week, and her question (the real issue isn't the pace, it's the direction) was the framing that held through the month, not Hoffman's efficiency case. AI-mediated education is happening regardless. The question Falik kept open is whether the metrics being optimized are the ones that matter, and a month later no one had answered it.

The "AI is the new mentor" framing got more complicated. Ellis in week three argued AI would let juniors operate at senior leverage; Natividad in the same week argued AI would let juniors skip the reps that build taste. By month-end, with Atlan's 4,000 agent runs and OpenAI's GDPval measuring expert-versus-AI on actual workflows, the honest read is that both are happening at once and the operator-level call is how much friction you intentionally preserve in your own practice. The mentor-versus-rep-stealer question did not resolve. It just got higher stakes.

What to Carry Into Next Month

The AI conversation has finished moving past capability and is now a craft conversation in disguise. Natividad and Ellis pulled in opposite directions on whether the next generation of operators will still get the reps that build taste. Mollick and Narayanan pulled in opposite directions on whether the human is the partner or the audience. The honest read across all four is that the question is going to be domain-specific, and the operator-level decision is how much friction you intentionally preserve. October's job is to track which companies build the deliberate-friction discipline into their workflows, and which ones discover six months later that the muscles they stopped using are the ones they needed most.

The founder identity arc is the leading indicator to carry into Q4. Mack and Signull in week one, Vora in week three, Milton and Poyar and Zuniga and McNayr in week five: this is a thread the inbox kept returning to in different costumes for the entire month. The next round of company formation will come from people who have built one thing and are explicitly unsatisfied with the version of themselves that built it. Watch for it in two registers: who leaves a comfortable role next quarter, and who publishes the essay before they leave. The essay is the lead indicator. The departure is the lag.

If you only carry three pieces from September into October, I would suggest Amanda Natividad's "AI can save you time, but don't let it steal your reps" for the cleanest framing of the AI craft question of the year, George Milton's "I Just Hired My Replacement" for the most honest founder essay of the month, and Ethan Mollick on Real AI Agents and Real Work for the cleanest read on where AI actually lands in expert workflows. The month told me three things in sequence: the AI conversation has fully rotated to craft, the founder class is reckoning with its own definition of success, and the work that survives the cut is the work that requires judgment. Sparse months reward the writers who knew what kind of month it was. September was that kind of month.