What Valley Founders Are Actually Thinking This Spring

Between February and May 2026, four pieces of long-form founder writing landed with enough weight to change how you might think about org design, mission framing, brand strategy, and your own exit timeline. None fit the format of a news story; the news cycle mostly ignored them. They were long arguments, written by founders for other founders.

Each essay below gets enough author context that you can calibrate how much weight to give the reasoning.

Jack Dorsey & Roelof Botha: "From Hierarchy to Intelligence"

Published: 2026-03-31 · block.xyz 1

Who wrote it: Jack Dorsey is the co-founder of Twitter and the founder and CEO of Block — the parent company of Square, Cash App, Afterpay, TIDAL, and Bitkey, with 2025 revenue of approximately $24 billion 2. Roelof Botha is the former managing partner of Sequoia Capital, which manages over $50 billion in assets, and sits on Block's board 2. Dorsey isn't theorizing from the sidelines — five weeks before publishing, on February 27, 2026, he cut approximately 4,000 Block employees (40% of the company's workforce), explicitly framing the move not as financial distress but as an architectural restructuring driven by AI 3.

"Our business is strong. Gross profit continues to grow, we continue to serve more and more customers, and profitability is improving. But something has changed." 3

Block's stock rose roughly 25% in after-hours trading on the day of the layoff announcement and 17% on the following close 3. This is the live context for the essay.

The central argument

The essay's opening move is historical: organizational hierarchy isn't a management philosophy, it's a technology — one invented to solve the problem of routing information across spans of people larger than one person can directly supervise (the Roman contubernium of 8, scaling to the Prussian General Staff, the first corporate org chart at American railroads, and eventually McKinsey's matrix structures). The fundamental constraint has always been human cognitive bandwidth: a manager can effectively supervise somewhere between 3 and 8 people. Every organizational innovation for 2,000 years has been an attempt to work around that bottleneck without solving it. 1

Dorsey and Botha's claim is that AI is the first technology that can actually replace the information-routing function of middle management — not augment it, replace it. 1

"Companies move fast or slow based on information flow. Hierarchy and middle management impede information flow." 1

The four-component model

Block's proposed architecture has four layers 1:

Capabilities — atomic financial primitives (payments, lending, card issuance, banking, buy-now-pay-later, payroll) with no standalone UI
World Model — two parallel models: a company model (Block's own operations, costs, decisions) and a customer model built from transaction data
Intelligence Layer — the system that, for a specific customer at a specific moment, assembles the right capabilities and proactively delivers a solution
Interfaces — Square, Cash App, Afterpay, and the other delivery surfaces

The human roles collapse to three: Individual Contributors (deep domain experts), Directly Responsible Individuals (DRI, cross-functional problem owners who rotate every 90 days), and Player-Coaches (senior people who both do the work and develop people). No permanent middle management layer. 1

The World Model is what makes this architecture distinctive rather than generic. Block sees both sides of millions of financial transactions daily — Cash App on the buyer side, Square on the seller side. That dual vantage point is the data engine. 1

"Money is the most honest signal in the world. People lie on surveys. They ignore ads. They abandon carts. But when they spend, save, send, borrow, or repay, that's the truth." 1

The question this essay actually asks early-stage founders

The sharpest part of the essay isn't the org design diagram. It's this passage on competitive moats 1:

"If the answer is nothing, AI is just a cost optimization story. ... If the answer is deep, AI doesn't augment your company. It reveals what your company actually is."

The question being posed: what does your company uniquely understand that nobody else can replicate? Block's answer is dual-sided transaction data at scale. If your answer as an early-stage AI founder is "we have GPT-4 access and some fine-tuning," you're in cost-optimization territory. The essay argues that AI doesn't create defensibility — it strips away everything that isn't defensibility, leaving whatever genuine proprietary insight remains exposed.

"Most companies are focused on AI as a productivity enhancer. Few are focused on the potential of AI to change how we work together." 1

This framing has implications for product strategy, hiring, and fundraising positioning that most tactical AI startup advice misses entirely.

Three critiques worth taking seriously

The essay generated substantial pushback in April 2026. Three independent essays raised challenges that aren't easily dismissed:

Anna Lecat (Substack, April 2026) 4: The essay addresses only one of the two things hierarchy has always done. Information routing is one function; resolving conflicts that peers cannot resolve themselves is the other. AI can do the first. It cannot do the second — not because it lacks capability, but because trust between people requires skin in the game. As researcher Mark Ryan's framing has it: what people feel for AI is reliance (like reliance on a reliable car), not mutual trust. Klarna's 2024 replacement of 700 customer service agents with AI, followed by the rehiring of human agents in 2025, is cited as early empirical data 4. When AI delivers unwelcome decisions, people accept them in the short run because they have no alternative. That is not the same as resolution.

Moore Dagogo-Hart (Medium, April 2026) 5: The essay's logic is "remove people + add AI routing = organizational intelligence." Dagogo-Hart's counter-argument is that the real constraint isn't the presence of middle managers — it's bad architecture. He defines a concept he calls "perspective loss": the systematic destruction of relevant viewpoints before they reach decision-making. His claim is that in a 4-week experiment, architectural changes (signal pathways, carrier analysis, outcome verification) produced measurable decision-quality improvements without AI replacing anyone. The summary line:

"Intelligence is not a trait of individuals or machines. It is an architecture." 5

And more sharply: "Speed without integration is just faster blindness." 5

Srikant Madhira (LinkedIn, April 2026) 6: The World Model can only capture what has been formalized and measured. Business environments are open systems — competitor relationships, informal interactions, human activity outside the model's data collection perimeter all determine future outcomes. More pointedly: the model designers decide what gets measured, what gets weighted, and what gets filtered out. The explicit human hierarchy gets replaced by an invisible one — the engineers and product managers who define the model's inputs and outputs. The information asymmetry doesn't disappear; it migrates from organizational roles to model infrastructure 6.

What this means for you

The most productive use of this essay is not as an org design blueprint (Block is a $24B revenue company with specific financial data assets; most of what Dorsey describes requires a decade of transaction data to be viable). The productive use is as a diagnostic prompt: what is the thing your company knows that nobody else knows? If the answer is compelling, the Dorsey/Botha model shows one path for how to build around it. If the answer is vague, that's the real problem to solve before spending any further time on org charts.

Sam Altman: two essays, one through-line

Personal manifesto (~April 11, 2026)

Published: approximately April 11, 2026 (shortly after the April 10 Molotov cocktail attack on his San Francisco home) · blog.samaltman.com 7

The essay combines three things that would normally be separate posts: a personal response to the attack, the launch announcement for Sora (a new video-generation app combining a model called Sora 2 with a social layer for creating and sharing video), and a statement of AI convictions 7.

The conviction worth paying attention to is the infrastructure framing:

"We want to create a factory that can produce a gigawatt of new AI infrastructure every week." 7

For context: a gigawatt is roughly the output of a nuclear power plant. Altman is describing an ambition to build AI infrastructure at a pace that has no historical precedent in civilian technology deployment. He also states that 2026 will see "systems that can figure out novel insights" and 2027 may see "robots that can do tasks in the real world" 7.

The structural observation embedded in the essay: Altman admits that conflict-aversion "has caused great pain for me and OpenAI" and frames it as a personal failure 7. This is notable less as personal revelation and more as a signal about what he thinks went wrong organizationally — and what he's trying to correct.

"Once you see AGI you can't unsee it." 7

What this means for you: If Altman's timeline is anywhere near accurate — novel-insight systems in 2026, physical-world robots in 2027 — the relevant question is not whether to use AI but what the 18-month window of differentiation looks like before those capabilities are commoditized. The "takeoff has started" framing is a timing argument, and timing arguments should affect how you prioritize speed versus perfection in your current product bets.

OpenAI's five principles (April 26, 2026)

Published: 2026-04-26 · openai.com/index/our-principles/ 8

This is a more formal document — a published statement of OpenAI's guiding principles: Democratization, Empowerment, Universal Prosperity, Resilience, and Adaptability 8.

The one passage with the most operational weight for founders:

"Power in the future can either be held by a small handful of companies using and controlling superintelligence, or it can be held in a decentralized way by people. We believe the latter is much better." 8

The essay also names "iterative deployment" — releasing capabilities incrementally rather than all at once — as one of OpenAI's most important discoveries, originating from nervousness around the GPT-2 release 8. That's an interesting piece of institutional history: OpenAI's deployment philosophy was partly accidental, born from caution that turned out to be the right approach.

What this means for you: The decentralization framing is not just a values statement. It has a practical corollary: AI capabilities will be broadly distributed, not concentrated. That makes the "we have access to the best model" moat temporary almost by design. If Altman means what he says here, the long-term competitive environment is one where model access is table stakes, not an advantage — which loops back directly to the Dorsey/Botha question about proprietary insight.

Paul Graham: "The Brand Age"

Published: March 2026 (no precise date published) · paulgraham.com/brandage.html 9

Who wrote it: Paul Graham is the co-founder of Y Combinator (YC), the seed accelerator that funded Airbnb, Dropbox, Stripe, and hundreds of other significant companies. He has been writing essays about startups and technology since the early 2000s; his essays function as reference texts for a generation of founders.

The essay uses the Swiss watch industry's response to the 1970s quartz crisis as a case study for what happens to companies when technology erases the substantive performance differences between products 9.

The quartz movement produced accurate, affordable watches at a fraction of the cost of mechanical Swiss watches. Omega, which had built its reputation on engineering precision, doubled down on engineering and went insolvent by 1981. Patek Philippe, which had begun investing in brand as early as 1968 — taking control of case design, expanding brand visibility from what Graham estimates as "8 square millimeters to 800" — survived and eventually thrived 9.

Graham's thesis is compact 9:

"Brand is what's left when the substantive differences between products disappear. But making the substantive differences between products disappear is what technology naturally tends to do."

The watch industry isn't the point. The point is that this pattern — technology commoditizing performance, brand becoming the differentiator — is "very much a story of our times."

What this means for you: For an early-stage AI founder, the uncomfortable version of this essay is: your current technical advantage over competitors using the same foundation models has a shelf life measurable in months, not years. Patek Philippe started building brand in 1968 — before the quartz crisis hit in the early 1970s. They had a head start. Omega was still betting on engineering performance when the floor dropped out.

The applicable question is not "how do I build a better model" (necessary but not sufficient) but "what is the thing about my company that would still matter after model performance equalizes?" That could be distribution, it could be a specific professional community, it could be a workflow so embedded that switching costs are structural. But it's worth naming explicitly now, not after the commoditization hits.

Elad Gil: "Random thoughts while gazing at the misty AI Frontier"

Published: 2026-04-20 · blog.eladgil.com 10

Who wrote it: Elad Gil is a serial entrepreneur (Color Genomics; Mixer Labs, acquired by Twitter) and one of the most prolific angel investors in AI — early backer of Stripe, Airbnb, Coinbase, and more recently a range of AI-native companies. He published High Growth Handbook, a widely-used reference for scaling startups past the early stage. His blog is notable for being unpolished by design — he publishes thinking-in-progress rather than finished arguments.

This essay is exactly that: 13 numbered observations, described by Gil himself as "some of which are probably wrong." 10 That framing is worth taking at face value — these are hypotheses, not conclusions. Four of them are directly relevant to early-stage AI founders:

Compute as currency 10: Gil argues that tokens (compute) have become the new unit of economic denomination in Silicon Valley — the way cash or equity once were. "What can you accomplish as an engineer?" is increasingly a function of your token budget, not your headcount. The implication for founders: your resource allocation decisions need a compute line item alongside salaries, not just a vague "API costs" estimate. Some companies, he notes, are "inference providers disguised as tools" — Cursor is his example of a product that is effectively subsidizing compute as a user acquisition strategy.

AI eats closed loops first 10: AI automation scales fastest where there's a tight feedback loop — where outputs can be quickly tested against ground truth, allowing the system to iterate. Software engineering has this property (code either runs or it doesn't). Gil frames this as a 2×2: fast-to-close-loop combined with high economic value = fastest AI labor displacement. The implication for hiring: roles with testable, quantifiable output criteria are being automated sooner. Roles that require judgment calls in novel situations (ones where the "right" answer is contested or context-dependent) retain more human value, for now.

Headcount will flatten, then shrink — but not at your stage yet 10: Gil reports that multiple later-stage CEOs told him they are choosing to let attrition shrink headcount rather than do visible layoffs, even as revenue grows 30–100%. This is a later-stage phenomenon; Gil is explicit that "true startups (e.g. a 5 person team) in the short run should continue to scale up headcount like in the olden days as they hit product/market fit but just with more leverage per person." The relevant number to track is not headcount growth but output per person. If your revenue is scaling faster than your headcount, that's evidence of the leverage working.

Most AI companies should consider exiting in the next 12–18 months 10: This is the most contrarian claim in the essay, and Gil presents it as a direct parallel to the 1995–2001 internet era: roughly 2,000 companies went public in that window, and only a dozen or two survived as independent entities long-term. His argument is not that AI is a bubble, but that the acquisition and exit market is favorable right now, while revenue is growing and competition hasn't fully consolidated. Founders running currently successful AI companies, in his view, should "take a cold hard look at exiting in the next 12–18 months, which may be a value maximizing moment." He explicitly exempts OpenAI and Anthropic from this advice; it's aimed at the infrastructure and application layer.

What this means for you: Gil's "closed-loop first" framing is a useful lens for product decisions — if you're building in a domain where feedback loops are inherently slow (healthcare outcomes, legal judgments, strategic decisions), you may have more runway before AI fully competes with you than in domains with fast, measurable feedback. The exit claim is worth pressure-testing against your own numbers: if you're at a moment of strong revenue growth, the question of whether to build toward independence or toward an acquirer is not just financial — it's a strategic commitment that affects team composition, product roadmap, and investor relationships.

A thread connecting all four

Each of these essays, from different angles, is circling the same underlying question: when AI can do most things, what remains valuable?

Dorsey and Botha answer with proprietary data and organizational architecture. Altman answers with infrastructure scale and democratic access. Graham answers with brand and switching costs. Gil answers with closed-loop defensibility and timing your exit before the market figures out which AI companies are durable.

None of these answers are compatible with "we're building an AI-first product and the AI is the moat." If any of these four founders is right — and the four arguments triangulate around similar conclusions from different starting points — the window for building differentiation before commoditization is shorter than most AI startup pitch decks assume.