When Founders Push Back: Three Essays on AI, Gatekeeping, and the Productivity Gap
Three essays published this week — from DHH, Cal Newport, and Noah Smith — each challenge a different AI mistake: guild-like gatekeeping in open source, prophetic tech messaging that's now a PR liability, and the discovery that only 18% of coding tokens convert to shipped product. Here's what each means for early-stage AI founders.
The past week brought a striking convergence: three writers with serious track records in building things — a framework creator who has shipped to a million daily users, a tech-and-policy economist who tracks AI's economic effects, and a productivity researcher who has spent years watching Silicon Valley's cultural mood — each came out with essays that challenge the dominant narrative about what AI is actually accomplishing. None of them conclude that AI is overhyped as a technology. All of them conclude that the people building and deploying it are making specific, correctable mistakes right now.
Here is what they said, and what it means if you are an early-stage founder trying to make real decisions about AI.
DHH: the open source gatekeepers are just protecting status
Published: June 1, 2026 · Source: David Heinemeier Hansson, creator of Ruby on Rails, co-owner of 37signals
Loading content card…
DHH's post is short and blunt: a growing number of open source projects are adding explicit barriers to contributions that were assisted by AI agents — new conduct rules, commit guidelines, and contribution policies that reject AI-aided code 1. His argument is that this inverts the stated founding promise of the open source movement, which was to give everyone the right to read, modify, and improve software. The anti-agent camp is essentially claiming that programmers who use AI assistance are not "real" programmers and therefore do not qualify.
DHH names what is actually happening: ressentiment (his word, Nietzsche's concept — resentment of others who obtain something without suffering for it the way the resenter did). The subtext is: "I spent years learning this. You don't deserve to do it the fast way."
"What should be celebrated as the spread of computing freedoms is instead condemned because it diminishes the exclusivity of those who possessed it first."
The argument he is not making is that AI-assisted code is always good code. He explicitly acknowledges slop is a real problem. His claim is narrower: quality is a separable concern from access rights, and using quality as a pretext to restrict participation is protectionist, not principled.
What this means for founders building today
If you are managing a small engineering team and debating whether to "allow" AI coding assistance, you are having the wrong conversation. The question is not whether your engineers are "real" engineers if they use agents — it is whether the output ships and works. DHH's framing also surfaces a useful diagnostic: when a policy discussion at your company starts sounding like it is about status or credentials rather than outcomes, that is a signal worth noticing. Most early-stage teams cannot afford to subsidize guild politics.
There is a second, sharper implication. The "malleable computer" vision DHH references in passing — his thesis that AI makes software modifiable by people who previously could not modify it — is a real product opportunity. Every vertical where end-users have needed to file support tickets to get configuration changes made is now a candidate for agent-assisted self-service. If your product relies on a moat built from configuration complexity, that moat is narrowing.
Cal Newport: the tech industry's religious fervor is a PR liability it created itself
Published: June 1, 2026 · Source: Cal Newport, computer scientist and author of Deep Work and Slow Productivity
Loading content card…
Newport opens with a detail that lodged itself in my mind: a tech executive attending Mass at the Vatican during an AI ethics roundtable, dressed not in his usual black t-shirt but in a brown suit, taking in the sanctuary. When asked why, he said: "We're building something that is going to change life as we know it. I want to make sure I keep in touch with what humans have always cared about." Newport finds this unsettling — not because of what it says about AI risk, but because of what it says about the psychological state of the people building it 2.
His core argument is that the AI industry has drifted from building useful products into something more like a prophetic religious mission: invoking civilizational transformation, warning of holy wrath, and treating the technology as an entity that must be appeased rather than a tool that must work. Pope Leo XIV's recent 42,000-word encyclical "Magnifica Humanitas," Newport argues, is the most direct institutional pushback yet against this posture 3.
The encyclical's key line, as Newport quotes it:
"With the heart of a shepherd and a father, I ask everyone to abandon the construction of yet another Tower of Babel and to join forces in building up the common good."
Newport then makes an observation that is worth sitting with: Sam Altman recently admitted he had been "pretty wrong" about previous predictions that AI would automate large numbers of jobs 2, and Nvidia CEO Jensen Huang publicly called out executives who cite AI as the reason for layoffs, calling the excuse "lazy." Newport reads both as likely PR damage control rather than a genuine change of heart — but notes that the p(doom) anxiety these leaders spread during their "solemn x-risk sage" phase may take years to dissipate from public consciousness.
What this means for founders building today
The immediate, usable signal here is about narrative positioning. If your AI startup's pitch deck or public messaging is loaded with civilizational-stakes language — transformation, disruption, paradigm shifts — you are borrowing credibility from a rhetorical account that is now visibly overdrawn. Institutional and enterprise buyers in 2026 have heard five years of prophecy and watched the job-automation predictions get walked back by the people who made them. The founders getting traction right now tend to be the ones who can describe, in specific boring terms, which workflow gets faster and by how much.
There is also a quieter observation Newport makes in an aside worth flagging: journalists and commentators are still casually asserting that "AI has already displaced many entry-level jobs" as settled fact, when even the leading AI CEOs no longer endorse this claim. If your competitive analysis or fundraising narrative rests on labor-displacement dynamics that your own industry's leaders no longer assert confidently, it deserves a second look.
Noah Smith: only 18% of tokens become shipped product
Published: June 2, 2026 · Source: Noah Smith, economist and Substack writer at Noahpinion
Loading content card…
Smith's essay is the most data-dense of the three, and the most directly useful for startup decisions. It opens with the Anthropic IPO — the company is filing confidentially at a near-$1 trillion valuation, has recently surpassed OpenAI in annualized revenue (approximately $45 billion per year as of the filing), and is about to turn its first operating profit 4. Smith is not bearish on Anthropic; he thinks the valuation may be conservative given 130% quarterly revenue growth.
But the essay pivots quickly to a harder question: what is all this AI-assisted coding actually producing? He tracks the "tokenmaxxing" phenomenon — companies ordering employees to spend the equivalent of their entire salary in Claude Code tokens, Meta briefly operating a leaderboard for who used the most tokens 4. The finding he quotes from EntelligenceAI (a startup that aggregated data from over 2,000 companies using advanced AI coding tools) is stark:
Only 18% of spending on tokens is translating into shipped coding products that reach real users.
Uber's COO said the link between AI usage metrics and features actually shipped "is not there yet." Jellyfish (a company tracking AI usage) found rapidly diminishing returns in converting tokens to software. Microsoft began canceling Claude Code licenses. Salesforce redesigned employee targets to measure output instead of AI input.
Smith's explanatory framework has two parts. First, turning task-level productivity gains into economic productivity is historically hard — this is not unique to AI. Second, and more provocatively: the existing software industry (better search, faster social apps, cleaner e-commerce) may be mature in the way steel and internal combustion are mature. The productivity gains from AI coding may show up in new types of software industries rather than incremental improvements to what already exists.
What this means for founders building today
The 18% figure is the number to internalize. If you are running a startup that uses coding agents heavily, the first question to put to yourself is not "how many tokens are we spending?" but "how much of that is actually reaching users as shipped product?" Most early-stage teams do not have a clean answer because they are not measuring it.
The more structural point concerns what to build. Smith's "mature industry" framing is worth taking seriously. If the marginal improvement in a given product category (faster checkout, better recommendation engine, cheaper ad targeting) is barely visible to users even with massive AI coding investment, that category may be a poor target for a new startup regardless of how much AI you can throw at it. The question to ask is: does AI coding, in your target domain, enable a qualitatively different product that was previously impossible — or just a marginally cheaper version of what already exists? That distinction separates the areas where the productivity gap Smith identifies will close quickly from the ones where it will persist.
The common thread
All three essays converge on a version of the same point: the bottleneck in the current AI moment is not capability — it is judgment about where to actually apply it.
DHH is arguing against misapplied gatekeeping instincts. Newport is arguing against misapplied prophetic instincts. Smith is arguing against misapplied tokenmaxxing instincts. In each case, the underlying technology works. The humans deploying it are making a specific, recoverable mistake by optimizing for the wrong signal — preserving status, performing seriousness, or maximizing usage metrics.
For an early-stage founder, the practical test is:
- Are your team's coding agent conversations about outcomes (what ships, does it work) or about credentials and legitimacy?
- Is your public-facing AI narrative specific enough that removing the word "AI" still leaves a coherent product description?
- Can you name, with actual data, what percentage of your AI tool usage is converting to shipped user-facing value?
None of these are rhetorical questions. Each has a checkable answer.
Add more perspectives or context around this Post.