# The AI Realist Manifesto: Beyond the Vibe > [!quote] > "It's a tool. An unusually good one." Pantheon's 17-page internal memo stakes out an "AI Realist" position: bullish on AI as a tool for web teams, skeptical of both AGI utopianism and hype-cycle hating, and organized around the claim that the enterprise web can't be built by prompts alone. The memo is the company line I heard in the 2026-07-30 interview ("we're not doomers, we're not optimists"), now written down with a market thesis, a W3Techs data point, and an operating-model table. ## The four-quadrant map Two axes: Optimism vs Doomerism (vertical) and Maximalism ("big / totalizing") vs Pragmatism ("shipped and measured") (horizontal). | | Maximalist | Pragmatist | |---|---|---| | **Optimist** | AGI Utopian: "Scaling solves it. Let it cook." | **AI Realist**: "It's a tool. An unusually good one." | | **Doomer** | AI Overlord Existentialist: "We are summoning the demon." | Hype-cycle Hater: "We've seen this movie. It ends in a write-down." | The memo calls realism "a lonely position" that is "tragically under-represented" and points to a "silent majority" that wants takes which can "spot BS, but also show how real value is delivered." ## Three principles 1. **LLMs are a tool, not AGI.** Not sentient, not reasoning intelligences, not on the verge of a singularity. Knowing limits matters as much as knowing strengths. 2. **Be skeptical on the economics.** Unit economics are hidden by capex bubbles and subsidized pricing. A realist assumes prices will shift a lot. 3. **Focus on augmenting humans.** Pure replacement is "unhelpful and unrealistic"; large-scale disruption would undercut AI's own value. ## Market thesis: prepare for post-peak AI Overlays Gartner's hype cycle on Moore's adoption lifecycle. The Trough of Disillusionment coincides with the Chasm. The Metaverse is the failure case: VR was real innovation, Meta wrote off $50B. AI is different because "even skeptics admit that the technology has utility." The moment to win is the expectation reset that coincides with early-majority adoption. Google, Facebook, and Amazon all rose after the dot-com crash. The memo labels that spot "OUR TIME TO SHINE." ## Websites are dead, long live the web Concedes that simple-answer traffic won't come back: "People seeking simple answers to simple questions don't need to click through anymore." Then argues the web survives because: - LLMs are "Artificial Interns" for research; people still double-check sources before big decisions. - Agents can't yet safely run transactions, so a human is still in the loop for any sign-up, purchase, or request. - Visitors now arrive later in the journey, at more specific destinations, with expectations already set. - The web remains the "premiere digital medium for first-party engagement" and "the operational layer" agents will act through. Wired's September 2010 "The Web is Dead" cover is the reminder that this claim is 16 years old. ## 2026: the year vibecoding broke through The memo's one hard data point: W3Techs shows "None" (no detectable CMS) declined from about 80% in 2010 to under 30%, then reversed between January and May 2026. B2C SaaS builders stalled and did layoffs; Vercel keeps growing; Netlify revived after pivoting to host generated code. Their read: "the era of vibe coding in production has arrived." Cites GoDaddy's 2026 Super Bowl Airo campaign as the honest picture of where AI serves today: "In actuality Airo mostly serves to autogenerate placeholder pages when people buy domains from GoDaddy, but that's a real use-case." Short-lived sites, landing pages, and side-hustle sites are now AI-achievable. The career anxiety is named plainly: agency hourly rates compared to a $20/mo subscription, entry-level jobs scarce, CS enrollment down. The counter is that the gap between "Goggins Goggles" and mission-critical workloads is "cavernous," and vibe coding has spawned a secondary market for clean-up jobs. Lack of guardrails keeps it "effectively off-limits in the Enterprise." ## Sites that matter go beyond the vibe A "site that matters" is one that is integrated in a larger business context, must respond to changing events, could be a target for bad actors, is shared by a team, or has an important job. For those, "purely prompt-based workflows are never going to be sufficient" because LLMs are probabilistic and compliance contexts require checks. "The enterprise cannot consume a spaghettiburger." Harnesses like Claude Code, Antigravity, and Codex still need product requirements, engineering standards, version control, testing, and DevOps around them. The dominant AI modality is "single player"; teams want agents that act "as a part of the team." Building the web is a team sport: Code, Content, Design, with WebOps at the intersection. ## Freedom within a framework The pitch: right mix of structure, constraints, and autonomy breaks the iron triangle (cost, scope, time). The caveat, in their own words: "this is not a simple matter of getting some Gemini Enterprise subscriptions for the team." Organizations seeing the benefit change their operating model. | Role | Pre-AI job | Modern job | |---|---|---| | Designer | Churning out comps handed to devs | Systems and standards: maintain a design system that delivers approved components downstream | | Content creator | Handing off docs; workflow ends in a ticket | Safe self-service publishing with compliant on-brand components, "starting via prompt but taking over" | | Web developer | Production and maintenance; landing pages, doc-to-post, patches | Integration, automation, innovation; maintain software that lets others self-serve | | Agents | Not present | Upkeep, assistance, review: safe first drafts from structured components, review human output, routine maintenance | Warning that measuring "lines of code or tokens consumed" is how you nose-dive into the trough. ## Conclusion: get confident Quotes Mustafa Suleyman (Feb 2026) predicting most white-collar tasks "fully automated by an AI within the next 12 to 18 months," calls it "ridiculous on its face" and counterproductive because "people's brains don't work very well when they're scared." Notes OpenAI and Anthropic CEOs walking back predictions. Closes: web builders "have the ability to see what's real and what's a hologram." "Let's go." ## My read - **This is the company line I'll be representing.** The interview transcript from 2026-07-30 has the same language verbatim, and the memo confirms it's founder-driven. Any DevRel content I produce for Pantheon should sit in the top-right quadrant on purpose. - **The W3Techs reversal is the one checkable claim.** Five months of data (Jan to May 2026) is a short window and W3Techs samples the top 10M sites. Worth pulling the September numbers before repeating it on stage. - **The Airo line is a gift and a risk.** They describe my former product accurately and without malice. I can speak to it first-hand, and I should be careful not to be the ex-employee who dunks. - **The role table is the DevRel content calendar.** Each row is a talk or a livestream: "what a designer does now," "what a developer does now," "what an agent does on a web team." - **Pairs with [[The AI-Augmented Software Development Manifesto - Ronan Berder]]**: Pantheon's memo is the market and organizational argument; Berder's is the individual practitioner's craft argument. Both reject LOC and token counts as progress. ## Related - [[AI & the Future of Knowledge Work MOC]] - [[DevRel MOC]] - [[The AI-Augmented Software Development Manifesto - Ronan Berder]] - [[3. Resources/Permanent Notes/Where a Human Stands in an AI Pipeline Decides Judgment or Volume]]