The Diary of a CEOThe Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron
CHAPTERS
- 0:00 – 4:26
Zitron’s core claim: generative AI is a “con,” pushed non-consensually
Ed Zitron opens by arguing generative AI is marketed as “magic” while being expensive, unreliable cloud software. He frames the current wave as a misleading, coerced adoption push driven by powerful companies and weak oversight.
- •Generative AI is sold as inevitable and magical, but isn’t autonomous or dependable
- •He views the rollout as the largest “non-consensual push” of technology
- •Claims media, analysts, and governments are being exploited by AI marketing narratives
- •Sets up a myth-busting structure for the discussion
- 4:26 – 6:22
Where the money really comes from: circular funding and hidden AI revenues
Zitron argues Big Tech’s AI revenue is opaque and propped up by a circular financing loop where Microsoft/Amazon/Google fund OpenAI and Anthropic, then count their spending as ‘AI growth.’ He criticizes the use of vague metrics like ‘annualized run rate’ rather than disclosed AI revenue and margins.
- •Most reported AI ‘growth’ is concentrated in OpenAI and Anthropic, which he says are unprofitable
- •Cloud giants allegedly fund these labs while expecting them to drive future cloud growth
- •Public companies don’t clearly disclose AI revenue; ‘run rate’ figures are criticized as misleading
- •Thesis: good news would be clearly disclosed—silence implies weak fundamentals
- 6:22 – 8:02
The real infrastructure story: data centers, GPUs, and extreme power requirements
The conversation turns to CapEx and what building AI infrastructure entails. Zitron uses specific examples (e.g., gigawatt-scale Texas build-outs) to illustrate the density of power, cost, and risk being concentrated into GPU data centers.
- •CapEx explained: long-term bets like data centers and GPU fleets
- •AI GPUs require massive energy, memory, and scale (tens/hundreds of thousands)
- •Example: a 1.2 GW data center and what that implies in physical density and cost
- •Claim: trillions in spending to generate only tens of billions in revenue is structurally irrational
- 8:02 – 12:03
Adoption vs coercion: ‘people use it’—but are they being forced and fear-marketed?
Steven points to explosive adoption and enterprise use; Zitron counters that much ‘adoption’ is forced via default integrations (Google, Docs, Word, Amazon) and fear-driven messaging about being left behind. He argues usage is often substituting for search and doesn’t justify the scale of investment.
- •Widespread use doesn’t equal voluntary, value-driven adoption
- •Default placements (Copilot/Gemini/etc.) create frictionless ‘forced’ usage
- •Media narrative pressures workers and companies to adopt to avoid falling behind
- •LLMs often used like search—helpful at broad trawling but unreliable on specifics
- 12:03 – 20:08
How AI pricing actually works: tokens, subscriptions, and hidden subsidy
Zitron explains token-based pricing and claims consumer subscriptions massively subsidize real inference costs. He argues that when enterprises are moved toward paying closer to actual usage, they balk—revealing weak ROI at true cost.
- •Tokens explained (roughly ~¾ of a word); billed for input, output, and ‘thinking’
- •Subscription tiers obscure real token burn; power users can consume far more than they pay
- •Example figures cited: $200 plan enabling thousands of dollars of token usage
- •Enterprise reactions: token budgets exhausted quickly; ROI becomes hard to justify
- 20:08 – 28:59
Can LLMs improve like cars/internet did? Hallucinations, benchmarks, and ceilings
Steven argues disruptive tech starts flawed but improves; Zitron disputes that LLM progress maps cleanly onto historic curves. They debate hallucination rates, benchmark validity, and whether improvements translate to reliable, real-world autonomy rather than test performance.
- •Disruptive-innovation analogy (cars/internet) vs LLM-specific constraints
- •Zitron challenges how ‘productivity’ and ‘improvement’ are measured in practice
- •Example hallucination: incorrect stock price output in a professional workflow
- •Benchmarks can be ‘rigged’ to model strengths; success rates can still be low on real tasks
- 28:59 – 31:31
Humans vs AI errors: trust, context, and why ‘output is all that matters’ is contested
The discussion shifts from raw correctness to trust and human context. Steven claims people care about outputs; Zitron argues process matters because humans provide empathy, shared learning, and dependable accountability that LLMs can’t replicate.
- •Trust is rooted in accountability, shared context, and iterative collaboration
- •LLMs may retrieve stored info but don’t ‘learn’ like humans without heavy scaffolding
- •Even rare wrong answers can be catastrophic in finance/medicine/legal contexts
- •‘More detailed’ AI outputs can increase error surface area and hidden failure modes
- 31:31 – 1:00:48
AI slop and declining software quality: SEO déjà vu, AI-coded bugginess, infrastructure strain
Zitron claims AI accelerates low-quality content and may contribute to worsening reliability in major platforms. They discuss how AI-assisted coding can flood repos with low-understood changes and increase operational instability, though quantification is challenging.
- •AI ‘slop’ follows a long era of SEO slop—LLMs scale it further
- •Claims Google/Search incentives worsened results; AI answer boxes intensify lock-in
- •Anecdotes: GitHub downtime, AWS incidents, general product bugginess
- •Open source and new coders can ship AI-generated changes without full comprehension
- 1:00:48 – 1:14:27
Job disruption claims: where impact is real, where it’s hype, and why robotics is different
Steven probes job displacement (drivers, lawyers, accountants); Zitron argues broad white-collar disruption data isn’t there. He distinguishes generative AI from robotics and other ‘AI’ domains, saying current disruption is concentrated among roles where bosses accept lower-quality outputs.
- •CEOs have incentives to predict disruption; workers’ experiences are less represented
- •Cites OpenAI research: token spend not correlated with revenue per employee
- •Disruption seen in commoditized creative/translation/transcription work more than broad white-collar roles
- •Robotics/autonomy may matter, but it’s not the same as LLM-driven data center buildouts
- 1:14:27 – 1:17:28
Security & ‘AI danger’ narratives: cyber risk is real, but ‘blackmailing AIs’ are media games
They examine claims of agents hacking and systems escaping control. Zitron accepts cybersecurity risks but argues sensational stories (blackmail, CAPTCHA coercion) often involve humans prompting systems or labs staging scenarios, amplifying fear for attention and investment.
- •Agentic hacking can scale brute-force discovery, but automated hacking predates LLMs
- •Sandbox ‘escapes’ often come down to human misconfiguration and governance failures
- •Blackmail/CAPTCHA stories framed as autonomy are often prompted or staged tests
- •Fear narratives can backfire but were used to create mystique and urgency
- 1:17:28 – 1:46:57
Myth-busting lightning round: growth, China race, job replacement, and ‘agentic AI’
In a rapid ‘myths’ game, Zitron delivers concise rebuttals. He argues economic growth claims are overstated, the China ‘race’ is ill-defined, and agentic AI is mainly LLMs plus scaffolding—marketed as autonomy.
- •‘Enormous economic growth’ is mostly GPU/data-center spend, not durable ROI
- •‘Trillions to beat China’ framed as panic spending with unclear objectives
- •‘Replace all jobs’ lacks evidence; impacts are narrower and uneven
- •‘Agentic AI’ described as LLM-to-LLM workflows with harnesses, not true autonomy
- 1:46:57 – 1:55:19
Bubble mechanics: overspending, circular incentives, and why 2027 becomes a breaking point
Zitron and Steven partially converge on the idea of overspending driven by scarcity of new growth narratives. Zitron argues continued progress requires ever-growing compute spend, and that the funding loop can’t sustain itself—predicting a major rupture around 2027.
- •Tech giants chase the next hypergrowth story; GPUs became the ‘banana tree’
- •LLM progress is tied to continuous massive spending on training and inference
- •Claims: training runs can fail at enormous cost; no clear path to drastic cost reduction
- •Prediction: momentum fades as funding and confidence tighten—especially around public-market scrutiny
- 1:55:19 – 2:08:25
Tech CEOs’ rebuttal and Zitron’s falsification test: what would change his mind
Steven reads CEO lines arguing underinvesting is riskier than overinvesting. Zitron responds that he’d need a huge cost breakthrough plus truly autonomous, reliable capability—‘indistinguishable from magic’—and condemns community harms from reckless data-center deployment.
- •CEO framing: build capacity early to meet demand; ‘not spending on a hunch’
- •Zitron’s bar: ~1000x cost reduction and major hardware breakthroughs
- •Additional requirements: real autonomy and reliability, not benchmark gains
- •Local harms: gas turbines, noise, community opposition, inequitable financing access
- 2:08:25 – 2:21:42
If OpenAI stumbles: IPO pressure, domino effects, ‘tech depression,’ and advice for the public
Zitron lays out a crash scenario: OpenAI struggles to IPO, funding tightens, and companies tied to its spend (cloud providers, Oracle, financiers) face cascading consequences. He forecasts a stock-market and retirement-portfolio hit, venture capital washouts, layoffs, and a long revaluation of Big Tech as mature utilities.
- •OpenAI’s need for perpetual capital collides with IPO timing and valuation constraints
- •Down/flat rounds become likely; comparisons to WeWork-style unraveling
- •Second-order impacts: Nvidia demand collapse risk, cloud guidance resets, VC AI portfolio losses
- •Public guidance: be skeptical of forward-looking hype, watch accounting games like ‘run rate,’ prepare for volatility
- 2:21:42 – 2:27:49
Closing: why he targets AI CEOs—and relationships as an antidote to cynical tech culture
Steven asks why Zitron is so hostile toward AI leaders; Zitron cites being misled and the real human cost of unchecked corporate power. The final question shifts to relationships, where Zitron argues community, appreciation, and uplifting others matter more than chasing hype-driven status and narratives.
- •Motivation: anger at perceived deception, weak accountability, and product/customer contempt
- •Critique: ordinary people face strict scrutiny while AI players get unlimited slack and capital
- •Relationship advice: build community, share appreciation, and ‘find your people’ amid negativity
- •Encourages direct encouragement to creators and loved ones; human connection over ‘machine promises’
