August 3rd, 2026

Tier 1 — PwC, EY, KPMG all caught publishing AI-slop reports with fake citations (MACA, CourseBuilds/Aria)

Verdict: verified shipped — story confirmed by Financial Times investigation, GPTZero detection, and PwC’s own acknowledgement. PwC Middle East reports contained fabricated citations, dead links, footnotes pointing at ChatGPT-generated sources, and outdated case studies passed off as recent. EY and KPMG have quietly retracted comparable AI-generated publications. This is the exact failure mode Roy has been designing MACA’s copy pipeline against for months — the market just gave him a live case study proving the “must pass human review” bar isn’t paranoid, it’s the actual moat. Action: cite this story on the CourseBuilds Aria pitch as the wedge — Zaicek’s team will remember it. For MACA, add “no fabricated stats, no dead links” as an explicit line in the ad copy QA prompt. Under 5 minutes’ work, permanent defence. Sources: TechSpot, Financial Times, GPTZero.

Tier 2 fill — Oracle’s $165B AI supercampus hit with multibillion-dollar cost surprises (Ben, MACA batch strategy)

Verdict: verified shipped — The Information’s exclusive on Oracle data center cost overruns landed Aug 2. SpaceXAI is simultaneously exploring major Texas expansion. This is the “compute is not getting cheaper this quarter” signal — reinforces the Sonnet 5 intro pricing dying Aug 31 warning from yesterday’s brief. Enterprise pricing has room to move up before it moves down. Action: for Ben and MACA, front-load anything you were going to run on Anthropic before Aug 31 to lock the intro rate. For the CourseBuilds pricing ladder, this backs the argument that AI budgets get more strategic in 2026, not less — Aria needs a partner who knows how to spend $200/month well, not a $50/seat SaaS. Source.

Tier 2 fill — Flock Safety AI license plate readers misread plates in 71% of alerts (CourseBuilds trust framing)

Verdict: verified shipped — Business Insider investigation of a California PD found Flock’s AI surveillance system sent misread plate alerts 71% of the time. This is the perfect counter-example when Roy is selling the CourseBuilds bespoke pilot into Aria: the pitch isn’t “AI is amazing,” it’s “AI without discipline is Flock Safety.” The wedge is Aria’s own documents in a Claude project delivered by a human who checks the output — not a bolted-on vendor tool. Action: file this in the CourseBuilds objection-handling deck (~/Reeve/docs/superpowers/specs/2026-04-14-coursebuilds-bespoke-pilot-design.md) under “what happens when AI is deployed without a wingman.” Source.

Your MACA ad-copy discipline is the exact pre-answer to the PwC scandal

While PwC, EY, and KPMG are publicly retracting AI-generated reports this week, you’ve been iterating for months on the “pass human review without being obviously AI-written” bar in MetaAdCreatorApp/docs/prompts/ — see Batch 3 (No Mirrors, No Phone Scrolling, Boxing is Therapy, Mirrors Extended, Apr 7) and Batch 4 (premium boxing-as-fight-training positioning, Apr 9). Active-projects.md line 37 literally says “differentiator must be that ads pass human review without being obviously AI-written.” That’s not a nice-to-have any more — the biggest four consultancies in the world just proved it’s the whole game. When the CMO agent build starts, this discipline gets promoted from a MACA quirk to the house style.

Three converging signals say “lock your token spend now”

This week alone: (1) Oracle’s $165B AI supercampus hit with multibillion-dollar cost surprises, (2) SpaceX confirmed the unpermitted xAI Memphis turbines stay until mid-2027 while building a new 1.2GW gas plant, (3) Hillsboro OR imposed a four-month moratorium on new data centers with permanent restrictions on the table. Layer that on top of Sonnet 5 intro pricing dying Aug 31 (29 days out, from yesterday’s brief) and the direction of travel is clear: frontier inference will get more expensive before it gets cheaper, permit friction is now real, and the labs will pass costs through. Worth 30 minutes to (a) run a fresh MACA per-ad unit-economics pass through MetaAdCreatorApp/api/lib/costs.ts and public/cost-dashboard.html, (b) check whether Ben’s daily-scan cost budget still holds at post-Aug-31 Sonnet pricing, © decide if any Trove/CourseBuilds template work should be Batch API’d through August to lock the intro rate.

“The Financial Times just caught PwC Middle East publishing AI-generated reports with fabricated citations and footnotes pointing at ChatGPT — and EY and KPMG have quietly retracted the same kind of thing. Big Four consulting just publicly failed the ‘must pass human review’ bar in the same week I’m building an AI training program pitched exactly around not doing that.”

Use case: Aria pitch prep with Michael Zaicek, RT AI role interview if it comes up (R53597), or any AI-pro conversation this week. Signals: you’re not selling AI hype, you’re selling AI discipline — which is currently the scarce good.

The Rundown just launched a Community AI Workflow Hub — front-of-newsletter feature for one submission a day

The Rundown (Rowan, 2M+ readers) shipped app.therundown.ai/community yesterday. Every day one reader’s submission gets featured in the newsletter with a direct LinkedIn link. Rowan explicitly says “this might fill your DMs — questions about the build, recruiters, job offers.” The pitch to you: you have three under-shared workflows that would eat this feature alive — (1) Ben’s PaperClip CFO-as-agent registration for autonomous UBX bookkeeping, (2) MACA’s 14-agent 4-wave ad pipeline with per-run cost tracking dashboard, (3) AI Edge itself (this brief, primary-source-first, wingman voice). Verdict: queue, don’t drop. You’re building Fillarup and about to start the UBX sale — no free cycles this week. But this is a straight LinkedIn/hiring-manager exposure play with almost no cost, and it aligns with the R53597 “show, don’t tell” strategy. Slot 30 min next week to submit MACA — it’s the most photogenic workflow and it aligns with the CMO Agent Build narrative if a founder DMs you about it. If the R53597 process goes quiet or the RT leaver decision accelerates, promote this to Tier 1.

• [The Information — “Anthropic’s multibillion-dollar IPO prep, Microsoft’s AI bug finder and Apple’s chip hunt”]: weekly stories-we-broke-first recap, every item covered in yesterday’s brief.
• [The Information — “China’s chip breakthrough, Claude Code’s lead and Stripe’s AI deal hunt”]: three-story digest, all three items already covered (China DUV 2026-08-01, Claude Code Reigns 2026-07-29, Stripe/OpenRouter 2026-07-30).
• [The Information — “Garry Tan’s Political Dreams Require This 32-Year-Old’s ‘Grunt’ Force”]: YC CEO political fundraiser profile, not AI.
• [The Information — “Robinhood Now Makes More Revenue From Predictions Than From Stock Trades”]: fintech, not AI.
• [The Information — “Inside Enterprise Software Deals and the AI Data Center Race”]: Pro-tier subscription pitch; SpaceXAI Texas + Oracle cost surprises folded into Section 3.
• [bagelbots — “The Prompt That Helps You Make Better Decisions”]: prompt-template edition; the PwC and Flock stories inside it were surfaced above, the prompt itself is decision-strategist boilerplate.
• [bagelbots — Anthropic security tests item]: covered in 2026-08-01 brief.
• [bagelbots — Dutch booksellers academic book orders]: interesting training-data anecdote, no action.
• [bagelbots — SpaceX/xAI Memphis turbines]: folded into Section 3 compute-cost signal.
• [bagelbots — Hillsboro data center moratorium]: folded into Section 3 compute-cost signal.
• [bagelbots — U.S. school cellphone-ban support]: not AI-relevant.

• Sources scanned: 13 newsletter senders (9 gmail.com, 4 prevailpartners.com.au fallback)
• Items extracted: 7 emails opened, ~15 discrete news items surfaced across them
• Items surfaced: 5 (3 Tier 1 slots + 1 anxiety-flip + 1 first-mover queue) plus Section 3 compute-cost synthesis
• Items skipped: 11
• Read time: ~5 minutes