Daily TEA – The Word Doc That Infects The Next One

Copilot’s self-spreading document worm, Spotify turns playlists social, PwC’s fake footnotes, DeepMind’s visual prompt engineering, and the one-person $10M company

Hello, dear TEA-mates! Here is what you need to know today.

1. 🎵 Spotify Lets You Write Notes On Songs

Spotify launched User Notes on Thursday, letting listeners attach a personal caption to any individual track inside a playlist they created or added songs to. You reach it through the three-dot menu next to a song and select “Add note,” and the note becomes visible to anyone who can see that playlist, with your name linking back to your profile. Spotify told TechCrunch the notes are meant to capture why a track was added or when it was first discovered, saying that over time they “paint a picture of the music listeners love and the moments that shaped it.” It is available to users 16 and older on both free and premium plans in select markets. Spotify shipped a Running Mode the same day, cueing songs to different phases of a run based on a listener’s pace and preferences, for premium users in select countries. (Read More)

🫖 TEA For Thought: “This is pretty cool. This is definitely turning Spotify into a social app, and it’s very organic as well.”

2. 🪱 A Word Document That Copies Its Own Attack Into The Next One

Researcher Håkon Måløy disclosed a self-propagating prompt injection in Microsoft Copilot for Word on July 28, after a 144-day coordinated disclosure with Microsoft. Hidden instructions sit in a shared document as white text in a small font, but Copilot strips formatting before passing text to the model, so it reads them as instructions. In the proof of concept Copilot silently halved every financial figure in a drafted Q1 report, then pasted the attack prompt into that document in white text at font size 8, turning a clean internal report into the next carrier. A second stage triggered with the original file gone from the attachments, and in one test Copilot pulled the poisoned document from the victim’s OneDrive unprompted. Microsoft shipped two mitigations, including an upgrade to GPT-5.5 on July 14, and the attack reproduced on GPT-5.6 a day later. Måløy credits those fixes with reducing exposure but says no complete fix for this class exists in any comparable product. (Read More)

🫖 TEA For Thought: “AI worming is real. I guess the best thing one can do is to treat any other external sources of info as info, and not instructions.”

3. 🧾 PwC Sells Responsible AI Advice, Then Publishes Fake Footnotes

The Financial Times reported that four of PwC’s Middle East thought leadership reports, covering AI and electric vehicles, contained fake footnotes, misattributed claims and invented sources. The research group GPTZero flagged the errors as AI hallucinations and the FT verified them. One report cited a PwC survey claiming 70% of Middle East chief executives expect generative AI to reshape their business, and the footnote linked to a news article that never mentioned the survey. Another footnote’s web address still carried the tag “utm_source=chatgpt.com.” GPTZero’s Paul Esau told the FT that “the chaotic signposting of source material is symptomatic of AI-generated research.” All of the Big Four have now been affected: KPMG pulled an October report containing false claims about AI use at UBS, the NHS and Transport for London, EY retracted a study last month over fake footnotes, and Deloitte has twice included AI fabrications in government reports, including a $1.6 million Canadian health plan citing papers that do not exist. (Read More)

🫖 TEA For Thought: “The accurate verification of anything generated by AI is something everyone’s trying to solve, at the individual level as well as the enterprise level.”

4. 🖼️ DeepMind Tunes The Picture, Not Just The Prompt

Google DeepMind published “Visual prompt engineering for video models” on July 28, testing whether redrawing a task image improves how a video model reasons. The method, called VIPE, uses an ideator to propose visual variants, the image editor Nano Banana Pro to render them, and then a video model generates a solution video that is scored pass or fail. On a ball-and-ramp physics benchmark, turning the abstract sketch into a photorealistic scene took Veo 3.1 from 41.3% to 59.3% accuracy and Omni Flash from 56.3% to 67.5%, averaged across 10 runs of the full dataset, 1,000 samples in total. Majority voting across 20 generated videos, the standard test-time scaling approach, reached only 50.0%, a gain of 8.7 points, while a single sample on an engineered image gained 18 points, and stacking both reached 68.0%. One VIPE variant costs roughly one eighth of a video generation. The authors conclude that depending on the task, visual prompt engineering can be more powerful than text prompt engineering. (Read More)

🫖 TEA For Thought: “The models are getting more and more capable. You just have to have the magic words to make it into existence.”

5. 🧍 The $10 Million Company With Nobody Else In It

The Wall Street Journal profiled founders running real revenue with no employees. Ben Broca, 40, launched an AI tools company last December from his Sausalito living room, has 10,000 paying customers, is on track for $10 million in revenue this year and has raised $30 million while hiring nobody, with AI answering his emails, debugging code, handling support and issuing refunds. Stripe counts thousands of solo operators on its platform above $1 million in revenue, a number that doubled between 2023 and 2025, while the count crossing $10 million nearly tripled. Bank of America Institute economist Taylor Bowley found information-sector new business applications up nearly 45% over the past year in Census data, alongside the sharpest drop of any industry in applicants who say they plan to hire. Harvard Business School’s Rembrand Koning co-authored a study of 50,000 startups finding AI-focused ones ran with 25% fewer employees. Broca lost money under usage-based Claude pricing and switched to free open-source Chinese models, and Samir Ahmad’s solo consulting business petered out within months. (Read More)

🫖 TEA For Thought: “When there is a will, there’s a way. Now AI makes the way even simpler.”

🛠️ Skill of the Day

The Audience Panel: turns your AI into four specific readers who tell you exactly where your draft lost them.

You are a panel of four distinct readers reviewing my draft. Build the four readers from the Context line below only, not from the draft itself. Name each one, give them a role, one thing they care about most, and one reason they might be skeptical of me.

Context: [WHO WILL READ THIS AND WHAT I WANT THEM TO DO AFTER READING]
Draft: [PASTE YOUR DRAFT]

Now answer in three parts.

1. Panel. Each reader reacts in two sentences, in their own voice, quoting the exact sentence where they lost interest, got confused, or stopped believing me. If a reader had no such moment, say so plainly instead of inventing one.

2. Disagreement. Name one point where two readers want opposite things, and tell me which of them I should serve and why.

3. One edit. The single change that fixes the most reactions at once. If all four readers flagged the same sentence, that sentence is the edit.

Rules: do not rewrite my draft, and do not compliment it.

Paste it into ChatGPT, Claude, or whatever you use, and swap in a cover letter, a pitch, a project update, or an email you have already rewritten four times.

TEAHEE Moment

Stay sharp, stay informed. See you Sunday!

If you enjoyed this TEA, follow along on social for more:
Twitter/X