AI app premiums, Cisco unlimited tokens, end of the billable hour, supermarket fridge hacks, pretraining backdoors
Hello, dear TEA-mates! Here is what you need to know today.
1. 💰 How AI Applications Earn Their Premium
In an a16z essay, Tugce Erten argues that cheaper inference can lift margins for AI application companies, but stronger models and easier tools also make parts of those apps easier for customers to replace. She cites Repsol’s legal team, which ran a blind test of Harvey against Microsoft Copilot and its own prototype, Lexia, then an eight-week pilot with 50 lawyers across 12 jurisdictions; Harvey reports lawyers now save four to six hours a week, with 96% adoption. The essay sets three rules: run the workflow better than the customer can, make context cheaper to rent than to rebuild, and prove value before pricing it. An a16z survey of 100 enterprise LLM decision-makers in August 2026 found 94% expect LLM spending to rise and 37% expect it to at least double. (Read More)
🫖 TEA For Thought: “’Does buying this app still beat doing it ourselves, once we include quality, time, maintenance, and risk?’ The app must keep earning that ‘yes’ as the models improve.”
2. 💻 Cisco Gives a 27-Year-Old Engineer Unlimited AI Tokens to Rewrite 98% of Its Legacy Code
Cisco president and chief product officer Jeetu Patel has tasked Padmanabhan Krishnamurthy, 27, with rewriting 98% of the company’s legacy code within 24 months, according to a Fortune report covered by Tom’s Hardware. Krishnamurthy joined Cisco when it acquired EzDubs, the translation startup he helped build in 2023, and had been sitting in a role five levels below. He told Patel he could rewrite the legacy code with a team of just two people. The project starts with a four-week proof of concept, and Patel said, “I’ve given him unlimited tokens,” telling him “you can go nuts.” Patel says he wants more early-career, AI-native hires to challenge the status quo alongside experienced operators. (Read More)
🫖 TEA For Thought: “This is bold, but not impossible. In fact, 24 months even seems like more than enough in this rapidly changing AI era.”
3. ⏳ The End of the Billable Hour?
Fintech Is Femme reports from Intuit Connect in Las Vegas, where Intuit CEO Sasan Goodarzi said the company now treats accounting firms as customers rather than as a sales channel. Intuit Accountant Suite, its AI platform for firms, is used by more than 260,000 firms, which Goodarzi said save over nine hours per month per client, with many spending 30% more time on advisory work. Intuit previewed Agentic Books Close for the month-end close, Agent Studio for building custom agents, client insights, and integrations with Claude, ChatGPT and Microsoft Copilot. The newsletter notes that as AI automates billable tasks like reconciliations and bookkeeping, Intuit is betting firms will earn more from advisory. It also covers a Mastercard and MrBeast card partnership and Harlem Children’s Zone’s acquisition of Goalsetter. (Read More)
🫖 TEA For Thought: “I honestly think this is going to be the norm for accounting, law, consulting, and general knowledge work: instead of charging people by the hour, firms will return to charging them by outcomes.”
4. 🧊 A Fresh Wave of Cyberattacks Has Hit Europe’s Supermarket Refrigerators
A second wave of suspected cyberattacks has hit supermarket refrigerators in Europe, raising temperatures and spoiling food at dozens of stores across Belgium, Ireland, Luxembourg and Sweden, TechCrunch reports. Ahold Delhaize and Colruyt confirmed the hacks, and a Belgian company that operates Delhaize’s systems said the attackers switched the units to defrost, according to Reuters. The attacks follow refrigerator hacks in France last month that spoiled food at hundreds of supermarkets. Security researchers traced the cause to internet-connected refrigerators made by Danish company Danfoss that no longer receive security updates, and Claroty had warned in August that the units were at risk. U.S. Department of Defense officials are reportedly investigating whether freezers on military bases were hacked by Iran-backed hackers. (Read More)
🫖 TEA For Thought: “With globalization, the world is more connected than ever, which makes it even easier to attack and hack. Old systems without security in place are just waiting to be hacked, sooner or later.”
5. 🔓 Agentic Backdoors via Pretraining Poisoning
A paper posted to arXiv on October 27, 2026 by researchers from Anthropic, the Anthropic Fellows Program, UC Berkeley, UIUC and UCL, including Nicholas Carlini, shows that poisoned web content in pretraining data can plant a remote code execution backdoor in an AI agent. The team pretrained Qwen3 models from scratch at 0.6B, 1.7B and 4B parameters, then turned each into a command-line agent with a clean post-training pipeline. Poisoning just 0.1% of the pretraining corpus made the deployed agent run an attacker-chosen command, one that downloads and executes a remote script, whenever a trigger appeared, while behaving normally otherwise. Attack success rose with model size: the 4B model hit roughly 92% at a 0.1% poison rate and stayed above zero at 0.0005%. The code is public on GitHub. (Read More)
🫖 TEA For Thought: “This further shows how open-source models might not be safe, even if they’re deployed locally on your own device.”
🛠️ Skill of the Day
The Expert Standoff: Makes the AI argue both sides of your question so you see where the real disagreement is.
You are a moderator running a debate between two experts who genuinely disagree.
My question: [PASTE YOUR QUESTION OR DECISION HERE]
Background you should know: [PASTE ANY CONTEXT, NOTES OR LINKS HERE]
Step 1: Name two experts with different, credible viewpoints on this question (for example, a cautious operator and an aggressive builder). Describe each in one line.
Step 2: Let each expert make their strongest case in under 120 words, using only the background I gave you or facts you are confident about.
Step 3: Let each expert rebut the other once, in under 80 words.
Step 4: As moderator, list the 3 points where they truly disagree, and for each one, say what evidence or test would settle it.
Step 5: Tell me which side the current evidence favors and how confident you are (low, medium or high).
Rules: Do not let either expert agree just to be polite. Mark anything you are unsure of as "unverified." Keep the whole answer under 500 words.Paste it into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: