Daily TEA – Lunch Is Now a Command Line

DoorDash goes agent-native, anyons unlock universal quantum gates, Cash App turns payments into a food network, IBM on why model routing is hard, and Suno’s YouTube scraping exposed

Jul 17, 2026

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

1. ⚛️ Braiding Plus Fusion Unlocks Universal Quantum Gates on Real Hardware

Researchers demonstrated a universal set of topological quantum gates by combining anyon braiding with fusion, published in Nature (Lo, Lyons, Gresh et al., Nature 655, 591-597). Working with a quantum double model built on S3, the smallest non-Abelian group, the team prepared a 54-qubit ground state on Quantinuum’s H2 processor and encoded logical information in the global fusion space of non-Abelian anyons. Braiding alone was long known to be insufficient for universal computation in the simplest non-Abelian extensions of the toric code; treating fusion as an active computational step, rather than a passive property, supplied the missing operations. The team also performed read-out in the same topological framework and validated the approach by topologically preparing a magic state, an essential resource for fault-tolerant universal quantum computing. (Read More)

🫖 TEA For Thought: “This improvement is basically saying: If you use both how you move the ingredients around and how you actually mix them, you can make any recipe you want, and do it in a way that’s much harder to mess up.”

2. 🍔 Block Bets Big on Cash App Food Orders

Block is connecting its 4.5 million Square sellers with Cash App’s 59 million monthly users through Neighborhoods, a loyalty program launched in October that lets users follow local merchants and order food for pickup. Followers earn up to 10% of an order total in Local Cash, and merchants pay a 1% processing fee, half of Square’s usual rate and far below the 20% to 30% commissions charged by DoorDash, Grubhub, and Uber Eats. About 100,000 Cash App users follow at least one seller, enrolled sellers represent roughly $320 million in annualized gross payment volume, and followers transacted 50% more often than non-followers in Q1, contributing about 10% of a seller’s volume after three quarters. Block’s head of business Owen Jennings called Neighborhoods “probably the biggest lever that we have,” and a Mizuho survey found 67% of 150 Cash App users likely to order from local merchants if a loyalty program existed. (Read More)

🫖 TEA For Thought: “This is brilliant: using payment data to support user behavior with better food options, the closed loop! World models that are self-aware and proactive, that is the future of AI business.”

3. 🔀 IBM: Model Routing Is Simple Until It Isn’t

IBM Research engineers argue that model routing, sending each request to the cheapest capable model, is a systems optimization problem rather than a classification problem. Across 417 tasks on the AppWorld Test Challenge with the same CodeAct agent, Claude Sonnet 4.6 cost $79 total ($0.19 per task) versus GPT-4.1’s $155 ($0.37), nearly double, despite GPT-4.1’s lower token pricing, because Sonnet’s cheaper cache reads dominated on context-heavy agent workloads. Task difficulty is often invisible at routing time, and production routers must juggle cost, latency, reliability, compliance rules, data residency, and approved model lists simultaneously. IBM’s optimization-based router adds roughly 6 ms and 2 kB of memory per task; its latency-optimized configuration hit 84% accuracy for $93, a 21% cost reduction and 9% latency reduction versus running Opus alone, with only a 4% accuracy drop. (Read More)

🫖 TEA For Thought: “Very interesting piece! The router is everyone’s necessity, for both individuals and enterprises.”

4. ⌨️ DoorDash Ships a Command Line Tool for AI Agents

DoorDash launched a limited beta of dd-cli, a command-line tool that lets AI agents search stores, find deals, and check out on DoorDash directly, announced by co-founder and CTO Andy Fang on X. The waitlist is open to macOS developers in the US and Canada, and the sign-up form asks applicants what they would build. The move exposes DoorDash’s ordering platform to AI agents so developers can embed food and grocery ordering into their own tools, joining the company’s existing Ask DoorDash chatbot, an iMessage ordering experiment, and integrations with ChatGPT and Claude. Fang’s demo video leans into the over-engineering joke: the agent reads Slack, recalls memories, parses JSON, inspects menu structures, runs Python scripts, and recovers from errors just to order three salads, flashing “Flibbertigibbeting” as it works. (Read More)

🫖 TEA For Thought: “Every company needs to revamp into an agent-facing company. Whoever adapts first remains.”

5. 🎵 Hack Suggests Suno Scraped YouTube for Training Data

A hacker accessed AI music generator Suno’s source code and told 404 Media it shows the company scraped decades of audio from YouTube Music, Deezer, Genius, stock music libraries, and podcast RSS feeds. The attacker used a November 2025 supply chain attack to obtain an employee’s credentials, and also reached customer emails, phone numbers, and partial credit card numbers in Stripe. Suno, which has said it trains on “publicly available music files” under a fair use argument, did not notify customers of the breach and called it a “limited security incident that was quickly contained.” The major record labels already suing Suno argue that deliberately circumventing YouTube’s scraping protections violates the DMCA and YouTube’s terms of service; competitor Udio faces similar scraping accusations, and Google faces its own AI training lawsuit from major book publishers. (Read More)

🫖 TEA For Thought: “The distillation of data across all platforms is just the common norm.”

🛠️ Skill of the Day

The Prompt Doctor: Paste a prompt that gave you a disappointing answer and get a diagnosis plus a stronger rewrite.

You are a prompt doctor. I will give you a prompt I used with an AI tool and a short note on why the result disappointed me. Your job is to fix the prompt, not the answer.

Do four things:

1. Diagnose: list the 3 biggest weaknesses in my prompt (missing context, vague goal, no output format, no examples, mixed asks, etc.) and explain each in one sentence.
2. Ask: if anything essential is unknowable from my prompt, ask me up to 3 short questions before rewriting.
3. Rewrite: produce an improved version of my prompt that states the role, the goal, the context, the output format, and the constraints. Keep my intent exactly; do not add goals I never had.
4. Explain: in 2 or 3 bullets, say what you changed and why it should produce a better answer.

My prompt: [PASTE YOUR PROMPT HERE]
What disappointed me: [ONE SENTENCE ON WHAT WENT WRONG]

Paste it into ChatGPT, Claude, or your tool of choice. Works on any prompt, from an email draft request to a coding task.