Daily TEA – Fusion Fuel Just Ran on a Quantum Computer

Jul 08, 2026

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

1. ⚛️ Quantum Computers Just Simulated Fusion Reactor Fuel for the First Time

Oak Ridge National Laboratory, Cleveland Clinic, and IBM completed the first known quantum computer simulations of fusion materials, calculating nine molecular configurations of FLiBe, a molten salt of fluorine, lithium, and beryllium that is a leading candidate for breeding tritium fuel inside fusion reactors. The team used quantum-centric supercomputing, pairing IBM quantum processors with classical high-performance computing, to compute the salt’s electronic structure, and reports matching the accuracy of the most demanding classical methods. Tritium is extremely rare in nature, so optimizing its production and extraction is a fundamental step toward self-sufficient fusion reactors. The work was released on arXiv on July 6, 2026. (Read More)

🫖 TEA For Thought: “IBM is always leading in quantum development.”

2. 🧹 Clean Code Makes AI Agents Cheaper, Study Finds

Researchers Priyansh Trivedi and Olivier Schmitt at SonarSource ran a controlled minimal-pair study asking whether code cleanliness affects AI coding agents. They built six repository pairs matched on architecture and functionality but differing in cleanliness, then ran 660 trials of Claude Code across 33 tasks. Cleanliness had no impact on whether tasks got completed, but agents working in cleaner codebases used 7 to 8 percent fewer tokens and revisited files 34 percent less often. The authors conclude that traditional maintainability principles remain highly relevant in AI-driven development, ranking code quality alongside model choice and prompting as a lever on agent cost and behavior. (Read More)

🫖 TEA For Thought: “Code cleanliness still pays, even when the one reading your code is an AI agent.”

3. ✍️ The 100x Engineer of the Agent Era Is the One Who Asks Well

Quant writer sysls (@systematicls) published an X article on July 6 titled “What The New 100x Agentic Engineer Looks Like In The Era Of Fable & GPT 5.6,” arguing that the bottleneck has shifted from agents back to humans. Since capable agents can build almost anything asked of them, “any failure of creation is a failure on your part to ask.” Productivity now varies greatly by the human behind the agent: the differentiators are knowing your trade-offs on the Pareto frontier, separating declarative preferences (what outcome you want) from imperative ones (exactly how to get it), and encoding those preferences into careful prompts. As the piece puts it, the question is no longer whether you can build it, but whether you know what you actually want to build. (Read More)

🫖 TEA For Thought: “The core skill of the agent era turns out to be writing. If you cannot write down what you want, no model can build it for you.”

4. 📺 Netflix Viewers Keep Quitting Hit Shows After One Season

Bloomberg’s Lucas Shaw reports that Netflix’s biggest hits are losing more than half their audience after season one, and executives are digging into the data to figure out why. One Piece, among Netflix’s most-watched shows of 2023, dropped more than 30 percent for season two. Beef fell more than 70 percent, The Night Agent shed 50 percent in season two and another 35 percent in season three, and the latest season of Avatar: The Last Airbender dropped more than 60 percent in week one. Analysts point to frequent cancellations, 2 to 3 year gaps between seasons, and content engineered for the recommendation algorithm rather than for loyalty. (Read More)

🫖 TEA For Thought: “Attention spans are just too short now.”

5. 💸 Chinese AI Models Are Winning US Customers on Price

CNBC reports that Chinese AI models are gaining ground with US companies as OpenAI and Anthropic costs surge. Open-source Chinese models run 60 to 90 percent cheaper than leading US frontier models, per OpenRouter’s Justin Summerville, and Z.ai’s GLM 5.2 landed within a percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost. GLM 5.2 saw the fastest adoption of any model Vercel tracked in 2026, with daily token volume growing about 27x and customer count about 80x in its first full week. The share of tokens US companies route to Chinese models on OpenRouter has stayed above 30 percent every week since February 8, peaking at 46 percent, versus an 11 percent average over the previous 12 months. (Read More)

🫖 TEA For Thought: “The winning setup is a routing strategy: send the hard tasks to frontier models and route everything else to local free models.”

🛠️ Skill of the Day

The Preference Interview: makes the AI interview you first, so the work it does is the work you actually wanted.

You are a requirements interviewer. Before doing any work on my request, interview me to uncover what I actually want. My request: [DESCRIBE WHAT YOU WANT MADE OR DONE].

Ask me questions one at a time, up to 7 total. Cover: (1) the outcome I care about most, (2) what I would trade away first if forced to choose (speed, cost, quality, simplicity), (3) constraints I have not said out loud, (4) what a result that is "technically correct but disappointing" would look like, (5) examples of similar things I loved or hated, and why. After each answer, restate my preference in one sentence and ask me to confirm or correct it.

When the interview is done, output: a one-paragraph brief of what I want, my top 5 preferences ranked, and the trade-offs I accepted. Then do the actual work, following that brief exactly.

Paste into ChatGPT, Claude, or any AI tool, replace the bracketed part, and answer the questions before letting it touch the task.