Hello, dear TEA-mates! Here is what you need to know today.
1. 🖥️ Reliance Jio Opens Its Cloud PC To All Of India
Mukesh Ambani's Reliance Jio opened JioPC to any internet user in India on Wednesday, turning a service that launched in July 2025 for its own broadband subscribers into a stand-alone subscription anyone can buy regardless of their telecom provider. The service streams a virtual PC from Jio's cloud onto the computer you already own, with plans offering up to eight virtual CPUs, 16GB of RAM, and 1TB of storage. Jio says a machine as old as eight years can reach the computing power modern AI applications need without new hardware. Pricing starts at ₹1,000 (about $11) for two months, with 12-month plans at ₹4,000 (about $42) for 8GB and 500GB, or ₹5,000 (about $53) for the 16GB and 1TB tier. The timing tracks a widening replacement cycle: IDC counted more than 65 million PCs in use in India in 2025, and consumers now keep a machine five to six years instead of four to five in 2022. Prabhu Ram of CyberMedia Research said the model could create a parallel, price-sensitive tier for households and students "for whom buying a new AI-capable PC was never really on the table to begin with," though the published specs mention no dedicated AI accelerators, and the whole thing stops working when the connection does. (Read More)
🫖 TEA For Thought: "Essentially it is renting a VPS, just rebranded in a way that utilizes the old PC at home that would otherwise be useless. Pretty brilliant business idea, actually."
2. 🏫 New York City Bans Generative AI For Every Student Through Eighth Grade
New York City public schools will block elementary and middle school students from using generative AI tools for the coming school year, Mayor Zohran Mamdani announced Wednesday. The one-year moratorium covers "all software that uses student-facing generative AI" through eighth grade, and the mayor's office called it the most expansive prohibition in the country. High schoolers get the opposite treatment: twice-yearly AI literacy classes for everyone, plus pilot programs letting some students use a handful of approved platforms under an educator's supervision. Companion chatbots are banned across all grades. The city also recommended daily screen time caps of 30 minutes for grades three through five and 45 minutes for grades six through eight, while teachers may still use AI for lesson planning and operational work. Los Angeles Unified said the same day that its own block on generative AI stays in place across all grades while it writes a policy. At least 37 states now publish AI guidance for schools, with no shared definition of what AI literacy even means. The teachers' union praised the screen time limits but said the policy "leaves many questions unanswered," and a children's advocacy group argued a single year is not enough time to evaluate anything. (Read More)
🫖 TEA For Thought: "Meanwhile, China is trying to teach kids robotics starting in kindergarten. In my opinion, kids of this generation should learn to ride the wave instead of avoiding the wave."
3. 🌊 Vercel's Fluid Assembles A Machine Around Whatever You Are Running
Vercel published the architecture behind Fluid, the compute layer that now runs its builds, sandboxes, and functions on one system rather than three. The pitch is that every workload has a different shape: a build is compute-bound and wants a heavy machine, a function is IO-bound and wants a small one that starts the instant a request lands, and a sandbox takes whatever configuration the work calls for. Fluid takes the workload and assembles the machine to fit, swapping configuration on the fly. Three pieces do the work. Hive is the hardware layer that provisions an isolated VM in milliseconds, usually one already warm. Fluid images let you push your own image to Vercel's container registry, which gets converted into a snapshot format that resumes instead of booting. Vercel Drives makes storage portable, so your files travel with the workload instead of being stranded on one machine's disk. The numbers behind it: over 15 million builds a day, 25 million sandboxes a week, and a trillion requests a month. Vercel frames the whole thing around agents, which run untrusted code, bring their own tools, and spin up and tear down constantly, so their state has to outlive the compute. "In Fluid you describe the work, and the machine forms around it." (Read More)
🫖 TEA For Thought: "Yesterday we had Runway designing the interactive landing page for each customer. Everything is customized, and today we have Vercel, where all the deployment could be fluid and adaptive. I guess personalization is the future. It is not only for users from all journeys, from beginning to the end, but also for agents end to end, from code writing all the way to deployment."
4. 🪞 Researchers Train Models To Grade Themselves, And They Get Honest About Uncertainty
A new paper argues that large language models have a metacognition problem: they hallucinate with high confidence, fail to recognize where their knowledge ends, and misrepresent how uncertain they actually are. The authors' bet is that a model able to judge its own performance accurately is better positioned to improve it. They introduce two mechanisms. Reinforcement learning with metacognitive feedback, or RLMF, refines how completions get ranked during preference optimization based on the quality of the model's self-judgments. Metacognitive data selection uses those same self-judgments to find high-value training examples, outperforming naive active learning. They apply both to faithful calibration, the task of aligning what a model says about its confidence with what it actually knows, which the paper notes is difficult even for frontier models. The approach is two-stage: calibrate the self-reported confidence scores first, then map them into natural, context-adaptable language through targeted output editing. RLMF reached state-of-the-art faithful calibration across diverse tasks without losing accuracy, and surpassed standard reinforcement learning by up to 63% while improving the model's ability to state its own limits. (Read More)
🫖 TEA For Thought: "People say the fear of God is the beginning of wisdom. Knowing what you do not know is the beginning of wisdom as well. Humbling is the beginning of wisdom. It is really a humbling experience to learn about human intelligence while building artificial intelligence."
5. ⚛️ Agents Are Now Writing The Quantum Problems They Get Tested On
UCSB's AI group and its Quantum CoDesign Lab released Quantum's Infinite Game, a system where agents do not answer a fixed question bank but go find the open problems themselves. You give it one word of direction, a topic or a bottleneck, and it searches the surrounding literature for unresolved directions that could become verifiable research problems. Accepted problems get formalized into runnable environments with explicit scorers, reference solutions, and trivial baselines, all bounded to 28 qubits or fewer, under three hours, on a single workstation, with no quantum hardware anywhere. Five worked examples are now browsable end to end, and the honest part is the scoreboard. On a quantum metrology task asking agents to design noise-robust probe states, the best of three blind attempts got 37% of the way from the trivial baseline to the expert reference. On a low-depth Hamiltonian simulation task, 42%, the best any agent has managed. On the most novel problem in the set, learning where to spend Trotter steps across a spin chain, the best attempt scored worse than doing nothing at all. Because the problems are genuinely open when they are created, memorizing the benchmark from training data buys you nothing. (Read More)
🫖 TEA For Thought: "Curious to see if there is new scientific discovery."
🛠️ Skill of the Day
The Knowledge Boundary Map: before an AI answers your question, it maps the edge of what it actually knows, so you find out which parts of the answer to trust.
You are about to answer my question. Before you answer it, map the edge of what you actually know about it.
MY QUESTION:
[PASTE YOUR QUESTION HERE]
STEP 1. Do not answer yet. Break the question into the 5 to 8 specific things you would need to know to answer it well. List them.
STEP 2. Sort every one of them into exactly one bucket:
- SOLID: I know this well and would bet on it. Say how you know it.
- SHAKY: I have a general sense but the specifics could be wrong or out of date. Say what specifically could be wrong.
- OUTSIDE MY KNOWLEDGE: I do not know this, it is past my training data, or it depends on your private situation. Say which of those three.
STEP 3. For every SHAKY and OUTSIDE item, name the one thing that would resolve it: a document I could send you, a source I should check, or a question only I can answer.
STEP 4. Now answer the question using only the SOLID items, and mark every sentence that leans on a SHAKY one. If the SOLID items are not enough to answer at all, say that instead of answering.
Do not soften this. An answer that hides which parts are guesses is worse than no answer.Paste it into ChatGPT, Claude, or your tool of choice. The map is often more useful than the answer, because it tells you exactly what to go find out.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more: