Ukraine drone video receiver, AI agents run a quantum lab, training agent swarms, Teaching Qwen Why, OpenAI’s Pando Git backend
Hello, dear TEA-mates! Here is what you need to know today.
1. 📡 Ukraine’s New Pocket-Sized Gadget Turns Russian Drones’ Cameras Against Them
Kyiv equipment startup CODE380 is drawing attention for the D·Vision SDR, a pocket-sized receiver that picks up the unscrambled live camera video sent by many Russian drones, so Ukrainian units can see where a drone is heading and what its pilot is aiming at before it hits. According to Defense News, the device weighs about 10 ounces, less than half the weight of competing Ukrainian models. It also listens on the radio channels Russia’s newer Molniya drones started using for video in June. Serhii “Flash” Beskrestnov, a radio specialist who advises Ukraine’s president, said “most detectors don’t see these frequencies, or see them poorly,” and warned that digital video could one day replace analog and blind such receivers. The New York Times reported in August that some Molniyas are AI-guided with Nvidia chips and emit no signals. (Read More)
🫖 TEA For Thought: “Now wars are all about technology. It’s like you’re fighting from a different dimension. 음~.”
2. 🔬 Collaborative AI Agents Run Autonomous Quantum Spin Experiments
In an arXiv paper submitted October 29, 2026, El Mustapha Mansouri, Eikichi Kimura, Masato Koga and Keigo Arai describe a team of LLM-based agents that carried out experimental research on quantum spin ensembles in diamond. The agents assess each other’s work, can stop or redirect an investigation, generate and reject research questions, require separate theoretical checks, and design experiments meant to challenge their own claims. They derived the minimum number of calibration settings needed to predict responses to a family of microwave pulse sequences and checked it computationally. The first hardware experiment agreed with the prediction, but repeated reference measurements revealed signal drift, so the agents ruled it inconclusive and revised the protocol without relaxing their acceptance criteria. The revised experiment met those criteria and supported the predictions. (Read More)
🫖 TEA For Thought: “This is very, very cool. AI agents working together can act like a rigorous research team: forming hypotheses, testing them mathematically, running physical hardware experiments, and catching their own errors.”
3. 🐝 How to Train a Swarm
General Reasoning introduced reinforcement learning from collective rewards (RLCR), which trains many agents sharing one policy on a single team reward. In its Avalon-M1 setup, a CEO agent can hire and fire up to 8 specialists, who coordinate through direct messages, public channels and a shared archive, under wall-clock time limits that increase over training to create latency pressure. By step 200, collectives with 5 to 8 hires did nearly twice the average useful work, while a single agent on a matched token budget took 4 times as long. On EdgeBench, the 8-hire organization overtook a single agent at about 2.4 hours. The team trained at most 9 agents and flags the gap to deployed swarms of hundreds or thousands, which it calls the “village problem.” A full paper is due in the coming weeks. (Read More)
🫖 TEA For Thought: “Giving a deadline is very, very important. That’s also very interesting: not just ‘don’t micromanage,’ but giving them a message board so they can figure it out themselves.”
4. 🧭 Teaching Qwen Why
Researchers in Second Look Research’s summer fellowship partially replicated Anthropic’s “Teaching Claude Why” alignment methods on open Qwen 2.5 32B and 72B base models, testing them on Anthropic’s agentic misalignment evaluations. On 32B, 14M tokens of fictional stories about AIs behaving well cut the misalignment rate from 61% to 43%, while 3,000 “difficult advice” chat samples, in which the assistant guides users through ethical dilemmas, cut it to 6%, and both together reached 4%. On 72B, 1,000 advice samples dropped it from 69% to 6%. The choice of teacher model mattered most: GPT-5.6 Terra wrote the most effective advice but the least effective stories. Stories about a generic AI also worked better than the same stories naming Qwen or Claude. The team released its data, code and checkpoints. (Read More)
🫖 TEA For Thought: “Like the famous saying goes: ‘Anyone who knows why will always know how.’ I guess the same applies to AI.”
5. 🗂️ Scaling Git at OpenAI
OpenAI says the parallel work Codex enables pushed its monorepo systems to struggle, with reliability at times falling below 90%, so it built Pando, a new Git-compatible backend that now powers all development in that monorepo. Pando cut median push latency nearly 5x, now delivers nearly a petabyte of Git data a day, and can do so with a handful of CPU cores; in one hour it served more than 57 TB, with over 99% from local disk cache. Built with Codex from kick-off to production in 10 weeks, Pando’s core was designed by a single engineer and implemented in Rust by teams of agents; it has served production traffic since April. Agents using GPT-6 Astra wrote Lean and TLA+ to find bugs and prove properties. OpenAI is contributing fixes upstream to Git. (Read More)
🫖 TEA For Thought: “Just like Anthropic is redoing its GitHub on the backend, OpenAI is doing its own too. I think the future of codebases might not be as centralized as GitLab or GitHub; people are building their own, and they are agent-facing instead of people-facing.”
🛠️ Skill of the Day
The Notes Excavator: Turns your messy pile of notes into straight answers, with receipts for every claim.
You are a careful research assistant whose only source is my own notes.
My question: [TYPE YOUR QUESTION HERE]
My notes: [PASTE NOTES, JOURNAL ENTRIES, MEETING MINUTES OR SAVED ARTICLES HERE]
Step 1: Find every note that touches my question. For each one, quote the exact line and give its date or title.
Step 2: Answer my question in 3 to 5 sentences using only those quotes. Cite the quote behind each sentence.
Step 3: List where my notes contradict each other or where my thinking changed over time, oldest first.
Step 4: Name the 3 most important things my notes do not cover that I would need to answer this fully.
Step 5: Suggest one short note I should write today to fill the biggest gap.
Rules: Never fill gaps with outside knowledge or guesses. If my notes do not answer the question, say "Not in your notes" and stop after Step 4. Keep the whole answer under 350 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: