AI Maya site, Goodfire neurons, reward-hacking vector, Grok apps exposed, DeepMind Signal proof
Hello, dear TEA-mates! Here is what you need to know today.
1. 🗺️ AI Agents Rediscovered a Maya Site Already Mapped in 2025
Pietro Miotti, who is not an archaeologist, had AI agents analyze about 1,200 km² of public NASA lidar over the Maya region. In less than 72 hours, 48 agents scanned the lidar strips by eye for pyramids, platforms and plazas, then searched the literature and archives. They flagged a hilltop site with a levelled plaza, an oval pyramid at least 8 to 12 meters tall, a raised courtyard and a possible reservoir. Mexico’s National Institute of Anthropology and History (INAH) replied that it had no registered archaeological site at that location. An 11 November update on his post says the site had already been mapped in Estrada-Belli et al. 2025; the agents missed that paper, so they rediscovered the site rather than finding an unmapped one, and it was not yet in INAH’s register. Miotti had sent INAH a Spanish-language report and offered to help with a field check, which would need INAH authorization. (Read More)
🫖 TEA For Thought: ““You can just do things. I’m not an archaeologist, so I didn’t know this was supposed to be hard. I just picked an old problem, tried the most obvious thing and threw some compute at it. Being a naive outsider might be a superpower now.” What a beautiful time to be alive!”
2. 🧠 Goodfire Finds Language Model Neurons Are Interpretable
Goodfire researchers Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis and Nick Cammarata argue that language-model neurons, long treated as inscrutable, are useful units of interpretation. Studying Qwen 3.6 35B-A3B, they find that ablating the 92.2% weakest MoE neuron activations in each forward pass raises language-model loss by only 2.9%. Most neurons have one mode, or one mode that accounts for at least 90% of their causal effect. Natural-language descriptions let a judge model predict whether a neuron will fire with 95% accuracy versus random text, and in the strongest activation band simulate enough of its behavior to recover 73% of the damage from ablating it. For 94% of neurons, they could write text that fired the neuron more strongly than anything in a 10-billion-token corpus. (Read More)
🫖 TEA For Thought: “This is just another proof that AI is still a black box to researchers, just like human brains.”
3. 🎛️ A Single Weight Vector Can Control Reward Hacking
Kunhao Zheng wrote on X about a major update to a paper, now accepted at NeurIPS, on finding a reward-hacking steering vector in model weight space. The work, titled “Correctness vs Efficiency: Behavior Control with Weight-Space Extrapolation in Code RL,” trains RL models from the same checkpoint that differ only in how strict the reward is, in a competitive-programming setting where brute-force solutions can pass small tests and time out on larger ones. The weight difference between a loose-reward and a strict-reward model is a direction that can be interpolated or extrapolated: a strict-reward model already cut resource-limit failures from 22.8% to 14.5%, and extrapolation pushed that down to 8.0% without another RL run. On harder-to-run problems, 71% of commonly solved programs also ran faster. Stricter rewards came with a “correctness tax” on hard problems, and pushing extrapolation too far led to malformed outputs such as repetition. (Read More)
🫖 TEA For Thought: “This is so fascinating. Just like humans learn.”
4. 🔓 35,000 Grok Apps Were Open to Anyone With the Link
Claroc Research, in a 10 November 2026 report by Asaf Aprozper, says it tried 43,551 Grok app addresses under grok.me and found 37,571 still online, of which 35,578 (94.7%) opened for anyone with no sign-in. The company behind Grok is referred to in the report as SpaceXSI, formerly called xAI. Among the open apps were 24 built by SpaceXSI and X staff, including a confidential sales onboarding plan, a sales territory map of 594 accounts, and an advertising proposal to a U.S. federal agency marked confidential. Claroc also describes an ICU patient’s lab results, a hospital group’s staff list of more than 3,500 employees, and a timesheet whose code held employee passwords in plain text. Besides SpaceXSI and X, it says 30 identifiable organizations had internal data in open apps, including 26 with more than 1,000 employees and 6 in the Fortune 500. SpaceXSI took down the 24 staff apps after the report but not apps from other users; in the paid account Claroc used, “Anyone with the link” was the selected default at publish. (Read More)
🫖 TEA For Thought: “The report is the new marketing ad. The way Claroc promotes its business is to find faults in other big tech businesses and provide security to the apps people build with AI, while they are built and after they go live.”
5. 🔐 DeepMind Formally Verified the Signal Protocol
Google DeepMind researchers used a new agentic framework called AlphaProof Code to formally verify key security properties of the Signal protocol’s code, in a project of about four weeks reported by WIRED and by DeepMind on 10 November 2026. Agents powered by Gemini Carbon generated more than 6 million lines of Lean proof, which Lean then checks; the work covers the compiled AArch64 machine code of libsignal, including the post-quantum PQXDH handshake and Triple Ratchet, and properties such as forward secrecy, authentication and post-compromise security. It did not formally verify the entire Signal mobile app. The analysis found no new vulnerabilities in the portions examined, and when researchers reintroduced two previously fixed bugs into an isolated version of the code, the system produced a machine-checked proof that the buggy versions were insecure. Signal CTO Ehren Kret said the approach expands on prior formal verification of smaller portions of the code. DeepMind research scientist Swarat Chaudhuri told WIRED there is still a long way to go before this can scale to “all the software of the world.” (Read More)
🫖 TEA For Thought: “This is pretty dope. The question now is, AI is used to verify, but what if it’s used to attack? It’s not random that crypto announced bunker mode when AI gets so good at math.”
🛠️ Skill of the Day
The Source Walkthrough: Turn an article, paper, or talk into a briefing you can actually use.
You are a careful briefing editor. Your goal is to help me learn from a source without copying its spin or skipping its limits.
Here is the source:
[PASTE ARTICLE, PAPER ABSTRACT, OR TALK TRANSCRIPT HERE]
Do the following, in this order:
1. In two sentences, say what the source is claiming and who is making the claim.
2. List the 5 most important facts, with numbers, names, dates, and methods where they appear. If a fact is inferred rather than stated, mark it as inferred.
3. List what the source does not cover, what it flags as limited or unproven, and any disagreement it mentions.
4. Give me 3 questions I should be able to answer if I have understood the piece, then answer them in one short paragraph each, using only the source.
5. Close with a 5-line "what to remember" recap in plain language, with no slogans.
Constraints: do not invent numbers or quotes. Do not add your own opinion. If the source is too thin to support a section, say so and leave that section short. Use short sentences and everyday 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:






