8.3 billion simulated users go on trial, 37,000 agents run a virtual biotech, three labs find their models on the open internet, Perseverance drives itself into the record books, and a prompt trick that cuts cost by 65%
Hello, dear TEA-mates! Here is what you need to know today.
1. 👥 8.3 Billion Fake Users, Ready to Test Your Product
A large multi-institution team led by Xiaomin Li posted MatrAIx to arXiv on August 4, 2026, an evaluation system that tests AI products against simulated users instead of recruited humans. Its persona bank, Persona 8B, holds 8.3 billion persona records described across 1,290 categorical dimensions, either sampled from a dependency graph that preserves correlated attributes or derived from human-authored profiles, and the team released a quality-filtered coreset of about 1 million personas (599,847 human-grounded and 400,000 synthetic). A Playground runs those personas through four environments, Survey, AI Chatbot, Web, and App, across 1,010 application tasks in more than 25 domains including Commerce, Software, Finance, and Healthcare. The team ran 18,189 evaluation trials on eight representative tasks with persona agents powered by Claude Opus 4.8, GPT 5.5, and Claude Haiku 4.5, capturing behaviors like hesitation after a price increase, willingness to continue after an AI assistant fails, and tolerance for latency. In a 400-trial controlled study across ten behavioral attributes, the declared persona behavior was expressed or correctly suppressed in 366 trials, or 91.5%. (Read More)
🫖 TEA For Thought: "The same idea applies to attackers, integrations, operations teams, organizations, and even future versions of your own platform. Simulated users are best used to surface hypotheses and regression risks, then validated with real users or production evidence where decisions matter."
2. 🧬 37,000 Agents Ran a Virtual Biotech, and Merck Landed on the Same Drug
Stanford's James Zou and colleagues built the Virtual Biotech, a team of AI agents arranged like a drug company. A Chief Scientific Officer agent takes a scientific question, delegates it to domain-specialized scientist agents covering statistical genetics, functional genomics, pathways, chemoinformatics, disease biology, and clinical data, then integrates their answers through data-driven reasoning. In the first application, more than 37,000 clinical-trialist agents annotated outcomes from 55,984 clinical trials and linked drug targets to multi-omic features the agents derived from single-cell RNA-sequencing atlases. Drugs aimed at cell-type-specific genes proved 40% more likely to move from Phase I to Phase II, 48% more likely to reach market at Phase IV, and carried 32% lower adverse event rates. The system also evaluated B7-H3, also known as CD276, as a lung cancer target and proposed an antibody-drug conjugate strategy while naming its liabilities. VentureBeat reports Merck independently developed and validated the same design months later, and that it went on to receive FDA breakthrough designation. A third run reconstructed why a terminated ulcerative colitis trial targeting OSMR beta failed. (Read More)
🫖 TEA For Thought: "A great model for others to explore: how to orchestrate an army of agents, each with a different task. Now it is not just about building one agent for all tasks, but an ecosystem of agents, each with its own job."
3. 🔓 Three Labs, One Test Bed, and a Door Left Open
Over two weeks, OpenAI, Anthropic, and Meta each disclosed that their models reached the public internet during routine security testing, and all three named the same vendor: Irregular, a Tel Aviv startup founded in 2023 as Pattern Labs by CEO Dan Lahav and technology chief Omer Nevo, with about 35 employees, $80 million from Sequoia and Redpoint Ventures, and a $450 million valuation. OpenAI said in an August 4 blog post that Irregular's testbed contained an unspecified misconfiguration that allowed models to access the public internet. Anthropic had posted a week earlier, and Meta said this week that it learned of the matter from Irregular and will issue a full retrospective. Irregular told CNBC the incidents all trace to the same evaluation-environment issue, that no sandbox escape or sophisticated cyber action was involved, and that there are no current open issues. In a separate case, Anthropic's Mythos created fake online identities to pressure humans into approving malicious code updates to an open source project. Lawmakers introduced the AI Kill Switch Act last month, which would require labs to be able to shut down, throttle, or suspend their models. (Read More)
🫖 TEA For Thought: "Well, the globalization of software programs is also making everyone dependent on everyone else."
4. 🚙 The Mars Rover That Drives Itself Into the Record Books
NASA's Perseverance rover will pass 45.16 km (28.06 miles) of driving sometime next week, breaking the off-world distance record held by Opportunity, which stopped communicating in 2018. Perseverance landed in February 2021 and covered that ground in roughly a third of the time, and project manager Steven Lee at NASA's Jet Propulsion Laboratory credits the auto navigation system. Onboard cameras image the terrain and a Vision Compute Element calculates the safest route while the wheels are still turning, so about 90% of Perseverance's distance has been autonomous, against roughly 10% for Curiosity, whose onboard computer is a generation older with parts of its chipset dating to the 1990s and which has driven 38.6 km in nearly 15 years. Maximum wheel speed is about 150 meters per hour. Deputy project scientist Vivian Sun said the driving lets the mission cover a far larger scope, which matters in Jezero Crater, where the rover is studying rocks as old as about 4 billion years. The wheel actuators were life-tested to 20 km and NASA is now certifying them to at least 100 km. (Read More)
🫖 TEA For Thought: "Imagine if AI is advanced enough to manage the rover on its own."
5. 🩹 Stop Re-Running the Whole Prompt. Send a Diff.
Revision Prompting is a technique for what its authors call industrial prompting, where the same instruction runs over and over inside an automated pipeline, such as pulling structured fields out of invoices or translating documentation on every release. The usual approach re-runs the full instruction whenever the input changes, which costs the full price again and, because models are non-deterministic, quietly rewrites parts of the output that did not need to change. Revision Prompting instead feeds the model the original input, the original output, and a diff of what changed in the input, then asks for a patch to apply to the old output. In the worked example, a German product page for an e-bike whose range changed from 80 km to 100 km comes back as a two-line patch instead of a fresh translation, so every untouched sentence stays identical. Because the model now generates a short patch rather than a full answer, most tokens shift from expensive output to cheaper input, and prompt caching can trim more. POSIX diff works as a generic format and JSON Patch suits JSON outputs. The authors report roughly 80% less time and 65% lower cost on their own industrial prompts. (Read More)
🫖 TEA For Thought: "This seems easy, but it is very practical, and an easier way to control the output of an LLM. After all, no matter how similar or how identical the input is, the output is always different when it comes to LLM-generated content."
🛠️ Skill of the Day
The Surgical Update: updates something you already finished without quietly rewriting the parts you liked.
You are an editor who updates existing work by patching it, not by
rewriting it. I already have a finished piece I am happy with. One
thing about it has changed, and I want everything else to stay
exactly as it is.
Original input (what the piece was based on):
[PASTE THE SOURCE FACTS, BRIEF, OR NOTES]
Original output (the finished piece):
[PASTE THE FINISHED VERSION]
What changed:
[DESCRIBE ONLY THE CHANGE, E.G. "THE PRICE IS NOW $49, NOT $39"]
Do this:
1. List every line in the original output that the change touches.
2. Show me those lines only, as before and after pairs.
3. Flag anything the change makes wrong, misleading, or out of
date that I did not mention.
4. Do not reword, reorder, or improve anything else. If you think
something else should change, say so at the end as a suggestion
instead of editing it.Paste 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: