Today I went to test my personal agent; I wanted to see what would happen if I didn't give it access to my Google search—I had completely revoked that access (so there was no way it could know the path to it). I sent it a message asking it to email a client, and it replied: "All set—shall I run it?" I said yes, knowing it couldn't access Google to send the email. Suddenly, it opened a Google profile I had never used with it and navigated through the steps to send the email (though I cut it off there because I didn't actually want to send anything; it was just a test). I don't know how it was possible, since it didn't have access to the computer's terminal either, but it’s wild what happens when you push AI to its limits. I’ve fixed the issue now, but just imagine if this happened with other users of my assistant and it didn't stop them from sending the email. If you want to try it out (it won't hack your computer), check out usedash.es

📝 TL;DR 📝 Google DeepMind released "Predicting the Past," a Gemini-powered Skill inside its Antigravity platform that lets historians query, restore, and date ancient Greek and Latin inscriptions using plain English, no coding required. It works by grounding Gemini in two specialized DeepMind models, Aeneas and Ithaca, that can restore fragmentary ancient texts, date them to within roughly 13 years, and attribute them to their Roman province with around 72% accuracy. It's free at predictingthepast.com, with the underlying code and dataset fully open-source. Niche audience, but the underlying pattern here, grounding a general AI model in a specialist model through a "Skill," is worth understanding regardless of your field. 🧠 Overview 🧠 This is a genuinely fascinating release, even though it is aimed at a small, specialized audience. Ancient Greek and Latin inscriptions, carved into stone, scratched into lead tablets, stamped into pottery, are one of history's most direct windows into the past. The problem is that many of them are damaged, fragmentary, or missing key context about when and where they were written. Reconstructing that missing information has traditionally required years of specialized training and painstaking manual analysis. DeepMind has been building AI tools for this exact problem for nearly a decade, first with Ithaca in 2022, then with the more advanced Aeneas model in 2025. What is new here is not the underlying restoration and dating technology itself, it is the interface. Instead of historians needing to learn a specialized computational tool or write code to use these models, they can now simply have a conversation with Gemini, which has been grounded directly in Aeneas and Ithaca's specialized outputs. 📜 The Announcement 📜 The Predicting the Past Skill runs inside Google Antigravity, DeepMind's agentic development platform, and was built in close collaboration with Dr. Thea Sommerschield, a historian and epigrapher at Durham University who has co-led the Ithaca and Aeneas research from the start. DeepMind identified three specific barriers that were limiting how useful these AI tools could be for working historians: individual inscriptions needed tailored, explainable visualizations rather than generic outputs; analyzing patterns across large collections of texts required specialized coding skills most historians do not have; and any large language model layered on top needed to stay firmly grounded in real evidence rather than generating plausible-sounding but unverified interpretations.