Nvidia AI module found inside Russia's new cruise missile
Ukrainian military intelligence says it pulled an Nvidia Jetson Orin — a module sold for robotics and self-driving development — out of Russia's newest cruise missile, Google researchers argue that frontier models' factual errors are mostly "lost keys" rather than "empty shelves," and Tsinghua open-sourced a code-repair model that tops the open-source leaderboard.
Ukraine's Defense Intelligence (HUR) says it found an Nvidia Jetson Orin module inside Russia's new S-71 "Monochrome" cruise missile, suggesting the weapon uses AI for its final targeting phase. The finding came in a War & Sanctions portal release cataloging 35 foreign-made components recovered from Russian weapons, alongside a passive radar seeker Russia has begun installing on Geran-2 drones, a Chinese camera in the jet-powered Geran-4, and parts of the Kinzhal's active homing seeker. HUR describes the S-71M as carrying a 250 kg warhead at 500–600 km/h over a 350–400 km range, launched from the Su-57, with the Jetson module handling terminal guidance — locking onto a target visually through machine vision rather than waiting on a human operator. HUR notes similar computer-vision guidance has turned up before in Russian weapons, including the V2U strike drone and the Shahed-236.
The significance is less the hardware than the pattern HUR is trying to break. Two years of cataloging on War & Sanctions shows Russia still cannot replace foreign high-tech components, and each identified part becomes ammunition for tightening export controls on the gray-market networks that route Western and Asian tech into Russian factories. There is a sharp irony here: a computing module Nvidia sells openly for hobbyist robots doing terminal guidance in a cruise missile — and the discovery lands as AI hardware itself becomes a contested export category.
Google researchers argue that frontier models' factual errors are mostly "lost keys," not "empty shelves": Gemini-3 and GPT-5 encode 95–98% of the facts they're tested on, yet fail to directly recall 26–34% of them. In "Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality," the team introduces knowledge profiling, a framework that classifies each fact by whether it is encoded and then by how accessible it is, along with WikiProfile, a benchmark of 2,150 Wikipedia-derived facts each probed by ten tasks across 13 models — roughly 4.5 million responses. Two familiar puzzles get reframed: rare facts are encoded almost as well as popular ones but are much harder to recall, and the so-called reversal curse largely disappears when models only have to recognize the correct answer rather than generate it. Since thinking recovers 40–65% of encoded-but-inaccessible facts, the authors argue the next factuality gains will come from better utilization of stored knowledge — post-training and inference-time reasoning — rather than from simply scaling models further.
Tsinghua's Shenzhen campus open-sourced VeriLoop Coder-E1, a 27-billion-parameter code-repair model that ranks first among open models at or under 32B parameters on the major repository-level benchmarks. Built by Liu Houde's Intelligent Robotics Lab at Tsinghua SIGS on a frozen Qwen3.6-27B base with a detachable adapter, it scores 85.20 on SWE-bench Verified, 62.38 on SWE-bench Pro, and 76.40 on Terminal-Bench 2.0. The distinctive part is the "Self-Harness" loop: the model recompiles its own code, folds test receipts and tool feedback into structured evidence, and iterates until the evidence holds — a scheme the team says is inspired by Karl Popper's falsifiability. Weights and adapters are open while the control stack itself stays proprietary, continuing the open-weights-but-closed-loop pattern Chinese labs keep shipping.
What to watch: whether HUR's catalog pushes Western regulators toward stricter export rules on edge-AI hardware, and whether Google's recall result shifts how labs spend on post-training.
An Nvidia module sold for hobbyist robots doing terminal guidance in a cruise missile — should edge-AI hardware face stricter export controls? Tell us in the comments.
