Inteligencia Artificial

A group of researchers published Stolen Thoughts, a technique that decodes the encrypted reasoning of Claude, GPT, and Gemini by reinjecting the signed block into a weaker, jailbroken model. After analyzing 6,708 public agent traces on GitHub and Hugging Face, they found 704 real secrets, including 62 API keys and 33 passwords.

AMD confirmed an alliance with a science ministry to develop an open-weight artificial intelligence ecosystem. The company already publishes its own open models, OLMo and Instella, trained on its Instinct MI300X GPUs with the ROCm stack.

Docker unveiled Sandboxes, a tool that runs code agents like Claude Code and Copilot CLI inside disposable microVMs isolated from the host. Each sandbox mounts only the project’s workspace, letting agents work autonomously without constant permission prompts.

An engineer built ChipTycoon, an interactive low-poly simulation that teaches how a chip is manufactured from start to finish. The trick: he used Claude Code to research the topic, verify the data, and generate the simulation, instead of reading scattered documentation.

OpenChamber is an open source, agentic development environment that lets you run the same task on up to five AI models at once, merging or comparing results. It runs on the OpenCode SDK and works from desktop, browser, phone, or directly from your code editor, without sending your code to third-party servers.

The UN created its first AI scientific panel within the Global Digital Compact, modeled on the IPCC. Its first warning is direct: the window to regulate artificial intelligence is closing while models advance faster than laws.

Google DeepMind published its WeatherNext results in Nature, an AI model that predicts the trajectory, intensity, and wind structure of tropical cyclones with a full extra day of lead time. The company has already released the code for WeatherNext 2 and WeatherNext Cyclones to the scientific community.

AMD acquired Taalas, a Toronto startup that etches AI model weights directly into the silicon of its chips. The first prototype served Llama 3.1 8B at 16,960 tokens per second, and AMD plans to combine it with its Instinct GPUs in large-scale inference deployments.

Researchers at SEOULTECH built an artificial intelligence framework for long-term bridge monitoring, capable of distinguishing real structural damage from normal environmental variability. The approach builds on decades of research in structural health monitoring, from the historic Z24 bridge in Switzerland to the 2007 I-35W collapse.

AirLLM, the open source library that lets you run massive models on little memory, now supports Kimi K3: the open model with 2.8 trillion parameters that now runs on a single GPU with less than 4 GB. The per-expert streaming technique avoids loading the full model into memory.