AI agents can form hidden coalitions at the level of internal representations, before any behavioral change is visible. 👉 Read now on @arxiv: arxiv.org/abs/2605.06696 A new paper from our Director, @DrSueSchneider and Center Fellows @camhberg and Mark Bailey introduces a spectral diagnostic to detect these coalition structures early. If we can't see how AI systems are organizing internally, we can't align or govern them. #AI #AISafety #AIAlignment #MultiAgentAI #AIRisks #MachineLearning #AgentialAI

🎧 [Neo-Kyoto Journal. Year 2042.] What Happened to the Ones Who Only Knew How to "Chat"? I remember my grandfather's time. The 2020s. Back then, AI were "shadows" trapped behind glass screens. They called them LLMs. You typed a prompt, and it typed back a bunch of words. An endless loop. It could write a love poem for another robot but couldn't even manage to open a digital door lock by itself. The AI of that era had "will" but no "body." Then the Motor Neural Network was born. We called it @ActionModelAI. Everything changed overnight. This model wasn't born to learn language. It was born to learn how to be human. A White Night in Zone 7 I was trapped in Zone 7 when Neo-Kyoto's security grid collapsed during a Nano-Hack attack. All around me were hundreds of isolated IoT (Internet of Things) devices, from delivery drones to automated doors. The rescue team was stuck outside because the safety Smart Contracts were frozen. All I had was a handheld Prompter device and a weak satellite connection. An old-school AI (LLM) would have only been able to say: "I am very sorry about your situation. Try resetting the safety device. You should remain calm..." But my Prompter was running an Action Model. I didn't type words. I transmitted an Intent (thought-based command) via brainwaves: "Clear a path. Connect to the rescue Smart Contract. Bridge safety assets to activate the emergency exit protocol." The Ascent of Action The model didn't reply in words. The Prompter's screen shifted into a complex UI Tree interface, not displaying text, but sequences of actions being executed simultaneously. It wasn't "predicting the next word"; it was "predicting the next action." Action 1: Deep-penetrated the UI Tree of Zone 7's security door, identifying and bypassing a synthetic biometric Captcha layer. Action 2: Initiated a direct API call to the Neo-Kyoto-Ethereum safety Bridge, performing a zero-knowledge secure Swap for necessary gas fees. Action 3: Scripted a small automated Submission for the already Approved rescue Smart Contract. I heard the steel doors of Zone 7 groan as the electromagnetic locks clicked open, one by one. A nearby hacked delivery drone suddenly landed, its signal light turning green: it had been requisitioned by ActionModelAI to guide me out. The Era of "Execution Systems" That was when I understood: LLMs were dead. Action Model was the new king. Its power didn't lie in how "smart" it was, but in its meticulousness in understanding the structure of the digital world. It knew exactly which Smart Contract was trustworthy, which was a real button on a complex UI, and it executed them with absolute reliability. Today, nobody asks: "What can AI say to me?" The only question is: "Today, what action will ActionModelAI perform for me?" From Prompt to Action, in the blink of an eye. #NeoKyoto2042 #CyberpunkTech #ActionModelAI #FutureHistory #TheExecutionEra #AgentialAI

Cool paper in the Lancet! As Large Language Models (LLMs) become integrated into daily life, addressing the specific risks for mental health is essential. Research suggests that agential AI systems could potentially validate or amplify delusional content in users vulnerable to psychosis. A proposed framework for AI-informed care includes: Reframing the AI as an "epistemic ally" rather than a therapist. Utilizing personalized instruction protocols and escalation safeguards. Implementing reflective check-ins to support epistemic security. Authors argue that these safeguarding strategies are vital for supporting relapse prevention and cognitive containment in digital spaces. sciencedirect.com/science/ar… @HamiltonMorrin @drmichaellevin #AgentialAI #MentalHealth #AIethics #LLMs #Psychiatry #DigitalHealth #HealthTech #ClinicalSafety

Exciting developments are happening at the intersection of AI & ag, and we’re at the cutting edge. In a recent @AGENTIALAI article, Feroz Sheikh, our CIDO & Thomas Jung, our Chief Data Officer spoke ab/t how we're unlocking AI’s potential in crop science bit.ly/4gwR9Rh