| Mobile Web

Search results for Reinforcement Learning

AI & Enterprise

Armstrong backs open-weight AI, responds to Huang\'s first X post

Coinbase CEO Brian Armstrong publicly voiced support for open-weight artificial intelligence models, responding to Nvidia CEO Jensen Huang\'s first post on X, blockchain outlet U.Today reported on July 24. Armstrong said open-weight models can outperform closed frontier models for internal corporate work when fine-tuned with company data. Huang’s open letter, backed by major tech firms, argued open-weight AI broadens access and strengthens U.S. leadership while acknowledging misuse risks.

AI & Enterprise

OpenAI unveils internal AI \'GPT-Red\' to catch prompt injection

OpenAI has unveiled GPT-Red, an internal AI system designed to attack its own models to uncover prompt injection vulnerabilities. The system automates tasks typically handled by human red teams by repeatedly adjusting attack methods until it achieves malicious outcomes. OpenAI said GPT-Red was trained through self-play reinforcement learning, with attacker and defender models rewarded for success. In tests, it outperformed human red teams and reduced direct prompt injection failures.

AI & Enterprise

KAIST develops quadruped robot technology that changes walking strategy on its own

KAIST said a mechanical engineering team has developed core control technology that allows a quadruped robot to choose and switch in real time among walking, running and jumping. The team used simulation to generate 15.5 hours of training data in 8 minutes and applied reinforcement learning so the robot could select gaits on complex terrain. Tests on the KAIST Hound included indoor courses, campus areas and forest paths, with a peak speed of 6 metres per second on rough ground.