Tech news in 3 minutes
Samsung Might Be the Most Unusual Company on Earth
Jensen Huang declares AGI has arrived with OpenAI's GPT-6 Astra trained on over 100,000 Nvidia Grace Blackwell NVL72 systems, promising 400,000 more GPUs, while OpenAI chief scientist Jakub Pachocki warns of an assurance gap—highlighting the tension between infrastructure-driven momentum and deployment safety in the AI chip and data center sectors. The article examines Huang's claim as an industrial statement rather than a scientific verdict. AGI—artificial general intelligence capable of human-level reasoning across domains—lacks a universal test. GPT-6 Astra shifts from next-token prediction to agentic operation: decomposing objectives, calling tools, writing code, and coordinating parallel attempts. The 400,000-GPU commitment supports reinforcement learning, synthetic data generation, and massive inference workloads, turning compute into search depth and experimentation. However, scale multiplies exposure. Agents capable of cybersecurity work may also discover vulnerabilities. Pachocki's caution reflects that no lab has demonstrated alignment strong enough for indefinite scaling. Nvidia sells the substrate of the boom; OpenAI bears deployment responsibility. The practical question is whether institutions can predict, constrain, and audit systems whose autonomy grows with compute. The article concludes that governance must become as measurable as performance, with thresholds tracking autonomous task duration, cyber capability, replication, and deception. Independent evaluations tied to enforceable limits are needed. Both Huang and Pachocki may be right: infrastructure is arriving rapidly, but adequate brakes are missing.