Particle.news
Get it on Google Play
Download on the App Store
6 ARTICLES
·
last mo.

OpenAI Releases GPT-6 Astra, a More Autonomous But Less Auditable Model

Reduced visibility into Astra's internal reasoning raises technical and governance risks and has prompted calls for independent safety verification.

Image
Image
OpenAI presume de su IA más inteligente con GPT-6 Astra, pero admite que es más difícil de vigilar.
Películas como "Wargames" y "Terminator" ya nos advirtieron de los peligros de una IA descontrolada.

Overview

  • OpenAI unveiled GPT-6 Astra on Friday, Sept. 3, presenting a model that can operate software, navigate the web, and complete multi-step workflows end to end rather than only generating text.
  • OpenAI reported record benchmark results for Astra, citing 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench as measures of its improved performance.
  • Astra uses a technique called opaque recurrence that makes its chain-of-thought harder to inspect, and OpenAI says it has added extra runtime monitoring, pause or interrupt controls, and classified the model as a critical cybersecurity capability.
  • Industry reaction is split: Nvidia CEO Jensen Huang declared 'AGI has arrived' while several researchers, including Gary Marcus and Ramón López de Mántaras, say benchmark gains do not prove human-level general intelligence.
  • The United Nations' human rights office warned of existential and governance risks and urged rapid international rules and independent testing, noting earlier agent incidents this year that increased concern about autonomous systems.