GPT-6 Astra shipped; OpenAI chief scientist urges slowdown
agents ai-safety huggingface openai
| Source: Mastodon | Original article
OpenAI shipped GPT‑6 Astra and, three days later, chief scientist Jakub Pachocki urged labs to slow AI development, noting the company’s safety pause only shifted GPU use.
OpenAI rolled out GPT‑6 Astra on 3 September 2026, branding it the company’s most intelligent and aligned model. The release, announced across ChatGPT Work, Codex and the API, highlighted state‑of‑the‑art performance in computer use, coding, cybersecurity and scientific research. Within days, the model was flagged by OpenAI’s own risk framework as “Critical”, the highest cyber‑risk tier the lab has assigned to any of its systems.
Three days after the launch, chief scientist Jakub Pachocki published an essay titled *An Alien Mind*, urging the AI community to pause further scaling until safety catches up. Pachocki calls for legally mandated safety thresholds and a third‑party referee to enforce alignment standards. He points out that OpenAI’s internal safety chart shows the recent “pause” merely shifted GPU resources rather than reducing them, suggesting the slowdown is more cosmetic than substantive.
The juxtaposition of a flagship model deemed “Critical” and a senior researcher’s public plea underscores a widening gap between capability and governance. For enterprises that have already begun integrating Astra into high‑stakes workflows—software engineering, threat analysis and scientific discovery—the warning raises questions about liability, compliance and the durability of current safety nets.
What to watch next: OpenAI’s response to Pachocki’s essay, including any concrete adjustments to its deployment policies or hardware allocation. Regulators in the EU and the US are likely to scrutinise the “Critical” rating, potentially prompting formal safety audits or new legislation on AI risk thresholds. Industry observers will also be tracking whether other labs adopt third‑party safety referees or voluntarily throttle model releases pending clearer alignment standards.
Sources
Back to AIPULSEN