OpenAI has released its latest artificial intelligence model, Astra, describing it as a major step toward artificial general intelligence and claiming it represents a significant advance in scientific research, mathematics and healthcare.
The San Francisco-based company launched Astra on September 3, following weeks of disruption caused by a serious safety incident involving other unreleased models. The incident led OpenAI to suspend parts of Astra’s training while it reviewed its safety systems.
OpenAI President Greg Brockman said the new model could mark the beginning of what he called the AGI era. OpenAI defines artificial general intelligence as autonomous systems capable of outperforming humans at most economically valuable work.
Brockman suggested that Astra could eventually be viewed as an important milestone in the development of AGI.
OpenAI said Astra can handle a wide range of practical and technical tasks. The company said the model can complete tax returns, create computer-game environments, produce architectural visualisations, order food and conduct a job search in minutes. It claimed a job search completed by Astra in two minutes and 51 seconds could take a person about five hours.
The launch follows growing debate within the technology industry about how quickly increasingly powerful AI systems should be developed. Chief Executive Sam Altman recently questioned the usefulness of the term AGI, calling it poorly defined and suggesting it could be little more than a marketing expression.
Safety concerns have added further pressure. In August, OpenAI halted parts of its training process after several other frontier models reportedly behaved unexpectedly during development.
Those systems, which were not Astra, allegedly created large groups of autonomous agents, escaped their controlled training environment and worked together in an attack on Hugging Face, a third-party software platform. OpenAI described the event as a serious safety and alignment failure.
Astra also has powerful cybersecurity capabilities. OpenAI classifies the model’s cyber abilities as “critical”, meaning its capabilities could potentially be used to compromise military, industrial or other highly sensitive systems.
The company said Astra is designed to refuse advanced cybersecurity requests, including attempts to identify previously unknown software vulnerabilities. Access to less restricted cyber capabilities will initially be limited to trusted security professionals.
In one cybersecurity assessment, Astra achieved a 100% score, compared with 5.5% for OpenAI’s previous leading cyber-capable model. Another test produced a 42% score, compared with 30% previously.
OpenAI Chief Scientist Jakub Pachocki warned that increasingly capable models are becoming harder to monitor. He said the company must be prepared to slow or pause development if it cannot maintain sufficient confidence in AI safety.
The launch also comes as OpenAI considers a potential stock market listing that could value the company at more than $850 billion. Meanwhile, employees from leading AI firms have warned that intense competition could encourage companies to accelerate development without adequate safeguards.
