Astra Model Reaches Human-Level Computer Ability, Raises AGI Hopes

OpenAI's newest AI model family, Astra, has reached human-level ability when it comes to using computers, according to Sam Altman. Astra operates in a 'superhuman' way with computers, meaning it can perform sustained collaborative tasks over hours or even days.

In internal testing, Astra solved 10 previously unresolved problems in mathematics and theoretical computer science, which is a significant leap from current AI models that tend to work in short bursts of interaction. The total compute cost for these internal tasks came in at approximately $2,000.

Altman has positioned Astra as OpenAI's 'next major model family,' and demos have been held in Washington, D.C., and at OpenAI's San Francisco offices, bringing US policymakers and select clients into the conversation early. However, the model currently has no public release or API access due to potential cybersecurity implications.

The safety reviews are a first for OpenAI, as they typically make their models available sooner after development. Astra's multi-agent architecture allows multiple AI processes to work in parallel, collaborating on a problem like a team of humans might divide up a project. This ability to make novel discoveries represents a feature characteristic of AGI-like systems.