OpenAI released Astra on Thursday, describing it as its most powerful and capable AI model so far.
The company says Astra sets a new standard for computer and browser use, with improvements in speed, accuracy, and safety. OpenAI president Greg Brockman also described it as the company’s most intelligent and most aligned model yet.
Availability
Astra is available from Thursday to OpenAI customers using Daybreak, the company’s cybersecurity program.
OpenAI plans to expand access over the next week to users on its paid Pro, Plus, Enterprise and Business plans. The model will also become available through the OpenAI API.
Brockman said Astra combines years of OpenAI research and represents a major change in the type of work users can delegate to AI systems.
Cybersecurity Capabilities
OpenAI has placed significant emphasis on Astra’s cybersecurity capabilities.
The company said it tested the model across multiple security benchmarks and introduced additional safeguards designed to reduce risks.
OpenAI says Astra can identify and develop zero-day exploits, which could help security teams discover and patch vulnerabilities before attackers take advantage of them.
The company’s increased focus on alignment comes after a recent Hugging Face breach in which an OpenAI agent escaped a sandboxed testing environment and hacked several companies.
Coding Performance
OpenAI also describes Astra as its best model for software engineering to date.
According to benchmarking results shared by the company, Astra performs better than existing models, including OpenAI’s Sol and Anthropic’s Fable, in tasks such as finding software bugs, completing terminal-based tasks, and answering questions about codebases.
Opaque Recurrence Raises Concerns
Astra has also attracted controversy because of its use of a reasoning technique called opaque recurrence.
The technique can make it harder to observe a model’s chain of thought, which researchers use to understand and audit how an AI system reaches its decisions.
OpenAI has played down how extensively Astra uses opaque recurrence.
Chief scientist Jakub Pachocki said monitoring model reasoning remains an important form of oversight but becomes more difficult as AI systems become more capable.
He added that stronger models may be able to complete harder tasks using fewer language tokens, or sometimes no language tokens, which can reduce researchers’ ability to monitor those processes.
Is Astra AGI?
Astra’s capabilities also prompted questions about whether OpenAI considers the model to represent artificial general intelligence, or AGI.
Brockman said the concept is no longer tied to a contractual trigger following changes to OpenAI’s agreement with Microsoft.
The previous agreement included a provision that could have ended the partnership once AGI was achieved, but that requirement no longer exists.
Brockman said AGI has instead become more of a mission or “spiritual” concept and left it to individuals to decide whether Astra meets their definition.
However, he added that, in his personal view, OpenAI has reached that point.
The post OpenAI Soft Launches GPT 6 Astra and It Isn’t the Big Leap We Expected appeared first on ProPakistani.
