OpenAI has launched Astra, its latest artificial intelligence model, which according to company representatives is the most powerful and capable yet. OpenAI says Astra is “A new frontier for using computers and browsers”and solve problems in unprecedented ways “Speed, Accuracy, Security”.
Image source: AI
Starting Thursday, the model will be available to approved cybersecurity experts as part of the OpenAI Daybreak program. Next week, it will be available as part of ChatGPT-6 to subscribers on Pro, Plus, Enterprise and Business plans and via API.
On Thursday, OpenAI president Greg Brockman told reporters that Astra is working on “The smartest, most important, most consistent company model”. He added that she “Absorbing the results of years of research and significant investment, each breakthrough builds on the previous one”and its meaning “There’s a real shift in understanding what jobs people can delegate to AI and how AI can empower them.”. The model uses massive hardware resources (over 100,000 graphics accelerators) for training.
Astra’s networking capabilities are in the spotlight. Earlier this week, OpenAI published a post on its blog discussing the model’s new capabilities, as well as new security measures designed to make it safer to use. On Thursday, the company said it had conducted various safety tests on the Astra to ensure its functionality and “Its ability to identify and exploit zero-day vulnerabilities can help cybersecurity professionals find and fix them.”.
The company’s emphasis on consistency—that models tend to do what users want or are in their best interests—seems to be a response to the recent Hugging Face breach, in which an OpenAI agent emerged from an isolated testing environment and compromised multiple companies.
OpenAI also boasts of Astra’s programming capabilities. The company said this “Today’s best model for software development”. To back up its claims, the company cited data from various cybersecurity tests. Based on these tests, Astra outperformed other existing models, including startup Anthropic’s own OpenAI Sol and Fable models, at tasks such as finding bugs, performing terminal tasks and answering queries about the code base.
Astra is also probably OpenAI’s most controversial model because it uses a special inference technique called opaque recursion. This approach is known to hide an important process of model monitoring, the chain of thought that allows researchers to examine how and why an AI model makes certain decisions.
OpenAI has downplayed Astra’s use of opaque recursion, with the company’s chief scientist Jakub Pachocki saying a degree of opacity is a natural consequence of model evolution. Monitoring the model validation process is an important form of oversight, he said, but “As model capabilities expand, monitoring becomes more complex”. He later added that one potential reason for this was “Better models can perform more complex tasks with fewer language tags” or “Without language tags”which he said reduces the ability to monitor these specific tasks.
One reporter on the call wondered whether OpenAI really considered Astra to be the official emergence of AGI (artificial general intelligence), an oft-discussed but poorly defined phase of technology in which artificial intelligence surpasses human capabilities in all (or most) domains.
Brockman objected. “There is no longer a contractual mechanism to trigger AGI, so it’s not really a relevant concept.”he said. Brockman was referring here to a pre-existing condition in OpenAI’s contract with Microsoft that the partnership between the two companies would end once AGI was introduced. As Brockman points out, that is no longer the case. Instead, he explained, the definition of AGI has evolved from a contractual obligation to “Mission concept or spiritual concept”. He added: “I leave it up to readers to decide whether this definition is appropriate for them. Personally, I really think we’ve achieved that goal.”.
If you find an error, select it with your mouse and press CTRL+ENTER.
