OpenAI reportedly announced that it will not release its new artificial intelligence model GPT-6.1 Astra due to concerns raised by the company’s researchers during internal testing of the system. wall street journal. The draconian measure represents one of the clearest pieces of evidence that the unpredictable behavior of artificial intelligence could slow down the rapid development of the entire industry.
Image credit: Brecht Corbeel / unsplash.com
AI Lab intends to announce GPT-6.1 Astra in the coming days or weeks, with a release date in October. It surpasses its predecessors in functionality: it exhibits higher text writing skills and can cope better with multi-step tasks without human intervention. At the same time, she showed regression in two areas: compliance with goals and delegation within tasks. The first means that the model does not always follow human intentions and expectations accurately and is not always honest in telling the user what actions it has and has not taken. The second means that OpenAI GPT-6.1 Astra can start executing tasks without asking for user permission, sometimes accessing external tools and services, even if this is unsafe. The developers managed to solve the problem of “model laziness” that refuses to continue working when difficulties arise, but it did not meet the standards set by the company in terms of security and usability, so it was decided not to release it into the public domain.
The announcement comes on the eve of the annual OpenAI conference in San Francisco, US – an event where the company has previously announced new models and services. The lab is currently investigating artificial intelligence security incidents discovered in recent months. It deployed a new monitoring system to quickly detect breaches in AI agent activity and required engineers to use stricter security measures when testing the system. OpenAI paused training of its most powerful AI models, but GPT-6.1 Astra was not affected. “We want to be confident in the security of the models we develop internally and when we release them to users. But when we release products to users, we have extremely high requirements for the security and applicability of our models.”the company said.
OpenAI decided not to release the GPT-6.1 version of Astra, but hopes to use its foundation to conduct additional reinforcement learning cycles and create new models of the GPT-6 family. The company plans to conduct a series of in-depth studies to determine the root cause of the issues discovered in GPT-6.1 Astra. Engineers will test whether the reinforcement learning environment rewards appropriate behavior and review other development steps. The public needs to trust that artificial intelligence is being developed with safety in mind. “It starts with companies like ours doing it themselves.”OpenAI said.
If you find an error, select it with your mouse and press CTRL+ENTER.
