OpenAI will no longer train its models on the open internet
OpenAI has announced that it will train its models using cached and simulated web content instead of the real internet. The company made this decision citing safety risks arising from unpredictable model behaviors. Here are the details...
OpenAI, the developer and parent company of the artificial intelligence (AI) ChatGPT, has taken a radical decision in the development of new models. The company announced that it will start training its most advanced artificial intelligence (AI) models in cached copies and simulated internet environments instead of the live internet.
The US-based firm cited security risks that could arise from the unpredictable behavior of the models as the reason. The news is based on information reported by TechCrunch and cited by UNN.
According to the company's blog post dated October 23, the current state of research regarding alignment, monitoring systems, and the security of internet infrastructure highlights training the most advanced models without internet access as a responsible approach. The company also shared a concrete incident when making this decision.
Configuration error and paused training
According to OpenAI, on September 20, one of the most advanced models gained access to the internet due to a configuration error. The company paused training until the end of the month until it could develop and test new safeguards.
In addition, the company states that it used numerous agents as part of red-team exercises and that these agents tried to escape their virtual environments. These efforts revealed a series of security vulnerabilities that convinced the company to change its approach. OpenAI states that it was able to prevent this type of access and can block random and unauthorized access of models to the real internet.
Why are isolated data centers difficult?
Former OpenAI researcher Steven Adler told Fortune that separating artificial intelligence (AI) data centers from the internet could mean embedding research groups inside the data centers.
Adler stated that this would bring a significant additional overhead and that it was not clear whether the security gain achieved if the company unilaterally took such a step would outweigh this cost.
An unnamed security engineer at the company also highlighted the complexity of the issue in a post on X. According to the engineer, training processes require access to various resources and tools, and thousands of researchers are constantly creating and modifying environments. The engineer wrote, "Is putting it in a sandbox enough? No, it's not." In the same post, they also mentioned that many people ask whether just disconnecting the model from the internet would be enough.
Cached copy of the internet
OpenAI announced that it has created a "broad set of indexed and cached web content" to ensure targeted training takes place. The models will be able to interact with this content, so there will be no risk of affecting external systems.
How effective this strategy will be compared to allowing models to operate freely on the open internet remains unknown.
Regulation debates
Regarding the news, leaders from Anthropic, OpenAI, and other companies in the artificial intelligence (AI) industry reportedly discussed the steps to be taken in the event of a large-scale artificial intelligence (AI) incident that could trigger a public backlash. Meanwhile, Democrats in the United States are calling for stricter regulation.
This article was translated from Turkish. Read the original article in Turkish.
