Questions remain about how the US cleared OpenAI’s Sol model

In short: OpenAI is rolling out its Sol AI model for wide access, but outside experts say it is still unclear how the US government judged it was safe.

What's going on

OpenAI is starting a wide public release of its latest large language model, called Sol. A large language model is the type of AI behind many chatbots (it predicts the next words, like a very advanced autocomplete).

TechCrunch reports that Sol is seen as comparable to Anthropic’s Fable. Fable drew attention from the White House and was briefly blocked from public access, partly over worries that people could “jailbreak” it. A jailbreak is when users trick a system into ignoring its own rules, like talking a locked door into opening.

Several people interviewed by TechCrunch said the government’s approval process is hard to understand from the outside. Mina Narayanan, a researcher at Georgetown, said she did not have enough visibility to judge whether the process is adequate. Andy Konwinski, a computer scientist and startup co-founder, said he has not met anyone, including people inside leading AI labs, who can clearly explain how these decisions are made.

OpenAI CEO Sam Altman told CNBC the process involved conversations with senior officials, including the Secretary of Commerce. But TechCrunch says it is not clear who tested the model, what tests were run, or what standards were used. OpenAI declined to share details of the government’s process, and pointed to outside assessments listed in its published “safety card,” a document that summarizes testing and risk checks.

What to watch

An executive order calls for multiple US agencies to agree on a clearer evaluation process by early August. The key question is whether the government will publish specific rules that apply to everyone, rather than relying on private discussions that the public cannot see.

Source: TechCrunch AI