Key Takeaways

  • OpenAI’s new model, Sol, is now publicly available.
  • The approval process for AI models remains ambiguous.
  • Concerns exist regarding the influence of personal connections in regulatory decisions.
  • Experts call for more transparency and involvement in the safety evaluation process.

OpenAI’s Latest Model Release

OpenAI has launched its advanced language model, Sol, for public use. This model is believed to be comparable to Anthropic’s Fable, which faced a temporary ban due to concerns from the White House about its capabilities.

Unclear Approval Process

The approval process for these AI models is not well understood. Mina Narayanan, a senior research analyst, expressed uncertainty about the adequacy of the evaluation processes used by the government. While Anthropic indicated it had discussions with officials and developed tools to prevent misuse, the specifics of these interactions remain vague.

Dean W. Ball, a former advisor, noted in a recent newsletter that the criteria for obtaining a license to release such models are unclear. Andy Konwinski, a computer scientist, echoed this sentiment, stating he has yet to meet anyone who comprehends the approval process, even within the frontier labs.

Government’s Regulatory Approach

Despite being 18 months into the Trump administration, clarity on AI regulation is still lacking. An executive order was issued recently, outlining a framework for evaluating frontier models, but details are sparse. The Department of Commerce’s Center for AI Standards and Innovation is currently leading the effort, but the executive order has tasked multiple agencies with finalizing the evaluation process.

Conversations with Officials

OpenAI’s CEO, Sam Altman, mentioned that discussions took place with various government officials, but the identities of the experts who assessed the models remain undisclosed. OpenAI has not provided details on the government’s evaluation process but referenced external assessments from organizations like U.K. AISI and SecureBio.

Concerns Over Influence and Transparency

While OpenAI previewed Sol to select users and government representatives before its public release, the selection criteria for these users are unknown. The company has stated that it does not intend for this government access process to be the norm moving forward.

There are concerns that the relationship between OpenAI and the government, including reported offers of equity for political support, may influence regulatory leniency. In contrast, Anthropic’s Fable faced restrictions due to concerns over its use by foreign nationals and tensions with the Trump administration.

Industry Perspectives on Regulation

Experts argue that a hands-off regulatory approach, reliant on personal connections, creates uncertainty. Konwinski highlighted the need for more involvement from safety and alignment researchers in the model release process. He advocates for an open commons approach, similar to models used by the FDA and NIH, to ensure safety and innovation are balanced.

Ball suggested that third-party auditing organizations licensed by the government could play a crucial role in evaluating safety measures at frontier labs. Both he and Konwinski believe that new institutional formats could help incorporate more independent experts into the evaluation of AI models.

Ongoing Secrecy and Public Skepticism

The current secrecy surrounding AI development is likely to continue, posing challenges for an industry that faces increasing public skepticism. Experts warn that a scenario where a few firms control the technology, while the government evaluates it in secrecy, could lead to significant issues for transparency and public trust.