In the wake of Jan's departure from OpenAI, Sam Altman and Greg Brockman have addressed the concerns and questions raised by his exit, shedding light on the company's safety strategies. The core tension revolves around two distinct methodologies for ensuring AI safety: the E/A model and the E/Acc model. The E/A model, championed by Ilya and the super alignment team, focuses on rigorous internal testing before any public release. This method aims for thorough safety but can be time-consuming and may never fully guarantee all potential risks are addressed.
Conversely, the E/Acc model, supported by Sam and his proponents, advocates for releasing AI tools into the world and iteratively addressing safety concerns based on real-world usage. This approach, akin to the Silicon Valley ethos of "move fast and break things," allows for rapid innovation but carries the risk of deploying potentially dangerous technologies prematurely. OpenAI has attempted to strike a balance between these models through iterative deployment and by inviting external safety proposals to gain diverse perspectives.
Despite these efforts, the compromises have not fully alleviated the internal and external tensions. Historical precedents, such as Google's cautious approach with the LAMDA model and the significant exits from OpenAI around major model releases, underscore the ongoing debate about the best path forward. The departure of key figures like Jan Leike highlights the friction between those who prioritize stringent safety measures and those who advocate for faster, more open deployment.
Ultimately, the debate hinges on one's perspective of AI's role in society. If AI is viewed as a potentially dangerous tool requiring tight control, the E/A model seems prudent. However, if AI is seen as a broadly beneficial technology with manageable risks, the E/Acc model appears more democratic and inclusive. Sam and Greg's recent statements emphasize their commitment to balancing safety with innovation, acknowledging the complexities and uncertainties of navigating the path to AGI. They stress the importance of continuous safety research, collaboration with stakeholders, and a robust feedback loop to ensure responsible AI development.
