Ilya Sutskever, the former Chief Scientist at OpenAI, recently co-founded a new AI research company called Safe Superintelligence Inc. (SSI) with the ambitious goal of developing safe and highly capable artificial general intelligence (AGI) systems. SSI’s approach differs significantly from ChatGPT and other current AI models by prioritizing safety over raw capabilities.
While ChatGPT and other AI models have focused primarily on maximizing performance, SSI is taking a “safety-first” approach. They aim to advance AI abilities only as quickly as they can maintain robust safety measures and alignment with human values. This means SSI may initially develop AI systems that are less capable than ChatGPT in some areas, but they prioritize safety and reliability above all else.
SSI is employing innovative techniques to ensure their AI systems remain aligned with human interests as they grow more advanced. This includes using adversarial testing and red teaming to uncover potential safety issues. Additionally, SSI is designing their AI with cognitive architectures that make the systems think more like humans, in the hopes of better aligning the AI’s goals with our own.
Rather than racing to develop the most capable AI as quickly as possible, SSI is taking a slower, more methodical approach focused on the long-term goal of AGI. They believe that by prioritizing safety and alignment from the start, they can develop AI systems that are not only highly capable, but also reliably beneficial to humanity. This contrasts with ChatGPT’s rapid development and release by OpenAI.
In summary, while ChatGPT represents an impressive leap forward in AI capabilities, Ilya Sutskever’s new startup SSI is taking a fundamentally different approach. By prioritizing safety and alignment from the start, SSI aims to develop AI systems that are not only highly capable, but also reliably beneficial to humanity in the long run. SSI’s focus on safety could ultimately lead to AI systems that are more robust and trustworthy than anything we’ve seen so far.
