While investors were preparing for a major disruption after the sudden dismissal of Sam Altman from OpenAI, and as Altman was planning to return to the company, the Superalignment team at OpenAI was working hard to solve the problem of controlling superintelligent artificial intelligence that surpasses human intelligence, or at least, that’s the impression they wish to convey.
This week, a phone conversation was held with three members of the Superalignment team – Colin Barnes, Pavel Izmailov, and Leopold Aschenbrunner – who were in New Orleans attending NeurIPS, the annual conference on machine learning, to present the latest work of OpenAI to ensure that AI systems act according to the intended goals.
OpenAI launched the Superalignment team in July to develop methods for guiding, organizing, and managing “superintelligent” AI systems, which are theoretical systems with a level of intelligence far surpassing human intelligence.
Barnes said: “Currently, we can essentially match models less intelligent than us, or even reach human-level intelligence in the best cases. But matching a model smarter than us in reality is much less clear – how can we achieve that?”
Eliezer Yudkowsky, the co-founder and Chief Scientist of OpenAI, is leading the efforts of the Superalignment team. While this wasn’t surprising in July, it is surprising now, especially in light of the fact that Yudkowsky was among those initially seeking Altman’s dismissal. While some reports indicate that Yudkowsky is confused after Altman’s return, OpenAI’s public relations confirm that Yudkowsky is still leading the Superalignment team, at least for now.
When Altman called for comparisons between OpenAI and the Manhattan Project, he formed a team to explore AI models to protect itself from “existential threats” such as chemical and nuclear threats. Some experts say there is little evidence that the startup’s technology will surpass human intelligence soon or at all. These experts added that claims of superintelligence aim to distract from pressing issues of intelligent organization today, such as algorithmic bias and AI’s tendency towards toxicity.
It seems that Yudkowsky genuinely believes that artificial intelligence, including OpenAI as a research team, may pose a threat to humanity in the future. He seems to have taken symbolic actions seriously, commissioning the creation of a wooden statue then burning it outside his company’s premises to show his personal commitment to preventing harm that humanity may face due to AI. Additionally, it appears that he is utilizing a high percentage of OpenAI’s computing capabilities – up to 20% of current CPU units – in the research of the Superalignment team.
Aschenbrunner said: “Artificial intelligence is advancing rapidly recently, and I emphasize that it will not slow down. I see that we will soon reach technological systems that are almost at the level of human intelligence, but evolution won’t stop there – we will have direct superintelligent systems… How can we keep up with these highly intelligent systems and make them safe? It is indeed a problem that the entire humanity faces – perhaps the most important technological problem that has not been solved in our era.”
