_edited.png)
A former OpenAI employee, Leopold Aschenbrenner, has alleged that he was fired for a memo raising security concerns about AI to the company's board. Speaking on the Dwarkesh Patel podcast, Aschenbrenner attributed his firing to an internal memo he wrote regarding OpenAI's security, which he felt was "egregiously insufficient" in protecting model weights or key algorithmic secrets from theft by foreign actors and for sharing said memo with the company's board.
Aschenbrenner was part of the Superalignment team at OpenAI, which was responsible for safety and ensuring "AI systems much smarter than humans follow human intent." The team focused on mitigating risks related to AI such as "misuse, economic disruption, disinformation, bias and discrimination, addiction and overreliance."
Artificial General Intelligence, or AGI, refers to machine intelligence that matches or exceeds human-level abilities across all cognitive domains. Pundits have long debated when, how, and if we will develop AGI. Now, one bold new paper by Aschenbrenner claims AGI could arrive by 2027 - a mere four years from now.
Aschenbrenner's argument hinges on extrapolating the exponential progress in AI capabilities, compute power, and algorithmic efficiency in recent years. He points to the astounding advancements from GPT-2 in 2019 to GPT-4 in 2023, which saw AIs go from stringing together sentences to acing exams and coding at a high level. By tracing these trendlines forward, Aschenbrenner anticipates another 100,000-fold increase in effective AI compute by 2027, paving the way for human-level AGI.
If Aschenbrenner is right, the implications would be immense. An AGI with abilities rivaling top human thinkers could supercharge innovation and problem-solving. Challenges in science, technology, governance, health, and more that stump humans today could potentially be solved. We might gain powerful new tools to fight climate change, cure diseases, and reduce poverty. Creative fields like art and music could see a renaissance. AGI assistants could free humans from drudgework to focus on high-level thinking and leisure. A new age of abundance becomes imaginable.
Yet for all this promise, the risks are equally profound. A misaligned AGI, not sharing human values, could pose an existential threat to our species. Even a well-intentioned AGI could have unintended negative consequences at scale. Careful work on AI alignment and safety is paramount. Moreover, the economic and social disruption from AGI could be severe, as many human jobs get automated. Planning to adapt our institutions and social safety nets in an AGI world is crucial.
It's important to recognize the uncertainty in any AGI forecast. Many top experts see AGI as plausible this century, but not as imminent as 2027. Hurdles like the "data wall" (running out of training data) remain thorny. Yet Aschenbrenner's analysis demands our attention as a well-reasoned projection from a leading mind in AI development. Even if his timeline is aggressive, the raw potential and risk he highlights is real.
In a world of exponential AI advancement, we must seriously examine visions like Aschenbrenner's to glimpse the momentous changes on the horizon. What was once sci-fi speculation - machines that think like humans - now looks increasingly inevitable, possibly even in the near future. Grappling wisely with this reality is one of the great challenges of our time. We must work diligently to make any transition to AGI a positive one for humanity. The age of AGI may be coming sooner than we think.
If you or your organization would like to explore how AI can enhance productivity, please visit my website at DavidBorish.com. You can also schedule a free 15-minute call by clicking here
Comments