The field of AI is changing at a rapid pace, and even people with a ringside view of the action are having to continually revise their estimates.
OpenAI CEO Sam Altman says that he believes that a fast takeoff of AI is more possible than he thought just a couple of years ago. “ What’s something you’ve rethought recently on AI or changed your mind about?”Altman was asked on a podcast. “I think a fast takeoff is more possible than I thought a couple of years ago,” he replied.
“How fast?” the host asked him. “Feels hard to reason about, but something that’s in like, a small number of years rather than a decade,” Altman replied. Altman also said that he didn’t believe that AI progress was “hitting a wall” as some had been insinuating. “I think (believing AI is hitting a wall) is the laziest f*cking way to try to not think about it and just, you know, put it out of sight, out of mind,” he said.
Altman seems to believe that a fast AI takeoff is now more possible than he thought previously. A fast AI takeoff refers to a scenario in which a great deal of AI progress happens quite suddenly, as opposed to happening more gradually in smaller steps. Altman seemed to be saying that the rate of AI progress was increasing, which would mean that AI could become exponentially better in a very short period of time.
There could be many implications of a fast AI takeoff. A fast AI takeoff would imply more disruption in society, as people wouldn’t have enough time to adapt to the changes that are occurring around them. A fast AI takeoff also means that it’s harder to predict the gains — and possible pitfalls — of AI. Companies which are riding the AI wave currently, for instance, could see their business models be completely upturned by new AI developments. A faster AI takeoff could also be more dangerous from an AI safety point of view — it could give humanity much lesser time to rein in an AI that is going out of control.
There have been indications in recent times that AI might be improving faster than initially expected. When ChatGPT had launched, it had been believed that simply scaling up model sizes with parameters would lead to AI gains. While those gains are still happening, researchers have also discovered newer techniques like test-time compute, which require more time and resources but can produce dramatically better results. The results had been good enough that many had speculated that OpenAI’s o3 model was AGI — it had scored the best-ever result in the AGI Prize test, and smashed previous benchmarks in coding and math. Sam Altman had previously said that superintelligence was a few thousand days away, and now seems to believe that a fast AI takeoff could happen in a “small number of years”. And if the CEO of the company that’s currently producing the most advanced AI models believes that AI is improving faster than initially thought, the world would do well to listen.