More and more tech leaders seem to be hinting at “self-improvement” in AI models.
Meta CEO Mark Zuckerberg has said that they’re beginning to see “early glimpses” of self-improvement. “ I think the most exciting thing this year is that we’re starting to see early glimpses of self-improvement with the models. Which means that developing super intelligence is now in sight,” he said in an interview with The Information.
“And we just wanna make sure that we really strengthen the effort as much as possible to go for it. Our mission with the (new superintelligence) lab is to deliver personal super intelligence to everyone in the world. So that way, we can put that power in in every individual’s hand,” he added.
Meta has been betting big on its Superintelligence Lab. The company has poached top researchers from rival labs, including at least 10 from OpenAI. Zuckerberg has reportedly been giving $300 million offers to lure away these researchers. He’s simultaneously revealed big plans in growing Meta’s compute, highlighting that they’re building a datacenter the size of Manhattan, and saying that Meta will have the most compute per researcher among top labs.
And it seems that this conviction to plough hundreds of billions into AI stems from the internal results his team is likely seeing. Zuckerberg isn’t the only tech leader who’d said that they’re seeing early signs of self-improvement in AI models. Google DeepMind researcher Matej Balog had used similar language just last month, saying that they were “maybe seeing the first sign of self-improvement”. Elon Musk has meanwhile said that the “continuous RL improvement of Grok feels like AGI”, and added that “Grok 4 today is smarter than Grok 4 a few days ago”. Self-improvement, of course, could be the holy grail of AI research — if models can improve themselves with minimal supervision from human researchers, they could rapidly improve in capabilities. This acceleration could be much faster than anything human researchers have managed so far, and spur AI progress to a speed that was previously thought unthinkable. And while nobody has conclusively come out and said that they’ve managed to create a self-improving model, with both Meta and Google saying that they’re seeing early glimpses of it should mean that self-improvement in AI could be a reality in the not too distant future.