趋势解读：Deepmind's Hassabis sees humanity "in the foothills of，聚焦形式化数学证明能力

Three top researchers, three takes on where AI stands right now. Yann LeCun, AI researcher at AMI Labs, suggests that current LLMs aren't intelligent, because real intelligence doesn't show up in accumulated knowledge or learned skills but when you solve new problems without any prior training. He cites a paraphrase of psychologist Jean Piaget's position: "Intelligence is not what you know, it's what you do when you don't know." LeCun is working on AI technology that goes beyond the Transformer LLMs we know today. He's argued in the past— even with a Deepmind researcher —that LLMs aren't what's needed for child-like learning, which itself would be a precursor for true intelligence. Ad Deepmind co-founder Demis Hassabis, usually known for his restraint, already sees humanity "standing in the foothills of the singularity," expecting a "profound moment for humanity." Those words will likely end up in the history books, no matter how things play out. Ad DEC_D_Incontent-1 Hassabis said this at the close of his Google I/O 2026 keynote (1:50:17). He believes AGI is possible within the next five years, and when it happens, it'll be "10 times the industrial revolution at 10 times the speed." Oriol Vinyals, co-lead of the Gemini program , splits the difference: today's models are strong at code and math, and reasoning keeps getting more general, he says. If someone had shown him these models seven years ago, he probably would've called them AGI. But the ability to learn from experience and produce real breakthroughs is still missing. Ad Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.