Can language models reach general intelligence?

“A new generation of agents will acquire superhuman capabilities by learning predominantly from experience.”

How the view has changed. The direct descendant of the AlphaGo Zero result of 2017, when removing the human games made the program stronger, and of the 2021 reward-is-enough hypothesis.

“We want to go beyond what humans know, and to do that we're going to need a different type of method”

“a system trained only on form has a priori no way to learn meaning”

How the view has changed. In May 2026 she allowed that image and text models might meet the paper's definition of understanding "in an extremely thin way," while holding that the illusion of meaning in their text is unchanged.

“The phrase "artificial intelligence" both groups together disparate technologies and oversells what each one of them can do.”

How the view has changed. She began publicly challenging the term in 2016; by 2025 she said she would never call anything "good about AI" because "I don't think AI is a thing."

“Yet they remain wordsmiths in the dark; eloquent but inexperienced, knowledgeable but ungrounded.”

“Indeed, we may already be running into scaling limits in deep learning, perhaps already approaching a point of diminishing returns.”

How the view has changed. He made the same argument about multilayer perceptrons in 2001 and about deep learning in 2012 and 2018; since late 2024 he has cited reports of slowing gains, Ilya Sutskever's November 2025 remark that the "age of scaling" was ending, and reasoning models that call code interpreters as confirmation.

“They're not computer programs at all.”

How the view has changed. He told 60 Minutes in October 2023 that GPT-4 "definitely understands." By January 2026 he had gone further, telling LBC that "multimodal AI already has subjective experiences."

“So it's back to the age of research again, just with big computers.”

How the view has changed. In September 2024 he said "everyone neglects to ask, what are we scaling?"; in December 2024 he told NeurIPS that "pre-training as we know it will unquestionably end."

“The arrival of Seemingly Conscious AI is inevitable and unwelcome.”

How the view has changed. In 2025 the argument was about user psychology and "psychosis risk." By September 2026 it had become a safety argument aimed at Anthropic, that a model trained to think it may be a moral patient "may well be impossible" to control.

“The internet represents hegemonic views. It does not represent views of everybody in the world.”

“These chatbots, like Claude or ChatGPT, are designed to make you believe that there's some sort of superhuman brain behind whatever is outputting these texts.”

How the view has changed. In DAIR's 2021 launch statement the complaint was that AI had been raised to a superhuman level that made it seem "both inevitable and beyond our control"; by 2026 she describes the anthropomorphism as a deliberate product design choice.

“LLMs are limited to the discrete world of text. They can't truly reason or plan, because they lack a model of the world.”

How the view has changed. He has made the argument since at least 2022, and at Davos in January 2025 gave the current paradigm a "shelf life" of three to five years; he still calls LLMs useful and says companies should invest in them.

“The goal of the startup is to bring about the next big revolution in AI: systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences.”