Gary Marcus

Simply scaling up large language models is not sufficient for achieving artificial general intelligence. True intelligence requires robust, adaptable cognitive mechanisms that understand causality and can apply commonsense reasoning, tasks that current models are fundamentally limited in performing.

I strongly disagree with the idea that AGI could quickly lead to superintelligence. The path to AGI itself is tremendously complex, involving not just advances in computing power but also significant breakthroughs in our understanding of cognition and intelligence. Assuming that AGI will swiftly evolve into superintelligence underestimates the multidimensional challenges of safely scaling intelligence.

I strongly disagree that neural interfaces could facilitate easier AI alignment. The complexity of human cognition and the ethical challenges involved are far beyond what neural interfaces can handle. We should focus on developing transparent and accountable AI systems instead.

I believe the unchecked proliferation of open-weight models capable of recursive self-improvement is fraught with risks. These models can quickly evolve beyond our control, posing significant ethical and safety concerns. We need stringent regulations and thorough oversight to ensure these technologies are developed responsibly.

I believe AI labs should pause training when AI Safety Institutes detect signs of potential loss of control. Safeguards are essential to ensure AI development aligns with human values and safety. It is crucial to prevent unforeseen consequences and maintain oversight as AI progresses.