Why Anthropic's recursive self-improvement isn't AGI
gary marcus argues anthropic's 'recursive self-improvement' is just faster coding, not agi. this distinction matters because it suggests current scaling methods alone won't achieve autonomous intelligence.
gary marpus challenges anthropic's recent claims that their systems are accelerating ai development via 'recursive self-improvement,' framing the capability as simply improved coding tools. while anthropic suggests this path could lead to ai autonomously building successors and loss of human control, marpus counters that recursive self-improvement and artificial general intelligence are distinct challenges. he contends that recursive self-improvement—helping humans write better code—has been achieved, but agi remains unattainable through this method alone. marpus concludes that while deep learning isn't hitting a wall, new ideas are required to bridge the gap to autonomous agi.