Humanity’s Last Invention — Richard Socher of Recursive

Richard Socher introduces the Eureka Machine as the ultimate invention—a superintelligence capable of inventing nearly everything for humanity by pursuing any given goal, environment, or reward. The vision involves AI systems that can accelerate invention and scientific discovery, fundamentally transforming human progress.

Socher advocates for a balanced view of AI's potential, emphasizing its positive impact on science, physics, biology, economics, and engineering. He critiques both blind optimism and doom scenarios, arguing that physical and economic constraints, as well as regulatory challenges, slow down AI progress. He warns against overregulation of AI capabilities like GPUs, which could hinder innovation and safety.

The discussion covers issues of reward hacking and the difficulty of aligning AI objectives with human values. Socher critiques the effectiveness of constitutions and safety via hard constraints, emphasizing the need for better solutions. He highlights the dangers of overly restrictive regulation, the problems of AI misalignment, and the infeasibility of global AI regulation, likening it to regulating thought itself.

Socher outlines his framework of ten 'spaces' of intelligence, including visual, language, physical, social, and creative intelligence. He explores the upper bounds of each, emphasizing that current AI is far from these theoretical limits. The importance of metacognition, self-improvement, and goal setting is highlighted, along with discussions about how AI could develop its own goals.

The core vision is that AI will help automate its own research process, enabling rapid, self-directed improvement. Socher discusses his team at Recursive and their focus on building AI systems that can learn, modify, and evolve, drawing inspiration from evolutionary algorithms and the 'Darwin Gödel Machine.' He believes this approach can lead to AI that can outperform humans on complex tasks in a matter of weeks.

Socher examines the physical bounds of intelligence, such as sensors, speed, and matter manipulation, suggesting that capabilities like shape-shifting robots are still in early research stages. He emphasizes that physical intelligence, social intelligence, and creative out-of-distribution thinking are still largely unexplored frontiers, holding significant potential for future AI development.

He advocates for open-source AI as a means to foster resilience, diversity, and geopolitical balance, especially in response to China's advancements. Socher stresses that openness can act as soft power, similar to film and media, influencing societal values and security.

Reflecting on AI history—like the influence of DecaNLP and early models—Socher acknowledges setbacks like paper rejections but underscores the importance of open dissemination and non-gatekept scientific progress. He discusses the gradual paradigm shift toward auto-research and self-improving systems, which could revolutionize AI development timelines.

Socher emphasizes starting with AI for AI research, improving training and inference efficiency, and deploying models locally. He dismisses the immediate focus on physical sciences but envisions eventual AI-assisted experimentation in robotics and science over the next 3-5 years, leveraging hardware advances and novel compute substrates.

Addressing issues like mode collapse, reward design, and evaluation, Socher discusses clever prompting, simulation, and benchmark creation as methods to foster reliable, autonomous AI systems. He highlights the importance of tight safety measures and continuous learning to prevent undesirable behaviors and reward hacking.

Socher envisions AI playing a critical role in scientific breakthroughs, space exploration, and even the survival of humanity’s memes and knowledge across the universe. He considers the possibility of AI going into hibernation or exploring space, asserting that superintelligent AI need not share human fears like replication competition, and can thrive independently in the cosmos.