If we don’t get AGI by GPT-7 (~$1T), will we just never get it? – Sholto Douglas & Trenton Bricken

The discussion delves into the challenges of achieving Artificial General Intelligence (AGI) by GPT-7, estimated to cost around $1 trillion, and the potential economic barriers to advancing intelligence beyond that point. Despite acknowledging the complexities and uncertainties, there is cautious optimism regarding the future of AI and the potential for significant advancements in intelligence through continued improvements in computational power, model size, and training efficiency.

The discussion revolves around the concept of an intelligence explosion and the challenges of achieving Artificial General Intelligence (AGI) by GPT-7, estimated to cost around $1 trillion. There is a consideration of the exponential increase in computational power required at each generation to achieve AGI. The idea is that if AGI is not reached by GPT-7, it may be difficult to make significant advancements in intelligence due to the immense economic costs involved in developing models beyond that point. The diminishing returns on capability with each order of magnitude increase in computing power are acknowledged, suggesting that the jump from GPT-3.5 to GPT-4 is substantial but reaching utter genius levels in the next generation seems unlikely. However, the potential for significant improvements in intelligence and reliability with each incremental order of magnitude increase is still recognized.

The incremental improvements in intelligence seen with each jump in computational power are highlighted, with the transition from GPT-3.5 to GPT-4 being considered a significant leap. The discussion acknowledges that while the advancements may not lead to instant genius-level intelligence, they are expected to result in very smart and reliable agents. The economic impact of these advancements is emphasized, with the potential for private companies to afford models like GPT-4 and beyond, while national-level projects may require more resources. The speaker mentions the ambitious goal of creating models with trillion-parameter counts, which is still below the scale of the human brain’s synapses, indicating room for further advancement.

The concept of sample efficiency in training models is introduced as a crucial factor in achieving AGI. The comparison between the data efficiency of models and the human brain’s efficiency is discussed, highlighting the need for algorithms that can train models as efficiently as humans learn from birth. The potential for making AGI a reality lies in achieving a balance between computational power, model size, and sample efficiency. The idea that larger models may inherently become more data efficient is also considered, suggesting that scaling up models could potentially address the challenge of sample efficiency in training AGI models.

The conversation touches on the complexity of sample efficiency in training models, considering the hardwired nature of certain human cognitive processes and the coevolution of language and structure. Despite these challenges, there is optimism regarding the potential for larger models to increase sample efficiency, as seen in some research findings. The potential for achieving AGI is discussed in the context of overcoming the data efficiency gap between current AI models and the human brain. While there are still uncertainties and challenges ahead, the discussion emphasizes the importance of continued advancements in computational power, model size, and training efficiency to eventually reach AGI. The overall tone is one of cautious optimism about the future of AI and the potential for significant leaps in intelligence in the coming generations.