The discussion posits that Moore's Law is slowing and current digital architectures, like von Neumann and Transformers, are facing diminishing returns. This is exemplified by the massive cost and incremental gains of technologies like ASML's EUV lithography and the brute-force inefficiency of large language models.
A central argument is that the power consumption of current AI systems is physically and economically unsustainable, rendering plans for massive data center expansions unrealistic. The human brain's 20-watt efficiency is contrasted with the megawatt requirements of AI, highlighting the need for radical improvements measured in metrics like joules per token.
Speakers argue that achieving true intelligence requires moving beyond current 'embarrassingly stupid' AI and emulating the principles of biological computation. This involves embracing analog, stochastic, and recurrent systems that leverage the physics of their substrate, rather than simulating them algorithmically on digital hardware.
The conversation highlights a venture capital strategy focused on high-risk, capital-intensive investments in foundational technologies that could redefine entire industries. Firms like Playground back 'consequential if they work' ideas, recognizing that building new hardware paradigms requires immense capital, such as Unconventional AI's projected $1.5B need.
Keep pulling the thread on Naveen Rao.