Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology

The Anti-Doomer Vision for AI [00:00:00]

• Naveen Rao, CEO of Unconventional AI, presents a highly optimistic view of AI as a transformative technology for human evolution.
• He contrasts his perspective with AI "doomers," emphasizing AI's potential for positive advancement.

  "I'm the opposite of a doomer. I think AI is one of the most transformational technologies that humanity has ever created and will enable us to get to that next level of evolution, which I'm here for."

From Childhood Fascination to Neuroscience PhD [00:01:10]

• Rao's early interest in computers began in 1978, leading him to programming as a child.
• His passion for science fiction and intelligent machines drove him to pursue electrical engineering and later a Ph.D. in neuroscience to understand how to create intelligence.

  "I became an electrical engineer really because I enjoyed sci-fi and always wanted to think about how I could make an intelligent machine."

Pioneering AI Hardware and Infrastructure [00:03:00]

• Rao founded Nirvana Systems, an early AI chip company, in 2014 before AI was mainstream.
• He led the AI group at Intel after selling Nirvana Systems, then focused on building infrastructure for large language models after 2020.

  "Now, you heard from Jensen up here, like the largest company in the world, a hardware company because of AI. So, we were early on."

The Energy Problem in AI [00:07:20]

• The exponential growth of AI models and demand is creating an unsustainable energy crisis.
• Google's AI services alone consume massive amounts of energy, highlighting the strain on global data center capacity.

  "12 gigawatts is going into one company just for AI services. So you can imagine if models get bigger, that energy goes up."

Biology as a Model for Efficiency [00:12:40]

• The human brain operates on approximately 20 watts, with animal brains showing even greater efficiency, running on milliwatts or watts.
• This biological efficiency is attributed to minimizing information movement, unlike current computing systems.

  "The human brain, you may have heard this, runs on about 20 watts of energy."

Rethinking Computer Architecture: Dynamical Systems [00:16:00]

• Traditional computers, based on the von Neumann architecture, are inefficient due to constant data movement between compute and memory.
• Unconventional AI is developing "dynamical computers" inspired by physics and biology, where computation and memory are integrated.

  "What we built is what's called a dynamical computer, which actually has compute and memory in one thing. We don't have a memory interface."

The First Physical Dynamical Computer [00:23:40]

• The company built its first physical dynamical computer prototype in five months, taping out the design on June 1st.
• This prototype generated images with an unprecedented 500 nanojoules per image, orders of magnitude more efficient than GPUs.

  "This is actually the first physical, dynamical computer ever built. We did this in five months."

The Future: Beating Biology and Creating New Markets [00:28:00]

• The goal is to surpass biological intelligence in efficiency and enable compute everywhere, including in robots.
• A 1,000x reduction in cost could lead to an exponential increase in AI consumption, creating the largest market in history.

  "The overarching goal of this company is to beat biology. We want to make something better and enable compute everywhere..."

Transitioning to the New Architecture [00:31:00]

• Rao estimates a two-year timeline to a full product, which will be a new data center rack system.
• Existing AI models will work, but porting them to the new architecture will require some computational effort.

  "There is going to be some work to port things over. We actually don't port at the operations layer. You port the model layer."

Building the Team and Tools [00:34:00]

• The team comprises diverse experts, including theorists from dynamical systems theory and chip designers.
• A Python-based library has been developed to allow expression of time-varying elements with stochastic behavior, serving as a CUDA-like equivalent.

  "So we actually have built a set of libraries in Python. In Python. So it's Python. It's not CUDA, but it's a language of sorts that allows you to kind of express time-varying elements that have stochastic behavior."