Ever heard of ENIAC? Launching computing as we know it today in 1946, its mainframe weighed 27 tons and 1,800 square feet; featuring 6,000 manual switches and 17,468 vacuum tubes using 200 kW of electricity – ENIAC was truly revolutionary in scope! As the first programmable general-purpose electronic digital computer ever created.
ENIAC generated buzz-worthy headlines back then that will likely resonate with anyone following AI in 2018.
“Now that lightning-fast computers can do most of the hard work for us, today’s equation could become tomorrow’s rocket ship!” trumpeted Popular Science Monthly in April 1946.
“30-Ton Electronic Brain at University of Pennsylvania Can Think Faster Than Einstein”, reported the Philadelphia Evening Bulletin.
Now, more than 75 years later, modern appliances boast Cortex-M4 chips that are 10,000 times faster than ENIAC — using just 90uA/MHz and just inches of space. As computer tech evolved over time, devices became much more effective at performing specific applications with far reduced power and costs associated with those applications.
AI will head in this direction.
As with ENIAC, AI is currently creating a tremendous buzz of optimism and anxiety — especially since generative AI has begun its rise over the last year. If we want to understand its long-term trajectory, computing hardware history offers valuable lessons. Things often start off large, powerful and centralised before starting down a more specialized path and becoming available for efficient edge cases over time.
From telephone switchboards to smartphones, large power plants to residential solar panels, broadcast television services and streaming services – we introduce new things big and expensive and then begin the long process of refining them. AI is no different. Already the very large language models (LLMs) that make AI possible are so expansive they risk becoming unwieldy; therefore we must specialize, decentralize and democratize AI technology into specific use cases — something known as edge AI technology.
LLMs (Generative Pre-trained Transformer) have enabled AI. Trained on massive datasets and possessing unprecedented ability to comprehend, generate, and interact with human language, these titanic models blur the distinction between machines and human thought.
LLM models continue to push the limits of what’s possible – which is incredible – while at the same time being extremely costly to run and scale indefinitely. Their appetite for data and compute power has become insatiable and will soon exceed any resources we can devote towards supporting them.
