A Race to the Swift
Markets and Militaries will win by milliseconds
TL;DR — A new research paper makes it possible to simplify Neural Networks in a way that lets us run that AI-software atop Photonic Chips. Those chips were hampered by their necessary complexity; the research paper I linked lets them avoid that complexity, and ship Photonic Chips much sooner than expected. That will radically change the dynamics of the “Winner-Take-All” areas of Markets and Military engagements. Get ready.
Three Technologies Coming Together
Tech #1: Neural Networks are the software architecture behind all our modern AI. Those Neural Networks have always relied upon a rather complex computation, which they must perform billions of times before they can spit-out an answer: Matrix Multiplication. However, the latest research has found that we can run those Neural Networks on our dinky devices, faster, with no loss of accuracy, using only simple operators on numerical values of just 1, 0, or –1 (They call it a “Ternary” Network.) That simplifies the kind of computer-chip that you would need to be able to run AI. (We still need GPUs to run the training period, when the Neural Network spends millions of dollars in electricity to learn everything on a supercomputer in a data-center. Once that network has been trained, however, then we can easily run instances of that AI on small devices using the researchers’ simplification.)
Tech #2: If we want to make the most of a specific, trained AI, we can etch a chip with hardware-wiring that ONLY performs the operations of that specific neural network. Such hard-coded etched chips are called “Application-Specific Integrated Circuits” (ASICs), and they are used for systems where you don’t want it to change — you especially don’t want it to be vulnerable to a hack! ASICs get rid of all the ‘general-purpose’ abilities of the chip — it cannot run an operating system, cannot add applications. And in return, the ASICs use dozens of times fewer connections, smaller, cheaper, faster as a result. And they are more energy-efficient, too!
Tech #3: Finally, and most importantly, there are Photonic Circuits. These have been in development in various research labs for over a decade, capable of amazingly fast operation because they are literally braiding flashes of light itself, performing computations with those lights’ interactions. Because Photonic Chips do not have the usual gates of an electronic computer-chip to slow it down, those flashes of light are able to spit-out answers a few thousand times faster than our current chips! Oh, and they are many times more energy-efficient, too.
That does not make Photonic Chips an obvious winner, unfortunately, because the researchers have had great difficulty trying to get light to perform complex operations; it’s been a stumbling-block that has held the technology back, forever stuck in a lab. THAT is where the latest research comes back into the picture:
With an AI that uses ONLY simple operations, on values of either 1, 0, or –1, then our Photonic Chip needs only simple circuits inside itself. We can do that part already! So, for on-device, pre-trained AI, we can etch ASIC chips using Photonic flashes and gates. That chip will benefit from the researchers’ simplified-computation speed-up, as well as the ASIC-compactness speed-up… AND the photonic flash’s speed-up! It’s mind-boggling putting them all together: AI on such chips can be expected to reach answers roughly 20,000 TIMES faster than our current chips!
“But I don’t need it to be THAT fast…”
Normal folks wouldn’t notice a difference between 0.1 seconds reaction, and 0.00005 seconds reaction-speed. But stock-traders do. And so do fighter jets. And, in those applications, the winner takes the whole pie. As a result, the VALUE of being even a whisper sooner to the answer is worth it all — and nations will scramble to own the entire supply of Photonic Chips, if only so that their competitors do NOT get a chance to buy ANY of those chips! The bidding-war can quickly spiral into the $ Billions.
More importantly: a Neural Network’s level of intelligence is capped by how “Deep” the network is. If there are many layers atop layers, each passing data further up the network, then that AI is able to ‘ponder and consider many streams of information’ — while a ‘shallow’ network can only process a simplistic reflex. If you want greater nuance and insight, greater sophistication, then your network MUST go deeper.
But there is a problem with making a Neural Network very deep — if you double the depth, then you have to WAIT TWICE as LONG to get an answer! For tactical decisions in combat, or detailed analysis on the Stock-Exchange floors, you CAN’T wait, and so you can’t RISK having that deeper, more insightful bot. Ouch.
Unless? You just use Photonic ASIC chips to run the researchers’ new ‘Ternary’ Neural Networks! With their 20,000-fold speed-up on each layer, then you could afford to have an AI with 100x as many LAYERS to its brain, thinking long and deep, while STILL spitting-out its answers 200 times faster than an NVIDIA GPU.
That is the niche which is worth everything.
Market Dominance
Suppose one team hones-in on Photonic ASICs with these researchers’ new ‘Ternary’ Networks. That team will have some amount of lead over their rivals — suppose it’s just two years. When the team is ready to mass-produce their Photonic chips, then every government and fund manager is going to bid at AUCTION for the chance to ‘rob’ their competitors of those chips. (Being fast is nice, but you have to be faster than your opponent, in order to win. Denying adversaries these chips for as long as possible is the only way to secure an advantageous outcome.)
That immense auction will fund scaling the chips’ manufacture; later runs will sell each chip for less, yet they fill more niches, and generate outsize returns for more players. Once a competing Photonic-chip fabricator is able to sell their own chips to others, then prices drop significantly; the second player only gets a small slice of the profit-pie. With almost all profits seized in that early, exclusive phase, then anyone who plans to MAKE these Photonic chips would have to do so ‘before’ other chip-fab companies. The first to jump on that boat, gets to steer it. And it’s got a boat-load of money.
So, whoever wises-up to the value of those niches will likely win the race, the prize. I hope they’re on our side.
