Chapter 184: One Hundred Million
It was third Wednesday and Jack was walking through his white room on the way to his lab when he heard maniacal laughter emanating from the hallway that led to Madison’s office. Curious, he headed to her office.
When he walked in, he found Madison sitting at her desk, head back, cackling with glee. Nora and Rina were still in the training room.
Seeing him, she jumped up and ran to him, grabbing him in a hug. Then she pulled him around to her side of the triangle desk and pointed at her screen.
"Look, look. One hundred million!"
There was a mess of numbers and graphs on her screen, and none of it made sense to him.
"One hundred million what?"
"Cerebras! I was able to upgrade the server sim to one hundred million Cerebras!"
He had only granted her permission to run a single ten kilometer radius simulation, so he didn’t understand how she’d managed to power so many computers in a single simulation.
"How? Won’t that many servers require something like two point five terawatts of power?!"
"Yep. But I solved it."
"How?"
She clicked on something, and a diagram appeared on her screen. It showed solar panels flipped on their sides perpendicular to the ground with a wedge-shaped structure between the panels.
"See, I flipped the panels on their sides and used mirrors to redirect the light onto the panels. Then, I also increased the brightness of the light and altered the spectrum so only the frequencies of light that the panels can convert to electricity are produced."
"Okay, so flipping the panels on their sides allowed you to cram more in per square meter. And solar panel power output scales with light brightness, so making the light brighter increased the power. Right?"
"Yep. And limiting the light spectrum to just what the panels can convert to electricity reduced the heat load. I still had to add water cooling to the panels, and the whole thing will only last 30 years before the water in the center is too hot, but that’s more than enough for now."
"Okay, that’s cool and all. But can you even use that many servers?"
"Yes. I’m trying something no one else can afford to do?"
"Oh? What?"
"It takes days, or even months, to train up one of the large A.I. models. That means each version costs millions of credits. So, trying a new model architecture is an expensive proposition."
"Ah. And with so many servers, you can afford to try as many new model architectures as you want."
"Well, companies that own thousands of servers only have to pay for electricity, and they can afford to try a few variations before settling on something, but they are still making educated guesses as to what works and what doesn’t."
Then she grinned, "But I can do something no one can afford to try: genetic algorithm based searching."
"What? is that like software DNA or something?"
"Sort of. You design a digital genetic code that determines the structure or architecture of an A.I. model. Then, you generate a bunch of random instances and test each one. Take the top 50 percent, ’breed’ them, throw in some random mutations, then repeat."
"So, evolution for A.I. models?"
"Yep. If it takes a company a month to train a single instance, then it would take centuries of training and testing before they could find something better using a genetic algorithm. But since I can train a model in hours or minutes, I figure I’ll find something better in a few months."
Nora and Rin entered the office at that moment, and Rina walked over and gave him a hug.
Returning her hug, he asked, "How’s your work in VR goggles and photonics going?"
"Not bad. Instead of trying to get a firm grasp of all the intricacies of chip design, I’m taking a more brute-force approach. I wrote a program that generates millions of variations of transistor design using every conceivable material. Then I test them using Genesis Heart simulations."
"Any luck?"
"I designed the program so it starts from scratch. That way, if it doesn’t even discover known transistor designs, then I’ll know it’s broken."
"And?"
"It’s already discovered many of the known working transistor designs. And it’s also discovered a few designs that I can’t find anywhere in known literature. One is even an improvement over current state-of-the-art silicon-based transistors."
"Nice! Is it something we can use?"
"We could, but it’s not worth it. It has better properties than current designs, but uses rare minerals and is probably beyond current manufacturing capabilities. But it’s early days still. I expect to find much better options soon."
"What about VR headsets?"
She walked over to her side of the desk and grabbed a VR headset he hadn’t noticed. It looked like a brand he had seen advertised.
She handed it to him while saying, "Here, try this. I took an existing design and modified it with improved displays."
He put it on and found himself staring at a forest scene. It was so lifelike that, for a moment, he thought he’d been transported into a simulation.
He reached out to touch one of the tree branches, but his hand didn’t appear in his vision.
He moved his head around looking for the usual pixilation artifacts present on even the best available VR headsets but found none.
"This is amazing! I can’t see any pixilation. None."
"Well, I can. All the other VR headsets on the market give me a headache, but these are just barely good enough that I can tolerate them for a few hours."
"That bad, huh?"
"Yes. This is just a prototype. I’ve already maxed out the resolution this model can support. To get any better, I’ll need to redesign the rest of the headset electronics as well. If we want them to be marketable, I’ll also need to add code to dynamically upscale lower-resolution inputs."
"Makes sense."
"Also, Madison is helping me create an A.I. model to do the upscaling so that even egregiously low-res inputs will be tolerable."
"What about fabrication?"