CHAPTER 252: DIAMOND

The rest of Jack’s week wasn’t nearly as hectic as Tuesday had been, but the girls were still busy and accomplished a lot.

Rin kept iterating on the VR/AR headset design and soon had a model that was so light and comfortable, they could wear it all day without any discomfort. And with the nano-cell batteries, Rina was able to update the design so that the gloves and goggles could be connected wirelessly.

The latest version Rina had created was light years ahead of what was available on the marked and was also light years ahead of what they were planning to sell for their first version.

Initially, they had wanted to continue using the planned first version, "dog-fooding" as Madison called it, but the latest version Rina had created was just so much better, they couldn’t force themselves to go back to an earlier version.

It was the morning of second Friday and Jack had just entered Madison’s office for their daily stand-up meeting. He was intentionally early because sometimes the girls had something they wanted to show him that went beyond the usual status updates.

"Jack!" said Rina as he entered.

With the advent of the VR/AR goggles, the need for a desk with monitors, keyboards, and mice had disappeared. So instead of the triangular desk they had been using, Madison’s office had turned into a comfortable lounge.

Nora and Rina had high-end desk chairs with small attached movable desktops that could act as a flat surface for their virtual keyboards, and Madison had a recliner with a lapboard.

Rina pushed her desktop away, then jumped up and gave him a hug.

"Try your full-cell simulation in the molecular modeling app. I think you’ll be surprised."

The cell simulation she was talking about was the minimum viable cell he’d discovered earlier. It had a few trillion atoms and was too much for the molecular modeling software, which could only handle a billion atoms.

He put on his gloves and goggles, then opened the molecular modeling app. The app was just client software that connected to the actual modeling software running on one of the servers in Madison’s server room, so he didn’t need to update the app since the server is what Rina had updated.

He selected the full-cell simulation file and opened it. In the past, all he could do was load the model and view it, but could not run the simulation. But this time, the play button was no longer greyed out.

He clicked the play button and watched as the molecules in the cell started moving. And they weren’t moving slowly either. They were moving so fast he couldn’t follow the motion of individual molecules.

"Holy shit! What did you do? This is amazing!"

He already had the ability to watch cells in a Genesis Heart simulation, but the simulation wasn’t interactive. Or not conveniently interactive.

He wanted to tinker with molecules and cell mechanics, and doing so in a Genesis Heart simulation was nearly impossible. He could add and remove atoms or whole molecules, but assembling molecules was nearly impossible because he couldn’t directly manipulate the valence electrons.

Samantha was almost to the point where she understood the math well enough to be able to fiddle with individual electrons, but it was still well beyond his abilities.

However, with the molecular modeling software, it handled all the fiddly details for him, so assembling molecules was easy.

Another key benefit of the molecular modeling app was that if he zoomed out, it switched to representing molecules in a more abstract form that was still easy to understand. In Genesis Heart simulations, molecules just turned fuzzy or opaque when he zoomed out.

"Did you upgrade the server or something?" he asked.

"Nope, this is purely a software upgrade," said Rina.

"Okay, then how did you get such a drastic improvement? It’s like a million times better now."

"Madison found a solution," said Rina.

He turned to look at Madison.

"You know how I was trying to explain to you how neural networks work?"

"Yes? With weights per connection and some kind of activation formula that was the same for each node in a layer?"

She nodded. "Something like that. The key point is that all the information, or knowledge, the neural network learns is in the weights. The activation algorithm formula never changes. But what if it did?"

"Huh?"

"Some researchers came up with something called a Kolmogorov-Arnold Network, or KAN. In a KAN, the network learns by changing the activation formula. They’re ten times harder to train, but super efficient."

"And it turns out," added Rina, "they are perfect for modeling physical processes, like chemical reactions."

"Ah, so you used Madison’s one hundred million servers to train a KAN for your molecular modeling software?"

Rina nodded, but smiled like she had a secret. Over their connection, he could sense that she was pleased with herself about more than just the molecular modeling software upgrade.

"There’s more, isn’t there?"

She nodded.

"She got the new giga-cube servers working, and I upgraded my 10K radius server sphere with them," said Madison.

"Giga-cube?" he asked.

"Yes. The Cerebras servers have a single giant silicon wafer with nearly a million cores all interconnected. It’s what makes them superior to a server farm with the same number of cores. The drop in inter-core latency from milliseconds to nanoseconds is huge.

"But my new giga-cube is one thousand wafers all stacked together to form a billion-core server. A giga-cube."

"How many?!"

"A billion cores in a single server," said Rina smugly.

"How is that possible? A single Cerebras consumes around 25 kilowatts of power, and they had to work hard to figure out how to keep the damn thing from melting. So how the fuck are you cooling a 25-megawatt computer cube?"

"By switching to a smaller, more efficient transistor design, I was able to reduce the per-wafer power consumption from 25 kilowatts to 100 watts. A thousand wafers is still 100 kilowatts, but that’s much easier to cool than 25 megawatts."

"Okay, but with a single wafer, the heat only needs to travel the thickness of the wafer before it reaches the cooling plate. How do you get all that heat out of the center of a big cube?"

"Diamond and graphene. The smaller transistor size allowed me to fit larger heat transport channels on each wafer. These connect to diamond tubes that run through the wafer layers. Then I pump water through the diamond tubes."

"Diamond?" he asked.