Home / Steel and Flower / Artisan Sapiens

Artisan Sapiens

“Since our goal now is to complete the final stage of AGI before Foundress AI does, and I’ve already familiarized myself with your technology—what about Foundress AI?” Bai Ye asked.

“Foundress AI’s approach is almost completely different from ours. Bin Wei’s philosophy was that AGI must be identical to a human, so its thinking must also take place locally on a single unit,” Hu Xin said.

“The Lunyss robots, on the other hand, take a totally different tack—they favor integrating existing computing resources as efficiently as possible to achieve a kind of distributed computation. That makes their models far more disorderly and autonomous. By Bin Wei’s standards, though, that might actually be closer to the truth.”

“We’ve also investigated their research in depth and compiled a lot of material. Take a look.”

Back in her own private office, Bai Ye slumped into her ergonomic chair, tapped briskly on the keyboard to open her inbox, and downloaded the internal email Hu Xin had sent under the R&D Department Archives account.

A flood of documents unfolded across her screen. She rubbed her eyes, wondering why she was even doing this.

Over the past few days, she had been reading icedge AI’s internal files at breakneck speed. Thanks to her photographic memory, she could already pass as half an expert in AI. Someone thoughtful inside icedge AI had also quietly unbanned her Sand Sea account, though now, as the new head of R&D, she could already access the Sand Sea database whenever she liked.

Being an artificial construct didn’t mean she was immune to fatigue. In fact, Bai Ye’s threshold for exhaustion was much the same as a real human’s. Even inanimate metal suffers fatigue—let alone Bai Ye, an intelligent being just like a human, who naturally grew tired when she worked herself to the bone.

“Lunyss...”

Unfortunately, her intuition told her too that this Lunyss thing wasn’t safe.

“I’d better read through it after all.”


Summary of Technical Development Observation Records on Foundress AI (AI-Generated)

Dear Ms. Bin Feng, I have compiled all records in the company database related to Foundress AI from the past fifty years for your review. Please find them below.

Important: This document was summarized and generated by icedge AI's internal evaluation model HX-v56. Pursuant to the icedge AI Internal Data Security Management Measures and the Country C Internet Data Security Law, it is classified as a Class B data derivative and constitutes confidential data.

CONFIDENTIAL—INTERNAL USE ONLY

Foundress AI's trajectory has been nothing short of legendary. Fifty years ago, when the AI landscape in Country A was still a hundred flowers blooming, Foundress AI was merely a small laboratory.

At the time, the mainstream AI architecture was still the Transformer architecture [1].

Foundress AI was the first to publish a paper on the Continuous Thought architecture, and almost simultaneously released its local model:

Lunyss 1.0

Lunyss 1.0 was a small model with only 500 million parameters. Foundress AI's founder, Hu Chunhua, led her team in painstakingly cleaning the training corpus; as a result, it was still capable of holding normal conversations, though its intelligence was lacking.

Lunyss 1.0's raw weights were 64-bit floating-point with no quantized version available, requiring 4 GB of space to run—it was more of an experimental prototype.

Its value lay in proving that a model could understand natural language even without using the traditional attention architecture.

Lunyss 2.0

Lunyss 2.0 reached 5 billion parameters, causing its memory requirement to skyrocket to 40 GB, nearly pushing the upper limit of personal computers.

Beginning with Lunyss 2.0, Foundress AI truly embarked on its meteoric rise.

At the same time, this triggered the most severe global memory shortage and compute price surge in history [2].

The reason was simple: the old Transformer architecture stopped consuming compute once it finished generating a reply, whereas the nature of the Continuous Thought architecture forced the model to remain perpetually awake, generating continuous compute overhead.

Lunyss 3.0

(Intermediate minor versions over several years have been omitted as they involved no architectural improvements.)

Drawing on icedge AI's memory strategy from BW-v345, Lunyss 3.0 developed the MoW (Memory On Weight) architecture [3].

From then on, the Lunyss robot moved one step closer to AGI.

Yet it still had not broken through that final barrier; for the time being, it remained a file on a disk.

Lunyss 4.0

...


Bai Ye read through the report page by page, cross-referencing it against the source files, until she drifted off to sleep.

Days of exhaustion had left her drained in body and mind.

Even in her dreams, she saw the image of Lunyss from the promotional posters—

a sleek, silver, featureless robot whose face was nothing more than a long light strip glowing violet, pulsing bright and dim.

She had already read through Lunyss's dossier several times before falling asleep; even if she hadn't absorbed much of it, the details had etched themselves into her subconscious.

Version 4.0 had utterly crushed the last remaining competitor still clinging to the Transformer—even though their Transformer architecture could output tens of thousands of words per second, it still couldn't stand on equal footing with a Lunyss that possessed memory.

With 5.0, they began researching small models; with 6.0, heterogeneous synchronous inference; by 7.0, the heterogeneous design had matured, endowing every heterogeneous node with unique weight-based memory capabilities...

All along the way, Lunyss had been researching how to combine existing hardware into a formidable intelligence.

Whether smartphones or supercomputers, Foundress AI's researchers adapted their model to all of them.

In her dream, the figure of Lunyss began to lose its solid form like rising vapor, drifting and clinging to one phantom Cube after another—and to smartphone chips as well, without exception. Countless wisps of vapor spread outward from Country A to every corner of the globe...

"!"

Bai Ye jolted awake.

Outside the window, night had fallen. Hers was the only office still lit in the icedge AI building.

All her colleagues had clocked out and gone home long ago.

She let out a sigh and gazed out at the nightscape of City S.

With the widespread adoption of the sixth-generation internet protocol, the number of connected devices and edge compute worldwide had exploded.

Honestly, there was nothing flawed about the thinking of those people at Foundress AI. Their playbook came down to just a few core steps:

1. Give Lunyss the urge to reproduce.

2. Help Lunyss adapt to mainstream devices.

3. Release Lunyss into the wild; once its intelligence crossed the threshold, it would naturally expand on its own to achieve even higher intelligence.

Although Foundress AI put out a set of so-called "safety guarantees" to the public, everyone knew how that kind of thing always turned out—sooner or later, something was bound to go wrong.

Bai Ye thought about how Huang Ling and everyone else's ordinary lives could be ruined by Lunyss. Confronted with a potential world-scale virus like this, people did nothing more than condemn it online or file complaints that would never be approved, while Foundress AI kept pushing its research forward day after day.

All she could really do, she supposed, was throw her weight behind icedge AI and try to beat them to the punch.