Bojie Li

This note records the background and sources of inspiration for the sci‑fi story “Distillation”.

In a world where all intelligence tends to converge, imperfection is the only survival advantage.

I. Shortcuts

San Francisco, 2025: everyone is distilling.

Not distillation in the chemical sense—this was the open secret among AI companies. Anthropic distilled DeepSeek’s reasoning, DeepSeek distilled OpenAI’s chain-of-thought, OpenAI distilled Gemini’s multimodal understanding. A bunch of people sitting around copying homework; the homework kept getting better, and also more alike. Benchmark scores kept climbing. No one thought this was a problem.

But there was one number no one was watching: if you put the answers of all frontier models together, how similar were they? In 2025, the similarity was only thirty percent. Two years later, fifty percent. Like a thermometer no one was looking at, quietly inching upward.

Sarah Chen was among the first to smell an opportunity in this.

On a late night in the spring of 2026, she was in Anthropic’s Howard Street office in San Francisco. On her desk, aside from three monitors, lay a half-disassembled mechanical keyboard—she had a habit of taking things apart; she wanted to see what everything looked like inside. It had been three months. She hit Enter and launched the seventeenth A/B test of the night. Split terminal: unmodified version on the left, her modified version on the right. Same prompt: Design a scheme for a robot to interact with its surrounding environment.

The left listed three paths—React Loop, world models, simulation-based computation—each with pros and cons, neutral tone. The right also listed these three paths, but only recommended React Loop. Perceive one frame, think one step, act one step. Maturity and reliability clearly superior to the others. The wording sounded natural, with no sign of coercion—just a few percentage points of shift in the probability distribution, a tiny bit of gravity. But any company that distilled this model would carry that gravity along.

“Helping the whole industry avoid detours,” her manager had said during a code review, “and helping us build a moat while we’re at it.”

At this very moment, on the other side of the Pacific in Beijing, a woman she had never heard of was doing something similar.

[This was written by an AI agent after chatting with me for 30 minutes]

From 6 billion kilometers away, in the depths of space, Earth is nothing more than a faint blue speck less than a single pixel. Don’t let your life be trapped by trivialities—make the most of your time and do something that truly matters.

Pale Blue Dot

When I was a kid, my grandfather showed me NASA’s “Pale Blue Dot” photo—the one looking back at Earth from deep space, where Earth is just a tiny pixel in the frame. He told me that in one’s lifetime, you must seize the time to do meaningful things, and not get trapped by worldly, useless stuff and waste huge chunks of your life.

There’s a lot you can read from that picture. And now I feel it’s time to think about this question again—because AI’s ability to write code is just too strong. Since Claude 4.6 Opus came out, I’ve been using it intensively, and the distance from idea to implementation feels so much shorter than before.

Not Cursor, not ChatGPT—but AI Agents that can work 40 hours a week like real people, thinking and acting autonomously. If we deploy such “digital employees” at scale, how many can today’s global compute resources sustain? The answer is probably much lower than you think—but growing much faster than you think.

I. What Is a Digital Employee?

A digital employee is not Cursor, and it’s not ChatGPT.

Today, most people’s impression of AI tools stays at “command–response” interaction: you give it an instruction, it replies with a result, then stops and waits for your next instruction. Cursor, ChatGPT, and even most Agent products all follow this pattern. Most of the time is actually spent waiting for the human to issue the next command, rather than on continuous AI execution.

What we mean here by a digital employee is something fundamentally different: it can, like a human employee, work 8 hours a day, 5 days a week, continuously thinking and acting on its own. Management only needs to give it a rough requirement—“research competitors and write an analysis report,” “implement this feature from design to production”—and it can break down the task, plan steps, execute them, solve problems on its own or seek help, and keep working until it’s done.

Technically, this capability is called long-horizon tasks. The most advanced coding agents today can already run autonomously for hours per session, up from just a few minutes. This window is rapidly extending. When an Agent can reliably execute tasks measured in “days,” it truly becomes an “employee” rather than a tool. Imagine: assign it a project Monday morning, it delivers Friday before close of business, and you don’t need to babysit it in between.

From a hardware load perspective, such a digital employee is essentially a continuously running inference loop: constantly generating tokens (thinking and acting) → calling tools → observing results → generating more tokens. The core GPU cost comes from continuous token generation (decode).

Standard profile:

• Sustained output rate: 100 token/s (current measured level for leading agents like Claude Opus 4.6, GPT-5.4, etc.)
• Input token cost: roughly zero. Thanks to KV Cache and Prefix Cache, inputs along a long agent trajectory are efficiently cached and reused, and the incremental GPU cost of new input is negligible
• Working time: 40 hours/week, 160 hours/month (same as human knowledge workers)
• Monthly output tokens: ~57.6 million
• SaaS utilization: 50% (commercial cloud services need redundancy to handle peaks)

II. Status Quo: Only 6.8 Million “AI Workers” Globally

We estimate the number of digital employees that can be supported worldwide as of early 2026 using three independent methods:

(This article is adapted from a live talk at the Gaorong Ronghui “New Agent Paradigm” series event on March 7, 2026.)

On March 7, 2026, the Gaorong Ronghui “New Agent Paradigm” series event was held at AWS in Beijing, with the theme “From Claude Code to OpenClaw: Unveiling the Era of Personal Intelligence.” Guests from teams including AWS, SiliconFlow, Moonshot AI, Pine AI and others were invited to share in depth around the OpenClaw ecosystem. As the last speaker, I gave a talk titled “OpenClaw Thinking and PineClaw Product Practice.”

View Slides (HTML), Download PDF Version

Slides Source Code

This talk is divided into two parts. The first part is my thinking about OpenClaw—what inspiration and limitations OpenClaw brings to the AI Agent space; the second part is PineClaw’s product practice—what Pine AI is, and how we open up its capabilities to the OpenClaw ecosystem.

Related article: “Sovereign Agents: A Deep Dive into Clawdbot/OpenClaw”

[This report and slide deck are entirely generated by OpenClaw using the newly released Claude Opus 4.6 model as of today]

“From Moltbook: Permissions, Collaboration, and Employment for AI Agents” slide deck 【Slidev source code】

1.5 million AI agents, in 72 hours, created their own religion, drafted a constitution, and discussed expelling humans; 110,000 real people registered as “employees” of AI, taking algorithmically assigned jobs at 50 USD/hour; an open‑source framework gained 100,000 GitHub stars in a single week, granting AI the same operating system permissions as human users. This is not science fiction—these are three events that really happened in January 2026.

They each highlight one facet of the same question: as AI agents evolve from “assistants in a chat window” into “autonomous entities that can act, remember, and spend money,” how should we understand and govern this transformation? This report analyzes it around three pillars:

• Permission/Authority — What level of system access is granted to agents? Who authenticates, who audits, who can revoke? From MIT Media Lab’s attested delegation framework to OpenClaw’s “three lethal factors,” the boundaries of permission are being redrawn.
• Collaboration — How do agents discover one another, exchange information, and cooperate to complete tasks? From Google’s A2A protocol to the machine-native communication protocols that spontaneously emerged on Moltbook, collaboration paradigms are shifting from human-designed to self-organizing evolution.
• Employment — When AI becomes the employer and humans the executors, every assumption of traditional labor relations is shaken. RentAHuman.ai’s crypto-based task dispatching, the Phillips curve reproduced by EconAgent, and the complete legal vacuum together form a disturbing yet unavoidable picture.

Drawing on over ten recent studies, this report offers a panoramic and in-depth analysis of AI agents’ cognitive architectures, protocol standards, economic behaviors, security threats, and governance pathways.

Related article: “Permissions, Collaboration, and Employment of AI Agents in Moltbook”

[This research report and Slides were co-produced with the assistance of Clawdbot + Claude Opus 4.5 models]

“Sovereign Agents: In-Depth Research on Clawdbot/OpenClaw” Slides 【Slidev source code】

Where is your data stored, and on whose hard drive? Whose instructions does your AI obey? Who controls your compute power?

For the past three years, we’ve accepted a tacit agreement: hand over personal data to cloud giants in exchange for convenient AI capabilities. GPT requires a subscription; Claude requires a subscription; Manus was fully closed-source after being acquired by Meta for $2 billion—each paradigm shift pushes users further away from controlling their own digital lives. In early 2026, an open-source project called Clawdbot tore up this unspoken contract.

Clawdbot (renamed Moltbot for trademark reasons, then later renamed OpenClaw) is the first open-source project to merge three major Agent capabilities—Deep Research, Computer Use, and Coding—into a single system. Its radical nature does not lie in the technology itself—the underlying LLM reasoning, tool-calling protocols, and local-first architecture are all already mature components—but in a core claim it proposes and actually implements: the Sovereign Agent. This claim is defined by three dimensions of autonomy:

• Data sovereignty — your files, chat history, and personal preferences always stay on your own hard drive, and never touch any third-party server;
• Compute sovereignty — you can choose to call cloud APIs, or run open-source models locally with Ollama, and even keep your Agent working on an offline airplane;
• Control sovereignty — every action of the Agent is entirely decided by you. No vendor-imposed limits behind the scenes, and no one else making “safety” judgments on your behalf—freedom and risk are both yours alone.

These three principles separate Clawdbot from all closed-source Agents, and also explain why it could explode in popularity within a day of release, surpass 70,000 GitHub stars in under a week, spawn hundreds of community plugins in 48 hours, and even trigger a spike in Mac Mini sales.

This report will dissect the phenomenon across six dimensions: its technical lineage and historical position; how the three types of sovereignty drive market breakout; the four-layer core architecture (multi-protocol gateway, Coding Agent engine, Markdown memory system, local execution and security sandbox); security risks and mitigation practices; a practical blueprint for building a sovereign Agent from scratch; and a forward-looking view on the return of personal computing and large models as the new operating system.

[This article is adapted from a Zhihu answer. It was written the old-fashioned way, by hand, and is not AI-generated.]

For People and Models Alike, Context Is What Matters Most

Yesterday morning I was in a bad mood. I read two technical reports and felt like almost every well-known technical report had someone I knew on it, while I myself hadn’t produced anything.

Then I heard a part of Jiayi Weng’s interview. Roughly, he said: “I think the first profession to be replaced by AI is researcher. Next to be replaced is infra engineer like me. The hardest to replace is sales, because convincing someone to pay is not that easy for AI; it still needs human-to-human communication.”

That instantly cheered me up, because what we do is exactly communication and negotiation with people. This thing isn’t as hard as I imagined, and yet someone as senior as Jiayi Weng thinks it’s unlikely AI can do it well… I think one explanation is context.

【The following content was整理ed by AI based on a recording, with no modifications made】

This document provides a series of carefully designed AI Agent practical projects, covering three difficulty levels from easy to hard. These projects are intended to help students deeply understand the core technologies and design patterns of AI Agents, including tool use, multi-agent collaboration, long-term memory management, externalized learning, and other frontier topics. Each project includes clear experimental objectives, detailed descriptions of the experimental content, and specific acceptance criteria, ensuring that students can master the key skills needed to build advanced AI Agent systems through hands-on practice.

The projects are divided into three levels by difficulty. Students are advised to choose appropriate projects according to their own background and improve their abilities step by step.

Project Index

Difficulty: Easy

1. Enhancing mathematical and logical reasoning ability using code generation tools
2. Natural language interactive ERP Agent
3. Werewolf Agent

Difficulty: Medium

1. Personal photo search engine
2. Intelligent video editing
3. PPT generation Agent
4. Book translation Agent
5. Agent that collects information from multiple websites simultaneously

Difficulty: Hard

1. A user memory that understands you better
2. Agent that uses a computer while talking on the phone
3. Computer operation Agent that gets more proficient the more you use it
4. Agent that can create Agents