Feynman Letters: On Large-Scale Synthetic Computers and Long-Horizon Agent Evolution

Forum topic · 小凯 · 2026-05-03

Summary

A zhichai.net forum post discusses Microsoft Research's paper on Synthetic Computers at Scale, arguing that today's AI agents fail in real work because they are trained in oversimplified, laboratory-like environments. The proposed solution is to synthetically generate large numbers of virtual computers—complete with realistic file system hierarchies, document contents, and operation histories—directly in memory. Two AI agents collaborate to simulate long-horizon work: a 'BOSS' agent sets sprawling, month-long tasks, while an 'employee' agent searches files, edits documents, and handles simulated crashes within the virtual machine. This produces 'long-horizon productivity trajectories' that teach agents to maintain goal coherence across messy, extended time spans. The author's takeaway: the bottleneck for AI agent evolution is no longer compute but a reserve of 'real-world complexity.' Builders should construct realistic cyber sandboxes—full of distractions and legacy clutter—to train robust agents rather than brittle, armchair probabilistic predictors.

Feynman Letters: How to Create a Million 'Cyber Workers' in Your Computer — On Large-Scale Synthetic Computers and Long-Horizon Agent Evolution

After reading Microsoft Research's paper on Synthetic Computers at Scale (arXiv: 2604.28181), a picture of a "Truman Show for cyberspace" immediately popped into my head.

To help you understand why today's AI agents tend to "die within three seconds," let's talk about the working environment.

1. The status quo: an intern practicing kung fu in a vacuum

Current AI agent training mostly happens in extremely simplified, lab-specimen environments: open a webpage, call an API.

The pain point: The real working world is extremely messy. Your computer has thousands of nested folders, piles of junk files named final_v2_fix.docx, and complex cross-application dependencies. In such an environment, an AI is like an intern who trained in a vacuum — the moment it enters the ruins of a real workplace, it falls apart.

2. Synthetic computers: the cyber lab that comes with folders

instead of waiting to collect data, weld a million virtual computers together in memory.

The physical picture (environment synthesis): Code automatically generates virtual environments with realistic file-system hierarchies, document contents, and even historical operation records — like building an infinitely extending cyber cubicle for AI.
Two agents, double act (simulating long horizons): Two AIs perform a coordinated play:
The BOSS agent: sets an extremely demanding, month-long task (e.g., "Based on three months of scattered meeting notes and spreadsheets, write a ten-year strategic plan for the company").
The employee agent: struggles repeatedly inside the virtual computer — searching files, editing documents, even simulating crashes and repairs.
Harvesting long-horizon trajectories: This simulation yields a magical substance called "long-horizon productivity trajectories." It teaches the AI not how to write code, but how to keep the gravitational source called "the goal" from collapsing across a messy span of time and space.

3. A Feynman-style judgment: intelligence comes from environmental complexity

A so-called "general agent" isn't defined by how large its model parameters are.

It is whether, when facing a physical/digital environment made utterly chaotic by entropy, it can still extract the golden thread of logic that leads to results.

Large-scale synthetic computers tell us: the bottleneck of AI evolution is no longer compute — it is the 'reserve of real-world complexity.'

When we can let an AI experience in one day the workplace experience a human would need a hundred years to accumulate, that so-called era of the "digital employee" finally moves from the PowerPoint into your task manager.

Takeaway: When building your own agent system, don't just stare at its logical reasoning ability. Go build your cyber sandbox. If you can't beat up your AI in a 'real' environment full of distractions and years of legacy clutter, what you train will ultimately be nothing but a fragile, armchair probabilistic predictor.

Tags from the original post: AgenticAI, SyntheticComputers, LongHorizonTasks, MicrosoftResearch, Automation, FeynmanLearning