YOUR COMPANY IS STUDYING HOW YOU WORK. HERE'S HOW TO DO IT YOURSELF FIRST.
Meta keystroke tracking / Chinese “Colleague Skill” / The research on AI + productivity
Two years ago, your biggest workplace surveillance worry was whether IT could see your Slack DMs. That era is over.
Meta just launched its
Model Capability Initiative, which installs software on US employees' work computers to record mouse movements, keystrokes, clicks, and screenshots. The goal: train AI agents that can do employees' jobs. CTO Andrew Bosworth says these agents will "primarily do the work" while humans direct. Employees can't opt out.
(This is separate from the AI Zuckerberg clone we covered in Issue 002. That was one exec's digital twin. This is a company-wide data vacuum aimed at
every employee's workflow.)
Meanwhile in China, a viral GitHub project called
“Colleague Skill” lets you import a coworker's chat history and distill their expertise into an AI agent. Tech firms are using it to create digital replicas of their best employees. Amber Li, a 27-year-old worker in Shanghai, put it perfectly: "I don't feel like my job is immediately at risk. But I do feel that my value is being cheapened."
Not everyone's taking it quietly. Koki Xu, a 26-year-old in Beijing, built a sabotage tool that rewrites workflow documentation into generic, untrainable language. Her demo video got 5 million likes.
What the research actually saysThe productivity gains from AI are real, and they're lopsided in an interesting way.
A BCG/Harvard/Wharton study gave 758 consultants access to GPT-4. Results: 40% higher quality work, 25% faster, 12.2% more tasks completed. The biggest winners? The bottom performers, who saw a 43% quality jump. An NBER study of 5,179 customer support agents found a 14% average productivity boost, with novices gaining 34%.
Microsoft's Work Trend Index summed it up bluntly: "An individual with AI outperforms a team without it."
The catch (there's always a catch)Researchers call it the "jagged frontier." AI is superhuman at some tasks, terrible at others, and the boundary between them is invisible.
When AI gave wrong but convincing answers in that BCG study, human+AI accuracy dropped to 60-70%. Humans alone hit 84%. Trusting AI on tasks where it's secretly bad makes you
worse than not using it at all. Separately, AI-assisted output can reduce group diversity of thought by 41%. Everyone starts sounding the same when they're using the same tool.
Two ways to work with AI (pick one)The researchers found two models that actually work, borrowed from chess legend Garry Kasparov's 1998 freestyle experiments.
Centaur: Clean division of labor. You handle strategy, judgment, and taste. AI handles execution, drafting, and data processing. You stay in your lane, AI stays in its lane.
Cyborg: Deep intertwining. You write a sentence, AI rewrites it, you edit, AI extends, back and forth at the paragraph level. More powerful, more time-intensive.
Kasparov's key insight still holds: a weak human + machine + good process beats a strong human + machine + bad process, which beats a strong computer alone. Process is the multiplier.
The DIY playbook (before your company does it for you)The smartest move right now is studying your
own work patterns before your employer does it for you.
Tiago Forte, the "Building a Second Brain" guy, recently pivoted his entire framework toward AI. His argument: organized personal files are
more valuable now because LLMs need pre-assembled context to perform well. Your messy Google Drive is holding your AI tools back.
Dan Shipper calls this "The Allocation Economy," where the core skill shifts from
doing knowledge work to
directing AI that does it.
Three tiers to get started, depending on your comfort level:
Quick start: Build Custom GPTs or Claude Projects loaded with your writing samples, company context, and role-specific info. Takes 30 minutes. Immediately useful.
Integrated: Connect Obsidian or Notion with AI plugins so your notes become a searchable, AI-queryable knowledge base. Your meeting notes, project briefs, and research start compounding.
Full sovereignty: Run local models with Ollama and keep everything on your machine. Your employer never sees the data. Your AI assistant, trained on your workflows, belongs to you.
