NDAs, AI Gold Rush, and Living in Brazil
The Call
West Coast number. Recognizable domain in the caller ID.
"We've been following your work. Want to talk?"
If you grew up on cyberpunk comics and dystopian future fiction, this is the call. The tech overlords noticed. They have projects. They need people who understand the weird intersection of AI, culture, and building things.
You sign the NDA before the second meeting.
The NDA Theater
Twelve pages. Standard boilerplate plus AI-specific clauses. Covers product roadmaps, training data, internal strategy, and basically everything else you might say out loud.
What you can say: "I work with AI companies." "It's interesting." End of list.
Here is the reality nobody mentions. They give this same NDA to thousands of contractors. Company is two years old, maybe three. Those contractors post on Blind under anonymized company complaints, talk on Discord servers, share on private Slack channels, write Medium articles "inspired by experiences," and give conference talks about "techniques I've seen."
The NDA is theater.
Not useless—they will enforce it if you directly leak product details. But thousands of people sign it. Information flows anyway, just indirectly. What actually stays secret: specific model architectures, maybe. Exact training data sources, sometimes. Near-term product launches, until someone leaks to TechCrunch.
What leaks immediately: how chaotic it actually is, what the culture is really like, which teams are disasters, what is actually being built in rough terms, who is getting fired, what the real priorities are.
The company wisdom: if a thousand people know it, it is not a secret. It is just not officially confirmed.
Brilliant People. Seven-Foot Bonghits. Ship Anyway.
The appeal is real. Brilliant people—legitimately the smartest you will meet. Cutting-edge tech, you are building the future. Compute budgets that would make universities weep. You see models before public release, play with capabilities that do not exist yet, understand what is coming before the world knows. A mission that is maybe world-changing, maybe civilization-ending, definitely something.
The reality is also real. Seven-foot bonghits called sprints. Ship faster than you can think. Five projects, all urgent, choose two to fail at. Everything changes at Thursday standup. Information firehose set to drown.
Survival mode: become automaton. Process everything. Sleep optional. Sanity negotiable.
You either adapt to machine speed or wash out in three months.
The Brazil Problem
Not Brazil the country. Brazil the movie. Terry Gilliam's bureaucratic nightmare where everything works and nothing makes sense.
You are living in a science fiction future: AI that actually works, models that pass Turing tests, technology that looked impossible five years ago, brilliant people solving hard problems. Inside a corporate structure: meetings about meetings, Slack channels that spawn sub-channels, documentation outdated before it is written, process optimization processes, agile sprints that feel more like marathons.
The contradiction is building revolutionary technology using conventional corporate systems designed for widget manufacturing.
Nobody Knows What They're Doing. The Field Is Two Years Old.
Technical skills: prompt engineering at scale, model evaluation, safety testing, production deployment, failure analysis when models do unexpected things.
Real skills: absorb information while drowning, say yes to the right projects and no to the career-ending ones, context-switch five times per hour without breaking, ship broken code that works well enough, iterate faster than you think possible.
The actual lesson is that nobody knows what they are doing. These companies are two years old. Best practices do not exist. The people who "invented the field" are making it up as they go. You are building the plane mid-flight. Engine is on fire. Passengers are complaining. Ship anyway.
The Volume Problem
Seven-foot bonghits equal sprints. The volume of information, decisions, and iterations packed into two-week cycles.
Monday 9am: kickoff call. New safety eval framework. Scope: "comprehensive testing protocol." Timeline: two weeks.
Monday 2pm: Slack notification. Previous project needs emergency fixes. Production issue. "Quick patch, shouldn't take long."
Tuesday morning: original project scope changed. "Actually, can you also include adversarial testing?" Sure. Timeline unchanged.
Wednesday: emergency meeting. Different team needs your input on prompt injection research. "Just 30 minutes." Turns into three hours.
Thursday: demo prep for Friday. You realize requirements shifted on Tuesday. Rebuild half the framework overnight.
Friday 10am: demo. Goes well. "Great work, let's expand this to..."
Friday 3pm: sprint retro. New project assigned. Starts Monday. Previous two projects still ongoing.
The only way to handle it: jump in, become the automaton, learn everything, process faster, satisfy the masters. Or drown in the firehose.
