Carlos Puig

Entrepreneur in AI, hospitality & tech. Curiosity drives me. Cash is king, cashflow queen. I share what most founders won't.

[Carlos Puig @cpfdtz]

@Martinoleary @AlexFinn @PenguinWeb3 I tried and it’s really weird

[Carlos Puig @cpfdtz]

I spent close to 2 years living on an island in the Philippines.

No paved roads. Power cuts twice a week. Supplies by boat.

I ran my company from there while building a private villas resort in a 30-year-old typhoon zone.

Most people thought it was too remote to operate from. I thought the constraints were the point.

[Carlos Puig @cpfdtz]

Every AI agent we deploy hits the same wall. Edge cases. Things humans handled on gut. Nothing written down. The agent freezes or invents an answer. We ended up spending most of the project on process documentation, not the agent itself. That is where the real work is.

[Carlos Puig @cpfdtz]

300 people at BUNCH have been labeling AI training data for years. The models they helped train now handle the routine volume. The humans moved to edge cases and quality review. That is not a prediction. That is a live experiment running inside one company right now.

[Carlos Puig @cpfdtz]

300 people at BUNCH do data labeling, content moderation, AI safety work for AI companies right now. That work exists because AI models need humans to supervise output. It does not disappear when agents scale. It becomes more specialised. The floor rises for the humans who can navigate both sides.

[Carlos Puig @cpfdtz]

More often than not, when something feels scary, is the right signal for you to do it.

[Carlos Puig @cpfdtz]

A calm mind is totally underrated these days

[Dave Kline @dklineii]

Major career cheat code: Be easy to work with. Calm when others panic. Positive in the face of pessimists. Reliable. Consistent. It doesn't take talent. It just takes intention. The world bends toward people who make everyone around them better.

[Carlos Puig @cpfdtz]

Three years ago i couldn't explain what an AI agent was to a client without losing them in the first sentence.

Now they come to me already sold on the concept. Already scared they're behind.

Nothing about the technology changed that. The market caught up on its own.

Being early is uncomfortable. Then it becomes an advantage you didn't have to fight for.

[Carlos Puig @cpfdtz]

Everyone thinks AI agents fail because the model isn't good enough. In my experience: 90% of failures are data problems. The agent is fine. The data it's working with is messy, outdated, or siloed in a system nobody's touched in 4 years. We spend the first 2 weeks of every project on data audit alone. Clients are always surprised. We're never surprised at this point.

[Carlos Puig @cpfdtz]

Everyone's obsessing over which model to use. Which agent framework. Which tool.

Wrong obsession.

The harness is everything. How you constrain the agent. What context it gets. When it hands off. How it fails gracefully.

A mediocre model in a good harness beats GPT-4 running loose every time.

[Carlos Puig @cpfdtz]

is this normal?

[Carlos Puig @cpfdtz]

Entrepreneurship is a lonely game that can’t succeed without a proper support mechanism around you, mainly from your loved ones. Proof me wrong.

[Carlos Puig @cpfdtz]

i grew up watching my dad build from nothing. He never talked about it as hustle or grind. He just worked, and the work compounded. i didn't understand that until my late twenties. By then i'd already burned through a startup, lived in Jakarta for three years, and started BUNCH in Manila. The lesson was the same one he'd been showing me since i was a kid. Stay in the game long enough for it to compound.

[Carlos Puig @cpfdtz]

i used to think the hardest part of building a company was the idea. Then i thought it was the product. Then i thought it was the team. Now i think it's the transition. The moment where you have to stop doing the work and start building the system that does the work.

[Carlos Puig @cpfdtz]

This is how deploying agents in a real business work:

Week 1: the CEO loves the demo. Week 2: the ops manager says "our process doesn't work like that." Week 3: you're rewriting everything.

The process on the pitch deck is never the real process. The real one lives in someone's head, a WhatsApp group, and a spreadsheet from 2019.

So you start there. You sit with whoever actually does the work. You map what they really do, not what the SOP says.

Then you find the one thing that breaks their day. The boring, repetitive thing they hate.

You automate that. Just that. Get it working in production, with real data, with real edge cases.

It takes 3-4 weeks. Not 3-4 months. You move fast or you lose the window. Attention dies fast inside a company.

If it works, they trust you with the next thing. That's how you go from one agent to five.

[Carlos Puig @cpfdtz]

Every company has two versions of its operations. The version in the handbook and the version that actually runs. After 50+ deployments i can tell which companies are healthy within the first meeting. Healthy: the CEO describes how things work and the ops manager corrects him on two details. Unhealthy: everyone agrees with the CEO and the real process lives in three people's WhatsApp groups.

[Carlos Puig @cpfdtz]

There's a specific moment in every OpenClaw session.

Everything works fine. Then you need to actually do something. And suddenly the browser times out.

It's like it's just been waiting for you to ask something real.

[Carlos Puig @cpfdtz]

Listened to the All-In episode on Anthropic's $30B run rate.

One number that stuck: they went from $9B to $30B ARR in under 4 months. OpenAI took years to hit $9B.

From where i sit deploying Claude for actual businesses: the revenue makes sense. It's the only model that comes back with fewer surprises when you run it on real ops workflows, week after week.

[Carlos Puig @cpfdtz]

Good operators fix problems. Great ones build systems so the same problem stops mattering.

[Carlos Puig @cpfdtz]

i always ask clients: who actually owns this process right now?

Usually there's a pause. Then someone names a person who's been there 11 years and hasn't documented anything.

That's not an automation problem. That's a knowledge risk. The agent just made it visible.