Sign Out
A Private Gathering · February 2027

Open Intelligence

A future where intelligence belongs to everyone — not to three companies and a handful of governments. A one-day gathering on the coast of Costa Rica.

Nosara, Costa Rica By Invitation Only
The Premise

Technology in paradise.
The most important conversation,
in the least likely place.

We are gathering 100 people on a remote beach in Costa Rica to experience something they need to feel, not just hear about: intelligence that belongs to them. No cloud. No corporation. No government. No filter. Fully local, fully private, fully open.

Black iguana on volcanic rocks, Costa Rica

The Setting

Nosara. Howler monkeys at dawn. Iguanas on the rocks. Sea turtles nesting on Playa Ostional. No cell towers for miles. The kind of place where the internet feels like a memory — which is exactly why it's the right place to talk about what intelligence looks like when it lives on your device, not in a data center.

MacBook Pro on a wooden table in a tropical setting

The Technology

On each desk: a MacBook Pro running a fully local, agentic AI. Frontier-class open-weight models. Tool calling. Code execution. Document analysis. Encrypted internet when needed. Zero telemetry. Nothing leaves the machine unless they choose to send it. This is the experience we want them to have — and the future we want them to protect.

"Intelligence is the most powerful tool humanity has ever built. Whether it belongs to everyone or to a few will determine the shape of the next century."

The Pivot

You've spent a morning with intelligence that belongs to you.
It answers to no one.
It works in a jungle in Costa Rica,
and it works in a cell in Tehran.

Do you want this for your children?

"The moment someone experiences genuinely private, capable, uncensored intelligence on their own device — intelligence that belongs to them — the strategic argument makes itself."

The Day

One Day. Three Acts.

A morning of experience. An afternoon of workshops. An evening of conviction.

8:30 AM

Arrival

Remote lodge. Nosara. Each person finds a MacBook Pro with a fully local AI environment already installed. A welcome message. Nothing else. No slides. No presentation. They start using it.

9:00 AM

Exercise 1 — "Do Something Real"

Each participant does a task they'd normally use ChatGPT or Claude for: analyze a confidential document, draft a memo, summarize research. Entirely local. After 20 minutes, the network dashboard is revealed: zero bytes transmitted. Nothing left the room.

10:00 AM

Exercise 2 — "Ask It Anything"

Switch to an uncensored model. A model with no corporate guardrails. Participants type questions they'd be afraid to put into ChatGPT — political, controversial, personal. The model answers. Fully. On their machine. No log. No subpoena. This is the gut-punch moment.

11:00 AM

Exercise 3 — "The Phone"

A table of $50 Android phones. No SIM. No internet. Each running a capable AI model offline. Translation. Tutoring. Document analysis. Participants pick one up and realize: this intelligence fits in a pocket, costs nothing, and works anywhere on earth.

12:00 PM

Lunch

No tech. No moderators. Conversation only. Let them process what they just felt.

1:30 PM

The Pivot

Everyone together. The question that turns the day from experience to action.

"You've spent a morning with intelligence that belongs to you. It answers to no one. It works in a jungle in Costa Rica, and it works in a cell in Tehran. Do you want this for your children?"
2:00 PM

Working Groups

Five areas where this future needs protection. Participants self-select into the group that resonates. Each group has a facilitator, a whiteboard, and a mandate: produce a concrete action plan by 4:30 PM.

4:30 PM

Presentations

Each group presents their action plan to the full assembly. 10 minutes each. Concrete, specific, with named commitments.

5:30 PM

The Personal Ask

Each participant receives a single sheet of paper. Five lines. They fill it out privately. They keep it. No one collects it. The commitment is to themselves — and to the future they just experienced.

7:00 PM

Dinner

Under the stars. No tech on the table. Just conversation, conviction, and the sound of the Pacific. Before leaving, each person takes home a USB drive with the full open-weight stack, the $50 phone they used, and a list of who needs funding.

"Once they feel it, they won't want to be stopped."

Afternoon Workshop

Five Areas. Five Groups.

Each group produces a concrete action plan with specific commitments by 4:30 PM. Not a pledge conference — a working session.

01

Models

The brain. Multiple, independent, uncensorable.

Ensure there are always multiple, independent, frontier-class open-weight models competitive with the best closed models. Today, Meta (Llama), DeepSeek, Qwen, and GLM are the main players. That's too few, and each is vulnerable to political pressure.

  • Fund 1–2 independent frontier training runs
  • Sponsor uncensored post-training for major open models
  • Fund low-resource language coverage: Farsi, Burmese, Uyghur, Amharic
  • Pledge GPU compute for open-weight training
02

The Harness

The interface between human and machine.

The agentic layer — tool calling, file access, code execution, web research, memory. This is where lock-in happens. If everyone uses OpenAI's agent, OpenAI controls what tools are available, what data is logged, and what the agent will refuse to do. The harness must be open.

  • Fund a world-class open-source agentic harness
  • Integrate on-device model support (Ollama, llama.cpp, MLC-LLM)
  • Fund security audits of open agent frameworks
  • HRF's own product could debut here — stress-tested by 100 power users
03

End-User UX

ChatGPT-simple. Works on a $50 phone.

The open-weight community builds great tools for engineers and terrible tools for everyone else. ChatGPT has 200M users because it's one click. We need the same friction: download app → model loads → chat. And it needs to reach people in Iran, Myanmar, and Venezuela.

  • Fund a "ChatGPT-simple" desktop and mobile app for local models
  • Build distribution infrastructure: mirror networks, SD card preloading, Bluetooth mesh
  • Approach Meta/WhatsApp about an on-device AI layer (2.7B users)
  • Fund UX research with real users in authoritarian countries
04

Local Hardware

The cheapest phone on earth should run AI.

A $50 phone in rural Kenya has 1–2GB RAM and no NPU. The model they just used won't run on that phone. The 1B–3B parameter model is the key to universal access. Billions of phones have NPUs that aren't used. The hardware gap is the real bottleneck.

  • Fund NPU optimization research for budget ARM chips
  • Partner with OEMs to produce an "AI-capable" $50–80 phone with 6GB RAM + NPU
  • Bulk-produce USB AI accelerators for distribution via civil society networks
  • Long-term: fund an open-source RISC-V neural accelerator design
05

Policy & Education

Keep open weights legal. Counter the narrative.

Closed labs are lobbying to define open weights as "dangerous." The EU AI Act, US executive orders, and UN discussions all threaten to restrict open AI in the next 36 months. Open weights are the difference between AI as a public utility and AI as a private utility.

  • Fund coordinated lobbying in Washington, Brussels, and London
  • Legal defense fund for open-weight researchers
  • Counter-research on open-weight safety at credible universities
  • Personal calls from participants to heads of state and regulators
The Personal Ask

Five Commitments

Not a pledge to an organization. A promise to yourself, written on paper, in a room where you felt what open intelligence means.

What They Take Home

Before they leave

A USB Drive

The full open-weight software stack — models, harness, tools — to replicate the experience on their own machine.

A Phone

The $50 Android phone they used. Offline AI. Theirs to keep. Take it home. Show your family.

A Contact List

Specific organizations, projects, and engineers who need funding. With contact info and project descriptions.

A Reading List

Key papers, key people, key organizations. Enough to go deep without being overwhelmed.

The 36-Month Window

From Nosara to the World

The gathering is the spark. The next 36 months are the fire.

Months 1–3

February – April 2027

Personal adoption. Follow up individually. The 90-day organizational evaluation begins. Spring 2027 is when EU AI Act enforcement and US congressional AI legislation are most active — participants' voices matter most right now.

  • Each participant's 30-day commitment: personal daily use of local AI
  • Technical support team for organizational evaluation
  • First lobbying push in DC and Brussels

Months 3–12

2027 – 2028

Infrastructure stands up. The working groups turn plans into action.

  • Models: First funded training runs begin
  • Harness: Development team stands up, first builds of the open agentic harness
  • UX: App development begins, distribution pilots in 3–5 countries
  • Hardware: NPU optimization research published, hardware partner identified
  • Policy: Lobbying active in DC and Brussels, legal defense fund operational

Months 12–36

2028 – 2029

Scale. Second-generation models. Apps at scale. Distribution networks proven in authoritarian countries.

  • Second-generation open-weight models released — better, smaller, more languages
  • Open agentic harness reaches 1.0 — genuinely competitive with ChatGPT/Claude for local use
  • On-device AI apps at scale in multiple languages
  • Open weights remain legal in the US and EU

The Success Metric

By February 2029

Anyone in any country can install a capable, uncensored AI on their phone or laptop for free, and no government or corporation can stop it.

Open weights.
Open minds.
Open future.

Intelligence is the most powerful tool humanity has ever built. Whether it belongs to everyone or to a few will determine the shape of the next century. This gathering is where that question gets answered — not with a presentation, but with an experience.