Local first

Start with a simple question. What should keep working if the internet goes away? In much of the forest, connection is slow, expensive, shared or simply absent. If the database, the assistant or the payment app only works through a distant server, the community does not really hold the tool.

At Parque das Tribos the morning before the workshop went into cables. We set up a Starlink dish and a Raspberry Pi, a computer about the size of a deck of cards, with Samuel and Eduardo, the two local champions who would run it. When we explained the setup to the group, we stressed one point above all: everything on this device belongs to the community. Taina's records live on it. A local champion decides who can use it. The community's data council decides what is shared and what stays private. The data sits on that device and not on GainForest's servers.

When we left, the kit stayed: the Pi with Taina running on it, the Starlink connection, AudioMoth recorders and phones. Samuel and Eduardo know it inside out. The point was always that the work would carry on after we flew home.

Local first does not mean refusing the cloud. It means the first copy, the first decision and the first chance to look at the data stay with the people it comes from. Tang makes the same argument about AI. Smaller, local systems "can be corrected more quickly, audited more meaningfully and aligned with local context." She also adds the caveat: "This does not mean every system must be local."

Taina shows the trade-off honestly. In 2024 she ran on Meta's open Llama 3.1 model on local machines, with voice transcribed by a privately hosted Whisper service. Today's version, Pi-Tainá, keeps the records, drafts and member list on the community's Raspberry Pi, but sends each photo to a cloud model (Google's Gemini by default) to identify the species. That identifies far more species than a model small enough for a Pi. It also means photos leave the device for a moment. Neither setup is right for every community, and the choice belongs with the data council.

Frugal tools also break in frugal ways, so plan for it. Some lessons from Manaus:

• Free is not the same as accessible. New Telegram accounts in Brazil needed a R$7 SMS payment. Some participants hesitated or could not join, so we brought phones with Telegram already installed.
• Uploads fail in the forest. On our second trip we changed the Android app so drafts survive a dropped connection and come back after the app closes.
• Connection costs money every month. In GainForest's projects, payments to communities have helped cover Starlink subscriptions. Someone has to pay for it after the grant ends.

The test is practical. Can the community open its own data, run the core tasks, make a copy, see what was sent out, and switch the system off without losing everything? If so, the tool is close to soil. It can be looked after where it stands.