Whose AI is it anyway?
Doug Belshaw
Tom Watson
Press H to see the keyboard shortcuts, S for the speaker notes

Doug Belshaw
Tom Watson
Press H to see the keyboard shortcuts, S for the speaker notes
Context
What you give it
Harness
The app around it
Model
The engine underneath
Bearing public dataset v2.1•354 tasks, 12 Apr to 2 Nov 2026•findbearing.org/data
GreenPT•when it runs

Waiting helps too. GreenPT’s own illustration: a 45-minute job due by 08:00 has about half the carbon if it starts at 02:00.
EcoLogits•where it runs•AI writing one email
4.17 watt-hours per email, in all three
Same email, same energy: ten times the carbon in Poland as in Sweden.
Estimates•CO2e: all greenhouse gases, counted as CO2•calculator.ecologits.ai, from the French non-profit Code Carbon•4 Nov 2026
You describe the job in plain words.
It weighs seven things:
It ranks 81 models from 27 providers, and says why.

Free•no account needed•open source, MIT licence
You tell Bearing
Bearing’s top pick
The trade-off
Case notes.
Must stay in-house.
GreenR (GPT-OSS)
Gain: privacy, openness, sustainability.
Lose: a little quality, and it costs a bit more.
Case notes,
on an ordinary laptop.
IBM Granite 4.0 Micro
Gain: cheapest, most open, runs on a laptop.
Lose: some quality.
Volunteer feedback.
Nothing about privacy.
OpenAI GPT-6 Luna Pro
Gain: the best quality.
Lose: openness, speed and sustainability.
Quality: how well Bearing expects the model to do this kind of job, from published tests and its own judgement•scores of 4 Nov 2026
swarm.fieldstation.xyz•built from SwarmLLM (now Pooled)•open source, MIT licence•tested with small models; 70B not yet proven
RELAY is how we want work to move between devices when you leave.
Where the work goes
Your own devices
↓
A co-operative server
↓
Big Tech
Only if needed, and only after asking you.

RELAY•now
Done when you were away
Consultation summary is ready: 40 of 40 responses.
Read summary

| 1925 | Court reporter, pen and notebook |
| 2026 | AI note-taker, transcript and summary |
Both give you one version of one meeting, without the wider context.
Neither links to the rest of the organisation
Every organisation that uses it.
RACK by FIELD STATION (shared handbooks are early)•plain text files, a copy on each computer
Without a handbook•the vendor’s defaults
Sarah Jones (DOB 03/04/1981) was signposted to HB via the LA following a s.188 assessment. SW to f/u re UC sanction and DHP application.
With your handbook•made with RACK
Your rules, not the vendor’s. Every tool that reads the handbook gets them.
“Co-operative models could work for smaller-scale infrastructure. Municipal ownership could work for community datacentres.”
Peter Wilson•4 Feb 2026•his own views
Co-op AI today
inference.coop
A member-governed co-op for AI. Open models, from $5 a month.
co/core
An experiment: members share spare computing power on their Macs to run open models for each other.
“AI Growth Zones: Gift Horse or Trojan Horse?”•peteski22.github.io•inference.coop•cocore.dev

Choices have consequences.
Try it: findbearing.org
Which of your tasks could wait until soon, or eventually?
Doug Belshaw
Tom Watson
contact@fieldstation.xyz
fieldstation.xyz
findbearing.org
tools.fieldstation.xyz/rack
It keeps the type of task and the choices made. It never keeps your own words, just labels Bearing writes about the task.
recurring or batch
were given a suggested series of steps
354 tasks, as of 4 Nov 2026•CC BY-NC 4.0

GreenPT’s carbon-aware service picks a cleaner time to run, before your deadline.
They say it can cut a job’s carbon by up to 50%. Early access.
Find out more
greenpt.com/carbon-aware-api


Computers you run
CodeCarbon measures the power, then applies your grid’s carbon.
AI services you call
EcoLogits estimates each request. Its calculator is on slide 6.
Free, open source•mostly for developers
codecarbon.io
Panel and synthesiser
Several models answer. One combines them into a final answer.
Vote
Several models answer. The most common answer wins.
Chain
One model drafts, another critiques, a third revises.
Retake
Drafts are checked. If none passes, the models try again.
Route
Each question goes to the model that suits it best.
Cascade
Start with a small model. Call a large one only if needed.
Patterns from the ScreamingFace toolkit, backed by OpenMined, as shown by Dave Buckley at MozFest 2026•redrawn by FIELD STATION•screamingface.ai
GitHub Copilot Fusion picks the models for each prompt. You pick the effort.
Three ways it combines them: one model on its own, a cheap draft checked by a judge, or a draft reviewed by a critic.
On all Copilot plans since 28 October 2026. You pay for each model it uses.

Diagram: GitHub, “Route and orchestrate models with GitHub Copilot Fusion”, 28 Oct 2026•github.blog/changelog

Freehold: Peter Wilson’s project. Developer preview.
freehold.earth

Mozilla.ai: open tools such as any-llm and llamafile.
docs.mozilla.ai
Mozilla.ai’s Otari: one door to many models, with a spending cap for each team.
otari.ai
An assistant you own.
Runs open models on your own computer.
Chats sync between your devices, encrypted.
Backed by
Mozilla Foundation. An anchor project in its Human-Scale AI incubator, with $300,000, says Tiles.

Early alpha. Desktop app is a nightly build. Mac (Apple Silicon) and Linux only.
tiles.run
AI doesn’t have to be a chat box.
A new kind of interface for your AI agent.
It builds dashboards on the fly, to suit the job.
At MozFest it was run entirely by voice.
Works with any model, including one on your own machine.
Open source, MIT licence•agent runs on Mac or Linux
television.run

A free AI assistant from AI Potluck, an initiative by Current AI, a non-profit.
People bring the open models, datasets and tools. Its X-ray shows what happened behind each answer.
alpha.aipotluck.org

Alpha’s X-ray panel for one answer, 5 Nov 2026•AI Potluck, an initiative by Current AI•aipotluck.org
Drawn by FIELD STATION from Mary Stevens’ insight brief•solid: more leads to more•dashed: more leads to less

Exmouth, 2023. Servers sit in oil. Their heat warms the pool water.
Lots of small sites, owned locally, could cope better than one giant one.
Source: AFP via Tech Xplore, June 2023