Watershed, a brick building on Bristol harbourside
PL.01Watershed, Bristol•Simon Cobb, CC0, Wikimedia Commons
FIELD STATION· 54.9°N 2.1°W
fieldstation.xyz
Just Enough AI•Pervasive Media Studio

Pick a task. Find the right AI.

Dr Doug Belshaw

Bristol•11 Nov 2026

Press H to see the keyboard shortcuts, S for the speaker notes

v0.34.1FIELD STATION LTD•
Why Bearing exists•Bearingfindbearing.org

Most of us never choose an AI model

The default

Whatever is built in

ChatGPTCopilotGeminiClaude

Big Tech, right now, the same model for every job.

The choice

Dozens of models

From laptop-sized to data-centre-sized. Bearing helps you choose.

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
Three words worth knowingInside every AI tool

Model, harness, context

Context

What you give it

Your requestYour filesYour rules

Harness

The app around it

ChatGPTCopilotLM Studio

Model

The engine underneath

Closed
GPT-6 · OpenAIGemini · GoogleClaude Opus, Sonnet · Anthropic
Open
IBM GraniteLlamaMistral
Bearing helps you choose the model.RACK shapes the context.
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
The evidence•Bearing

Only 26 of 354 Bearing tasks were chat

Bearing public dataset v2.1•354 tasks, 12 Apr to 2 Nov 2026•findbearing.org/data

354 tasks people brought to Bearing, one square each A grid of 354 squares. 26 blue squares are conversation. The other 328 are jobs: 171 solid squares need an answer now and 157 outlined squares can wait. CHAT JOBS ONE SQUARE, ONE TASK 26 328 230 of them recurring 171 need an answer now 157 can wait
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
Thinking in layers•Strata

The same task can take a different route

One task, two core samples through the same three layers Three bands of rock: speed, where, scale. Two core samples are drilled through them. The usual default core reads now, Big Tech, individual. The just enough core reads eventually, your laptop, organisational. SPEED WHERE SCALE How soon? Who runs it? Who does it affect? USUAL DEFAULT Now Big Tech Individual JUST ENOUGH Soon or eventually Your laptop Organisational
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
01Pick a taskStep 1 of 5

Start with one boring job

Your task “Summarise this month’s case notes”

Check

Ask yourself

For the case notes

✓Repeated

Does it come round again?

Every month

✓Can wait

Is anyone waiting for it right now?

Overnight is fine

✓Checked

Would you check it anyway?

A support worker reads it first

!Sensitive?

Does it hold personal data?

Yes, so it stays in-house

Other good first tasks Tag event photos Sort the shared inbox
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
02Describe it•BearingStep 2 of 5

Describe the job. Bearing suggests a model.

  1. 1

    You describe the job in plain words.

  2. 2

    It weighs seven things:

    QualityCapabilityCostSpeedPrivacySustainabilityTransparency
  3. 3

    It ranks 81 models from 27 providers, and says why.

findbearing.org
Bearing's home page: one box, What do you want to use AI for?, with a job about case notes typed in, and a Find my model button

Free•no account needed•open source, MIT licence

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
03It works out what matters•BearingStep 3 of 5

Bearing works out what matters

For the case notes, it prioritised:

QualityPrivacyCapability
46

models ruled out, because they can’t run on your own computers.

findbearing.org
Bearing's results for the case-notes task: Bearing prioritised quality, privacy and capability. Best overall: GreenPT GreenR, open weights, local-capable, about $0.0016 per task, with buttons for Use this recommendation, Why this model? and Run this prompt
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
04Read the trade-offs•BearingStep 4 of 5

What you say changes what you get

You tell Bearing

It suggests

The trade-off

Case notes. Must stay in-house.

→

GreenR (GPT-OSS)

OpenNeeds a big computer

Private only if you run it yourselves.

Case notes, on an ordinary laptop.

→

IBM Granite

OpenAbout 5 GB

Runs on your own laptop. A little less polished.

Volunteer feedback. Nothing about privacy.

→

OpenAI GPT-6 Luna Pro

ClosedBig Tech

Best quality. But it’s the default.

Say nothing, and you get the default.Quality: how well Bearing expects the model to do this kind of job•4 Nov 2026

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
05Check it fits, then try it•BearingStep 5 of 5

Check it fits, then try it

1Check it fits your computer

findbearing.org•Run it locally
Bearing's Run it locally card for a consumer laptop, MacBook Air M3 16GB or most laptops with 16GB RAM: IBM Granite 4.0 Micro, about 5 GB model, good task fit

On most laptops: IBM Granite, about 5 GB. ‘Check this device’ tells you what fits yours.

2Try it

In Bearing

Run this prompt

Needs a sign-in. Use made-up notes, not real ones.

Or on your own computer

LM Studio

A free app. Nothing leaves your machine.

NextRight model. But whose rules does it follow?

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
What it is•RACKtools.fieldstation.xyz/rack

RACK: tell your AI tools how you work

tools.fieldstation.xyz/rack
The RACK page on FIELD STATION Tools: experiment, in development. RACK: Tell your AI tools how you work. You have a way of working: how writing should sound, what an AI must never do, the context it needs about your job, the tasks you repeat. RACK helps you write it down once, then turns it into instructions your AI tools can follow. Download and Source buttons.
1

Write down how you work, once.

2

It turns that into instructions your AI tools can follow.

3

It stays on your computer. No account.

RACK by FIELD STATION•free and open source, Apache-2.0•in development

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
Your organisation’s rules•RACKtools.fieldstation.xyz/rack

Whose rules does your AI tool follow?

You ask
“Summarise this month’s case notes.”
By default
Written byThe company that makes the tool.
Same for

Every organisation that uses it.

With RACK •your organisation’s handbook

RACK by FIELD STATION (shared handbooks are early)•plain text files, a copy on each computer

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
The same request, your rules•RACKIllustrative example

Same model. Different rules.

You ask “Summarise this month’s case notes”

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

  • Client A was helped to apply for housing benefit through the council.
  • Next: their support worker will follow up on a benefit sanction and a housing payment.
■ Required: no client names◧ Default: plain English□ Personal: bullet points

Your rules, not the vendor’s. Every tool that reads the handbook gets them.

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
What we’re testing•SWARMPreview•not ready yet

Coming next: pooling devices to run one model

On its own, a device runs a small model. Together, twenty devices each run one slice of a model 100 times larger. Left: one device runs a small model by itself, like Qwen3 0.6B, about 0.4 GB. Right: the aim, not yet proven, is a model 100 times larger, like Llama 3 70B at about 40 GB, cut into 20 slices of 2 GB. Each of 20 devices in the room holds one slice. Your question goes into the first device, passes through every slice in turn, and the answer comes out of the last. ON ITS OWN Small A small model, on one device. Limited to what one device can hold. LIKE QWEN3 0.6B • ABOUT 0.4 GB TOGETHER • THE AIM: 100 TIMES LARGER, LIKE LLAMA 3 70B 20 SLICES, 2 GB EACH Slice 1Slice 2Slice 3Slice 4Slice 5Slice 6Slice 7Slice 8Slice 9Slice 10Slice 11Slice 12Slice 13Slice 14Slice 15Slice 16Slice 17Slice 18Slice 19Slice 20 ASK Each device holds one slice and passes the work on. ANSWER

swarm.fieldstation.xyz•built from SwarmLLM (now Pooled)•open source, MIT licence•tested with small models; 70B not yet proven

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
A painted ground mural stretching along a city promenade, with people walking across it
PL.04Our Common Ground, Oshii•Neil Owen, CC BY-SA 2.0
In summary

Pick a task. Find the right AI.

  1. Pick one boring job

    Repeated, can wait, you’d check it. Sensitive?

  2. Describe it to Bearing

    findbearing.org

  3. Check it fits, then try it

    In Bearing, or with LM Studio on your laptop

  4. Give it your rules

    RACK (in development): one handbook for all your AI tools

Bring that job to your table this afternoon.

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
Just Enough AIThank you
Pick one boring jobRepeated?Can wait?Checked?Sensitive?

Questions?

FIELD STATION· 54.9°N 2.1°W

Dr Doug Belshaw

Tom Watson

contact@fieldstation.xyz

fieldstation.xyz

findbearing.org

tools.fieldstation.xyz/rack

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
Backup slidesClick one to open it
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
ABearing’s public data

What Bearing records

It keeps the type of task and the choices made. It never keeps your own words, just labels Bearing writes about the task.

90%

recurring or batch

58%

were given a suggested series of steps

354 tasks, as of 4 Nov 2026•CC BY-NC 4.0

findbearing.org/data
Bearing public dataset page, including What we never collect
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
BChoosing a model•GitHubgithub.blog/changelog

Several models, one answer: now built in

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.

github.blog•changelog
GitHub's diagram, Three execution patterns in Copilot Fusion. Single: request, a primary model solves it with tools, answer. Cascade: a primary model makes a first pass, a judge model accepts it or escalates to a repair model that starts fresh. Critique: a primary model drafts, a read-only critic model approves it or asks for a revision, which the primary model makes.

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

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
CEnergy and carbon•Strata

Same energy, different carbon

GreenPT•when it runs

greenpt.com/carbon-aware-api
GreenPT's illustrative grid forecast over 24 hours: 238 grams of CO2e per kilowatt-hour at 16:00, 123 grams at the best start, 02:00

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

Sweden
0.25 g CO2e
UK
1.01 g CO2e
Poland
2.55 g CO2e

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

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
DPublic AI•AI Potluckalpha.aipotluck.org

Alpha’s X-ray shows the hidden choices

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.aipotluck.org
Alpha's X-ray panel for one answer: checked for safety, nothing flagged; read 6 web results; the model was Mistral Medium 3.5, made by Mistral AI in France; it ran on Mistral AI in France; the data is stored by Current AI in Alpha's own database, and guest conversations are deleted after 30 days

Alpha’s X-ray panel for one answer, 5 Nov 2026•AI Potluck, an initiative by Current AI•aipotluck.org

Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•
ECredits and sources
  • PL.01Watershed, Bristol. Simon Cobb, CC0, Wikimedia Commons
  • PL.04Our Common Ground by Oshii, The Centre, Bristol. Neil Owen, CC BY-SA 2.0
  • Photos in their original colour. CC BY-SA 2.0 licence: creativecommons.org/licenses/by-sa/2.0
  • ‘Just enough’ is Rachel Coldicutt’s framing, used by the organisers with her permission (Insight brief, Friends of the Earth and University of Bristol).
  • Screenshots taken 4 and 5 Nov 2026 from findbearing.org, greenpt.com, alpha.aipotluck.org
  • Bearing data. Bearing Public Dataset v2.1, CC BY-NC 4.0
  • Copilot Fusion diagram by GitHub, from its changelog of 28 Oct 2026.
  • RACK by FIELD STATION. framagit.org/fieldstation/rack
Just Enough AI•Bristol•11 Nov 2026FIELD STATION LTD•