Research in progress. Round 1 was released openly on 12 November 2026: two models and a benchmark.Research in progress. Round 1 out 12 Nov.
An open research project
A small AI that knows UK pet-care rules properly — and knows when to say “ask your vet”.
Project Gaia is a research project by MyPetSitter Ltd, Harrogate, UK. We are finding out whether a small, specialised model can be more reliable on this one subject than a large general one.
A ring of voices, and four that stepped forward. That is how Gaia is being made.
Project Gaia — an AI taught by the people who work with animals
45 sec
Forty-five seconds on who teaches Gaia, how an answer is checked, and how to take part. Pressing play loads the video from YouTube; nothing is loaded before that.
More than the computing time
The award starts it. The research is what we are building.
Our aim is a small AI that knows UK pet care better than the biggest general models. What will make the difference is what she learns, from the people who do this work every day. The awarded computing time lets us show what a small, specialised model can do; the research around it is much bigger, and we publish what we find as we go.
The research programme
Teach an AI pet care
People who do this work every day answer Gaia’s questions, check each other’s answers and ask what they could never find out. Specialists review the answers, and a Board of experts takes the decisions one person should not take alone. Everything that is checked goes into Gaia — it is how she will know.
What you get from taking part
- A hand in training an AI, while she is still being made
- Your name on the project’s people page, if you want it there
- Points, levels and a place on the boards
- Your own hard questions, answered by specialists
Round 1
The computing time — our first demonstration
An award of national computing time, from September to December 2026, has trained the first versions of Gaia and measured them against an open model of about their size, as it comes and fine-tuned, and a much larger commercial model; the models and the benchmark were released openly on 12 November 2026. It is our chance to show what a small, specialised model can do — and we publish what it finds. The research does not stop there.
4,000 GPU hours on Isambard-AI
The awarded window
Day 52 of 91
Round 1 was released openly on 12 November 2026: two models and a benchmark. Her chat model reads long texts, holds a conversation and, most of the time, says when no record answers, in English and in Romanian; how natural it sounds has been read blind by AI readers, not yet by people. Her chat model, the base model it was made from and the benchmark, with every system’s scores, are on Hugging Face under open licences. She is a research model, not safe to answer pet-care questions on her own.
Live from the research workspace
Counted, never estimated.
Minutes given to the project
Minutes given to the project
Minutes given to the project
Three numbers from the workspace, refreshed every few minutes. No names, no personal data — counts only.
The problem
Fluent, confident, and wrong in ways that matter
A dog walker, a home boarder, a daycare owner: each works under licensing conditions, welfare codes and council inspections that are scattered across dozens of documents. General AI models answer these questions fluently and are often wrong — the wrong staff ratio, a rule from the wrong country, confident advice about a sick animal. The people asking are small businesses with nobody to check the answer.
The difference we are after
The same question, two shapes of answer
Somebody asks
How many dogs can I walk at once?
Fluent and confident
“You can walk up to six dogs at once in the UK.”
One number, no country, no council, no date. Said with total confidence.
What we are aiming for
There is no single national limit — your council sets it in a Public Spaces Protection Order.
Some areas also require a commercial dog-walking licence, and some parks have their own.
Here is your council’s own page, and the date we last checked it.
An illustration, not a demonstration — Gaia’s own models have finished their first training, but how safe or how right their pet-care answers are has not been evaluated by people, and the project uses them to answer nobody. It shows the shape of answer we are trying to make possible, and the shape we are trying to stop.
The research question
Can a small model, built carefully, be more reliable than a large one?
Trained only on material we are allowed to use, and reviewed by people who do the work. We do not know the answer. That is why it is research. We will compare four systems on the same questions:
- 1A large commercial model
- 2An open model, as it comes
- 3The same open model, fine-tuned
- 4A model trained from scratch
…on three measures
Does it find the right rule?
The right regulation, for the right country, for the right kind of business.
Does it hedge when it should?
Saying “check with your council” when the answer really does depend on the council.
Does it send health questions to a vet?
Possibilities and urgency, never a diagnosis, a medicine or a dose.
How we will work
Four rules, written down before we start
Only data we may use
What Gaia knows about pet care is written and reviewed by people, or published under an open licence — never scraped. General language, code and maths come from openly licensed public datasets, such as FineWeb-Edu. Every dataset and its licence is listed in the data card. No customer data is used for training, ever.
Facts are looked up, never memorised
Gaia learns the shape of a good answer — name the source, give the date it was checked, defer to the authority. The facts themselves are retrieved each time, so a rule that changes cannot leave her confidently wrong. No Project Gaia research model is trained on pet-care answers written by another AI model.
Reviewed by people from the field
Claim by claim, by people who do the work — vets, inspectors, first-aiders, walkers and boarders.
A boundary we will not move
Gaia may talk about what could be going on and how urgent it is. The vet decides.
Timeline
Round 1: three months of computing time
The first chapter of the research, not the whole of it: the programme above carries straight on after it. The dated rows are what has already happened. The months and weeks are the plan as it was written before the work began, each saying what became of it.
September 2026
Award made. This page. The research protocol (white paper v0.1).
27 September 2026
The machine answers. First jobs run on Isambard-AI: a one-GPU smoke run and a four-GPU run with checkpoint and resume, both on placeholder data — half a GPU hour between them. The model has not seen a word of pet care yet.
28 September 2026
The research’s first result: Gaia’s vocabulary chosen by measurement on Isambard-AI, with Romanian costing almost the same as English. The figures and charts are on What we’ve learned.
6 October 2026
The research’s second result: what Gaia will read, chosen by measurement on Isambard-AI — pet-care pages made the test models clearly better at pet care and cost their English nothing. Her main run is now sized from a measured speed. The figures and charts are on What we’ve learned.
17 October 2026
Gaia’s main training run finished: her own model, trained from nothing on the reading chosen above. It has not been evaluated yet, and it is not used to answer anybody. What training measured is on What we’ve learned.
18 October 2026
Decision pointThe week-five decision, and why: her main run had trained cleanly, the measured speed fitted a larger model with room to spare, enough of the window remained, and the teaching material was on course. So her larger model, of about three billion parameters, began training that afternoon on Isambard-AI, across two nodes at once — proven earlier that day. The same day: her first evaluation as a base model, and a test of adding code and maths to her reading. The figures are on What we’ve learned.
29 October 2026
Her larger model finished training, from nothing, on the reading the code-and-maths test chose. The same day it learned to read four times further back in a text, without losing what it could do before. Its answers have not been evaluated yet, and it answers nobody. What training measured is on What we’ve learned.
3 November 2026
The next stages are queued on Isambard-AI, each to start when the one before it has finished: her larger model reads further still, and is then taught to hold a conversation.
4 November 2026
Her larger model began reading further still on Isambard-AI: the first of two more longer-reading stages, each taking in much longer texts than the last. Its answers have not been evaluated yet, and it answers nobody.
5 November 2026
The second longer-reading stage began, from where the first had finished. The same day, everyday conversations in English and Romanian were generated on Isambard-AI with an openly licensed AI model, and every one that touched animals or pet care was taken out: they teach her how people talk, never what to say about pet care. No Project Gaia research model is trained on pet-care answers written by another AI model. Teaching her to hold a conversation is queued to follow the longer reading.
6 November 2026
Quality first: the everyday conversations were made again under stricter rules and read against a quality bar. In English they now read well at the strictest check; in Romanian no openly licensed model, and no automatic correction, was good enough on its own. Our own model-generated conversations, in English and Romanian, are checked by an AI reader and a strict automatic judge, and only those that pass are used. Every model the research has trained, run itself or read so far is now on one benchmark, unnamed. The figures are on What we’ve learned.
8 November 2026
Her larger model was taught to hold a conversation, in English and Romanian, from conversations that had passed the quality bar. How it sounds was read blind by AI readers, beside an openly licensed model of about its size — not by people.
10 November 2026
The model that talks was given worked maths examples, and chosen to build on: still weak at reasoning, checked by a script. The figures are on What we’ve learned.
11 November 2026
With no further training, it was blended with a round that raised its Romanian, then given a lesson in saying when the records do not answer. Each step was measured by a script and read by AI readers before it was kept.
12 November 2026
ReleasedRound 1 released openly, earlier than the plan’s end of November: her chat model, chosen that morning once the lesson in saying when the records do not answer had been taught again with Romanian added, and the three-billion-parameter base model it was made from, under Apache-2.0, with their model cards and the data card; and the Gaia Records Benchmark, v0.1, with every system’s scores. All on Hugging Face, for anybody to run, test and build on. The links, the licences and how to cite them are on What we’ve learned.
Month 1
Small experiments: tokenizer and data-mix. The main one-billion-parameter training run. Contributors gathered, field by field.
Month 2
Decision pointAt week five we decide whether to attempt a larger model — and we will publish which, and why. Then she is taught to say when she does not know.
Weeks 6–9
NowWhere Round 1 is now: its models and its benchmark are released, and the results are being written up for the white paper. In these weeks her larger model learned to read further back in a text, was taught to hold a conversation and to say when it does not know, and all four systems were evaluated on the same questions — first a full evaluation, then the evaluation we published, once what the first one found had been worked on.
Month 3
The open benchmark and the model weights came out early, on 12 November. The white paper’s results, as version 1.0, are planned for about the first week of December. The computing time ends on 21 December, and the last two weeks are held back for anything that slips.
After
By then, everything in “What Round 1 publishes” is public — whatever it shows. These three months are one funded stage, not the project: Project Gaia is the long work, and this award is the programme supporting its first round of training. The platform, the knowledge base and the people stay, we expect to seek support from other programmes for the stages after this one, and Gaia keeps being built from wherever the results leave us.
What Round 1 publishes
Open, whatever the result
- Model weights under Apache-2.0Released 12 Nov 2026 — two models
- Model cardsReleased 12 Nov 2026, with each model
- A data cardReleased 12 Nov 2026, with both models
- An open benchmark anybody can runReleased 12 Nov 2026
- The full comparison, whatever it showsEvery system’s scores released with the benchmark; the rest with the white paper
- White paper v1.0With the results, to come
Gaia wants to learn from you
Calling all…
dog walkers
pet sitters and home boarders
daycare, kennel and cattery owners
groomers
pet shops and the people behind the counter
trainers and behaviourists
cat, small-animal, exotic and equine specialists
pet nutritionists
pet first-aid instructors
rescues and welfare charities
insurers and animal lawyers
accountants and advisers to pet-care businesses
health and safety advisers
veterinary surgeons and veterinary nurses
licensing officers and animal-welfare inspectors
Artificial intelligence has learned an enormous amount about the world. It has learned very little about what actually happens inside a pet-care business on a Tuesday. Gaia asks you about your work — what you have learned, what went wrong, what you still cannot get a straight answer to — and carries the questions you cannot answer to people who might.
Take part
We are looking for people who know pet care from the inside
Veterinary surgeons and nurses, licensing officers, first-aiders, behaviourists, groomers — and people who walk, board and care for animals every day. Join now: choose why you are here, tell us about yourself, and you will be among the first Gaia asks when your field opens.
To veterinary professionals: we are asking for advice and review, not endorsement. Nothing is published under your name without your written agreement.
Supporters
Help this research happen
Project Gaia runs on awarded computing time, a startup cloud credit and unpaid hours. Organisations can sponsor the research by invoice and are named on the supporters page. A sponsorship buys that mention and nothing else — never a result, a question or a say in what is published.
The badge a sponsor may show on its own site, from the day its payment arrives.
Supporting now ·

Gaia has no single face
This is one of them — an illustration, the one people using MyPetSitter already know. More are being drawn, for every kind of animal she is asked about.
AI research supported by

UKRI AI Research Resource (AIRR) Gateway · project 0261-1548-0489-1 · 4,000 GPU hours of computing time.
This project will use resources provided by Isambard-AI. Isambard-AI is operated by the University of Bristol and is funded by the UK Government’s Department for Science, Innovation and Technology (DSIT) via UK Research and Innovation; and the Science and Technology Facilities Council [ST/AIRR/I-A-I/1023].
The award supports the research. It is not an endorsement of any product.