An open, decentralised initiative
AIFG connects the people building open source AI, autonomous agents and robotics, funds the proofs of concept that would not otherwise get built, and publishes what comes out of them. Kuala Lumpur first. Southeast Asia next. Anywhere after that.
Move to steer · click to fire a signal
Free to join, and free to leave. We convene, we fund and we publish — we do not lobby, certify or sell anything.
Why now
The frontier moved from models to agents. Systems now hold funds, execute transactions and call other systems with no human between intention and consequence. The tooling is a year old. The accountability infrastructure does not exist.
The talent is distributed, the opportunity is not. Southeast Asia has deep engineering capacity and a fast builder culture. What it lacks is the room where a robotics researcher, an agent developer, a regulator and a family office end up at the same table.
Capital and policy are both looking for somewhere to land. Patrons want AI exposure that is not a venture bet. Governments want capability-building that is not a procurement contract. Neither can easily find a neutral, competent counterparty.
How we work
Without a legal entity, transparency is the governance. These are published so they can be used against us.
Code, notes and proceedings are published unless a partner requires otherwise.
A charter that is adopted, never granted. There is no centre to ask permission from.
Under ten paid people, permanently. Capacity comes from contributors, not headcount.
A working prototype beats a position paper. Every event produces an artefact.
No commercial affiliation, no exclusive vendors, no paying to be on a stage.
Individuals are named on everything they help produce. Always.
Focus
Four layers. We work across all of them, because the interesting problems sit where they meet — and because a region that only ever works on one layer stays dependent on whoever owns the others.
Chips, GPUs, and what compute actually costs
The layer that quietly decides who gets to build at all. We do not design silicon — that is a fab's job and a billion-dollar one. We work on everything sitting above it: squeezing large models onto small hardware, getting useful inference running on devices that cost less than a laptop, and publishing honest cost-per-token numbers across providers. Malaysia already assembles and tests a meaningful share of the world's chips, which makes this a room we can convene better than most.
Open weights, and the freedom to look inside them
Open weights, open code, open evaluation. Models you can inspect, run yourself and build on without asking anyone's permission — plus the regional and low-resource language work that no large lab is ever going to prioritise, because the commercial case for it does not exist and the cultural one does.
Agents, and systems that act on their own
Where a model stops answering questions and starts doing things. Systems that hold funds, call other systems and act with real consequence, with no human sitting between intention and outcome. This is where the accountability gap lives, and it is where our flagship work sits.
Where the stack touches the ground
Everything below, embodied. Physical autonomy is the point where a mistake stops being a bad output and becomes an actual event in the world — which is also why it draws a crowd and attracts talent. It is led through university partners, because that is where the research capacity already sits.
The arena
The stack is what we work with. The arena is what we point it at — and year one commits to a single one, deliberately, so the work goes deep rather than everywhere.
Open satellite data is free. The work is computer vision and geospatial machine learning that runs on a laptop — which means flood mapping, deforestation detection and air quality modelling genuinely fit a small grant and a ten-week build.
It is also where regional agencies and climate philanthropies already hold budget lines, which matters when the money has to come from somewhere.
A second arena opens only once the first has shipped work. Candidates under consideration: agriculture and food systems, public health, and public infrastructure.
Not in scope: semiconductor design and space hardware. Both are legitimate fields and neither fits a small grant, a ten-week cycle, or a team of eight. Focus is the only real asset a small initiative has.
Interactive · How learning works
Nobody writes the rules into a model. The rules are what is left once the noise has been squeezed out of enough data.
What you are watching is a latent space forming. Every point starts in random position — no order, no meaning, no relationship to its neighbours. Then it settles. Points that belong together end up near each other, and the distances between them start to carry information they did not carry before.
That is the whole trick, and it is why these systems are difficult to hold to account. There is no line of code to point at that says what the model believes. There is only geometry — learned, distributed across billions of parameters, and readable only by running it and watching what comes out. Which is exactly why we fund people to run it and publish what came out.
Programmes
An annual summit, quarterly domain forums, a monthly practitioner workshop series and two to three hackathons a year — plus the quieter work: a member directory, structured introductions between builders, capital, universities and regulators, and closed-door technical briefings where officials can ask engineers direct questions with no procurement relationship in the way.
Design rule: every convening produces an artefact. A published note, a prototype, a dataset, a documented commitment. No event exists purely as an event.
An open call, a reviewed selection, a micro-grant with compute credits and mentor access, eight to twelve weeks of building, and a demo day with investors and university partners in the room.
Small cheques, fast. The point is to get a thing running that did not exist before — not to fund a study about whether it could.
Technical notes from every grant, open training resources, and an annual State of AI Builders survey for the region — the thing that gets AIFG cited rather than merely mentioned.
The obligation: grantees publish code or a full technical write-up. No publication, no further grants. An unproven prototype is just a demo, and demos are cheap.
Open AI access
A small build budget disappears fast against centralised API pricing. AIFG routes member projects to open and independently operated inference and compute providers, so the money goes into the work. Access is free to members with a funded proof of concept.
Routing to open-weight models through an independent provider network. Our default endpoint for agent and language model work where no single operator should sit in the path.
Marketplace compute for fine-tuning and training runs that will not fit a laptop. Priced well below hyperscaler rates, which is the entire point for a ten-thousand-dollar grant.
Model and dataset hosting for published grant outputs. Every AIFG-funded project ships its weights or its write-up here, under the publication obligation.
Subnet access for teams working on distributed training and evaluation. Being assessed for reliability and cost before we route grant work to it.
Burst capacity for hackathon weekends, where a hundred teams need compute for seventy-two hours and then none at all.
A supported toolchain for running open models on member hardware. Zero marginal cost, no data leaving the machine, and the most underrated option for most POCs.
AIFG does not resell access and takes no margin. We hold no exclusive relationship with any provider, and a provider's presence here is not an endorsement of its token, if it has one. Assessment notes for every provider are published, including the ones we decided against.
Community
Conferences, workshops, training and hackathons. Free or near-free to members. Recordings and materials published afterwards.
Materials from past sessions are published in the resource library. Members receive the calendar by email — no other mail, ever.
Stories
Not announcements. What we learned, what broke, what a number actually meant once somebody had to spend it.
A grant that size does not survive a centralised inference bill — it disappears in about nine days of serious experimentation. Here is the arithmetic that pushed us toward open providers, what it cost to switch, and what the first three teams found when they ran the same workload both ways.
Closed-door briefings work because nobody is selling anything. What officials want to know turns out to be far more specific, and far more practical, than the policy papers assume.
Build logOpen Sentinel imagery, four laptops and a weekend. What flood mapping looks like when nobody is waiting on a procurement cycle — including the parts that did not work.
How we workA charter that is adopted rather than granted removes the centre entirely — and with it, the power to shut anybody down. We think that trade is worth making, and here is why.
Year one ambition
Progress against every one of these is published annually, alongside a full account of money received and money spent. If we miss, that is published too.
The network
To start: three named core contributors resident in the city, the charter adopted in public, one open event within ninety days, and a page saying who you are. Nobody grants this. You adopt it and tell us, and we add you to the register.
You decide: your own funding, your own domain priorities, your own programmes, your own governance. AIFG Kuala Lumpur has no authority over AIFG anywhere else, and no city speaks for the others.
The one sanction: the Network Assembly can delist you from the register and withdraw use of the name. It cannot suspend you, because there is no entity that could. A reputational sanction is the honest limit of what a body like this can enforce, and we would rather say so than pretend otherwise.
First city. Programme, grants and secretariat for the opening ninety days.
Contributors being recruited. Targeting phase one.
Contributors being recruited. Targeting phase one.
Contributors being recruited. Targeting phase one.
Early conversations. Contributors welcome to put a hand up.
No permission required. Read the charter, find two others, and tell us you have started.
Roadmap
Phases are sequential and gated. Each one has an exit condition, and we do not move until it is met.
Months 1–3 · Malaysia only
A credible standing start, not scale.
Months 4–12
Exit: three cities running, twelve grants, five hundred members.
Year two
Governance moves from the founders to the network.
Year three onward
The founders should be replaceable by this point.
Get involved
Anyone working in or near the focus domains. Free.
Senior practitioners who own a workstream and carry real authority in it.
Event crew, translation, mentoring, content, local support. No commitment beyond the task.
From a guest lecture to a joint lab, at whichever rung suits you.
Technical access to builders with no procurement relationship attached.
A front-row seat at the frontier, with no commercial conflict and no fund manager.
Support the work
Patrons do not choose grantees, set programme content, or receive commercial rights. We say this in the first meeting rather than the third, because it is precisely what makes the initiative worth funding.
Tier contributions and the full budget breakdown are in the patron prospectus, available on request. Sponsorship of individual events is separate, non-exclusive, capped at three sponsors per event, and never buys a speaking slot.
Questions
No. It is an unincorporated, decentralised initiative. Where a legal counterparty is required to receive funds, that is handled through a fiscal host arrangement with an existing registered organisation. The host holds and administers funds; it has no authority over programme.
No one. The charter is a document anyone may adopt, and the name goes with it. Nobody granted AIFG Kuala Lumpur its standing either — it adopted the charter first, which is the only thing that distinguishes it.
Yes, and you do not need our permission. Find two other core contributors where you are, adopt the charter in public, hold an open event within ninety days, and publish who you are. Tell us and we will add you to the register.
Very little, and we would rather be straight about that. The Assembly can delist a city from the public register and withdraw use of the name, which is a reputational sanction and not a legal one. That is the real price of having no centre, and we think it is worth paying.
No. Grant decisions sit with a rotating panel of core contributors against published criteria, and the outcomes are published. This is structural and it is not negotiable at any contribution level.
No. AIFG does not issue tokens or financial instruments, and no city may do so in the AIFG name. Some of the infrastructure we route grant work to has an associated token; listing a provider is not an endorsement of it, and we take no margin on any of it.
No equity, no cohort model, no commercial interest in what gets built. Grants come with a publication obligation rather than a term sheet. The network is the product.
Both are legitimate fields, and neither fits how we work. Deep tech spans quantum, biotech, semiconductors, materials and energy — a team of eight cannot be credible across that. Space hardware does not fit a ten-thousand-dollar grant or a ten-week cycle. The one exception, Earth observation, sits inside our climate arena already, because it is computer vision on free open data.
Programme is phased against confirmed income, and a scaled-down operating plan is held in reserve. Grants and publishing are protected first; travel and the summit are cut first. What we can afford is published either way.
Start here
Membership is free. Grant calls open in phase one. The charter is available now, and adopting it is the only thing required to start AIFG where you are.