Thor
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$26,500 left to reach the funding goal.
Max is an AI agent built on a single bet: that general reasoning can be made fully transparent. Instead of opaque neural weights, every capability Max has is an explicit, human-readable rule, so each decision can be traced to a specific principle, and the effect of each principle can be measured by switching it off. Max runs entirely offline on one local machine, with containment and operator goal-approval built in from the start, and its progress is measured on the ARC-AGI-3 benchmark. I’m an independent researcher building it solo.
Frontier AI is powerful but opaque — we can’t read why a model does what it does. Max is the opposite approach: an agent assembled entirely from explicit, general principles. Every choice traces back to a named rule, nothing is hard-coded to a specific puzzle, and each capability has to earn its place as a genuinely general principle rather than a shortcut.
I build it under a strict test-first discipline. Before writing a capability I write tests for it — hundreds of them — and I replay recorded episodes through the agent to check its behaviour. No capability is accepted unless there’s a test that fails without it, which keeps the whole system honest and makes any regression visible.
Safety is structural, not added afterwards. Max runs on one local machine, fully offline, with no network access at all. Containment is physical: the agent is fed information only through physical media by me, the operator, and I approve its goals and sub-goals. By design it cannot choose harmful means, cannot copy itself, cannot alter its own rules or checks, and must stop on command.
The longer-term reason to pursue transparent general reasoning is the problems it could eventually help with — above all scientific and medical discovery. An agent that can learn and reason in genuinely new situations, and whose every step a human can read and check, is the kind of tool that could one day help accelerate research into disease. That is exactly why it must be built safely and transparently from the ground up. This grant funds the concrete first step toward that — not a medical result, but the trustworthy reasoning engine such work would depend on.
For this stage, success means:
Measurably better results on ARC-AGI-3 than today — more tasks and levels solved, reported as clear before/after numbers.
A capability-test suite that keeps growing and stays green, so every new ability is backed by a test that fails without it.
A public write-up of the method, the full set of principles, and the results, detailed enough that others can inspect and critique it.
Demonstrable safety properties: every decision traceable to a rule, and the containment design holding throughout.
If the numbers don’t improve but the method and negative results are clearly documented, that is still a real, shareable outcome.
A local research workstation. The work is currently bottlenecked on hardware: ~$7,200 (50,000 DKK).
A part-time stipend for me for 12 months, so I can work on Max consistently: ~$1,600/month gross (roughly $1,000/month after Danish tax), ~$19,200 total.
Target: ~$26,500. Minimum to keep going: ~$7,200 (the workstation alone).
I’m an independent, self-taught researcher working on this solo. My track record is the artifact itself: a working agent developed under a disciplined test-first regime, a growing suite of capability tests, measured results on ARC-AGI-3, and an explicit safety architecture (physical containment, operator approval, and immutable self-rules). I’m offering the method and the results for open scrutiny.
The most likely failure is that the first-principles approach plateaus before reaching human-level generality on ARC-AGI-3. As a solo project it also carries key-person risk. But even then the work produces value: a documented, reproducible method, honest benchmark numbers, and a concrete demonstration of an interpretable, contained design — results that are useful to others whether or not the capability goal is reached.
None — the project has been entirely self-funded so far.
There are no bids on this project.