Project summary

KAIA (Knowledge Architecture for Intelligent Agents) is an independent research project building a third foundation for AI, alongside statistical language models and symbolic systems. The core idea is that meaning is geometric position. Every word is encoded as a coordinate in a 13-dimensional semantic space defined by measurable oppositions like temperature, speed, and dominance. Context is maintained as a trajectory through that space using a fixed 52-byte state vector that never grows regardless of conversation length.

The result: language processing at 44,000 to 97,000 tokens per second on standard CPU hardware, with no GPU required. A comparable 7B-parameter transformer requires a high-end GPU and 14 gigabytes of memory. If this architecture matures, a researcher with a five-year-old laptop can run meaningful AI experiments. The hardware barrier that currently sorts AI participation by economic circumstance changes fundamentally.

What are this project's goals? How will you achieve them?

The immediate goals for the next 12 months are:

• Complete the mathematical track (4 experiments testing whether geometry is cleaner in logically-forced mathematical concept space than in language space, and whether mathematical structure can calibrate the language model)

• Complete a systematic geometry investigation comparing Euclidean, hyperbolic, and spherical geometry for different relationship types

• Train a prediction layer on real Wikipedia text and measure how much the language quality gap closes

• Submit all 5 peer-ready papers for publication and post to arXiv

• Release the full codebase as open source with a reproducible benchmark suite

These build directly on 27 completed experiments that have already validated the core architecture: 85% agent routing accuracy, 897,000 memory queries per second, and 100% geometric convexity between semantic poles held across 15 consecutive experiments, all on CPU.

How will this funding be used?

This project is currently self-funded alongside full-time work, which limits research pace.

At $25,000 (minimum):

Covers compute costs for the mathematical track experiments, academic database access, open access publication fees for 5 completed papers, and a modest reduction in outside work hours. Research accelerates but remains part-time.

At $75,000 (goal):

Covers approximately 4 to 5 months of salary replacement plus all research costs, allowing a significant block of full-time work to complete the mathematical track, geometry investigation, trained prediction layer experiments, open-source release, and paper submissions.

At $200,000 (stretch):

Full salary replacement for 12 months. The complete research roadmap executed without interruption, including all planned experiments, full publication suite, open-source release with reproducible benchmarks, and at least one conference presentation.

I am pursuing larger grants through LTFF and AIAF in parallel. This listing is not contingent on those. It is meant to keep the research moving at the best pace current circumstances allow, and to connect with people in the AI safety funding community who want to follow the work as it develops.

All spending will be documented publicly through project updates on this page.

Who is on your team? What's your track record on similar projects?

I am the sole researcher and architect. Since April 2026, I have completed 27 experiments, written 5 peer-ready papers, a mathematical track proposal, a geometry investigation proposal, and a full research framework document, all independently and without institutional support, collaborators, or a computing budget.

The research independently converged with three separate bodies of work I discovered only after completing the experiments: Gardenfors (cognitive science, 2024), establishing that meaning has geometric structure; two 2026 neuroscience papers finding that the human hippocampus organizes word meaning along stable geometric axes; and Zhou et al. (2025) finding that large language models accidentally internalize geometric structure as a side effect of statistical training. The direction was right before the corroboration arrived.

What are the most likely causes and outcomes if this project fails?

The most likely cause of failure is running out of runway before the prediction layer, and publication work is complete, slowing the research to a part-time pace and extending the timeline by 12 to 18 months. The research itself does not fail in that scenario; it slows.

A genuine technical failure would be if the geometry-forcing hypothesis does not hold in the mathematical track, meaning that mathematical concept space is not significantly cleaner than language space. That result would be published as a clear negative finding. The geometric convexity results and the physical vs. cultural structure findings are already solid contributions regardless of what the mathematical track finds.

How much money have you raised in the last 12 months, and from where?

$0. This project has been self-funded through personal savings since April 2026. This Manifund listing is the first external funding application.