About the project · Frozerti

Goal

Make useful general-purpose local AI more accessible on hardware owned by people and small teams: laptops, home computers and small servers.

We care about the quality of work within a device’s resources. We consider memory, data movement, computation and response time, alongside the needs of software running with the model.

Why local AI

Local operation gives more control over data, less dependence on external services and the ability to work without a constant connection. And when a model can do more within the same resources, a useful assistant becomes available on hardware people already own.

Current work

We study a model where only a small portion of its experts — parts of the model with different abilities — needs to reside in VRAM for the current task. The goal is to reduce GPU memory requirements while preserving answer quality and speed.

The idea is currently being tested on small experimental models. We develop solutions and examine how they learn new tasks and transfer results to unfamiliar examples, separately for single-step and multi-step tasks.

The latest recheck showed an improvement on new tasks. The next goal is multi-step tasks and a stable result across repeated trials.

How we verify results

Compare solutions under comparable conditions.
Test new examples and assess complete answers separately.
Distinguish tool repairs, diagnostic observations and confirmation of a hypothesis.
Publish both positive and negative results, together with the limits of each conclusion.
Check conclusions against the original data.

Research journal

The journal holds the key entries of the current study: what we tested, what we found, what we fixed and what the recheck showed. Early trials are summarised as a short history.

If a published conclusion is refined, a dated addendum appears beside the entry, so the history of checks stays visible.

The project is in research and experimental development.

Research journal