Project summary

NadIntellect Lab is currently an independent research project in the field of AI safety.

The main thing we are researching is whether we can create a safe architecture where a very powerful AI can remain very intelligent, but at the same time not have the ability to influence humanity, society, politics, and civilization. We want such a system to be used mainly as a scientific tool.

Our idea is not to make AI less intelligent or just hope that it will follow the rules we give it. We want to test another approach: whether we can build a system of restrictions around a powerful AI so that it cannot get through these systems and get the ability to act independently in the outside world.

We have not yet proven that this approach is safe. We already have a technical hypothesis, a specification, and the system is prepared for the first live experimental test.

At the next stage, we need to test this approach in practice and collect data that will help us understand whether it works or not, where exactly it does not work, and which of our assumptions may be disproved.

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

Our main goal is to test whether we can limit the ability of a powerful AI to influence and interact with the outside world, but without affecting its intellectual abilities.

We have already passed the theoretical stage. We created a technical specification for the first experiment, implemented it, and carried out internal technical checks. But this is not proof that the architecture is safe. It still needs to be tested in practice.

Now we need to move to live testing. The first experiment should test our assumptions about the architecture on a real LLM and show whether we can measure and detect the transmission of information that we do not want to allow through a constrained channel.

If the experiment shows weak points or shows that our assumptions were wrong, we are not going to hide these results. This is exactly what will help us understand what needs to be changed in the architecture or what we need to give up.

If the first stage goes well enough, in the next experiments we will gradually make the conditions more difficult and try to break our own assumptions about the architecture and the system as a whole.

How will this funding be used?

For the next stage of the research, we need funding for live experiments, technical development, and testing of the architecture.

The first part of the funding is needed to conduct the first live experiment. After that, we will analyze the results, do the necessary technical work on the system, and prepare the results so that the experiment can be run again and checked.

Regardless of the result of the first test, we then plan to conduct more difficult experiments, look for weak points, and try to disprove our own assumptions.

We also need funding for independent technical review of the results and for work related to intellectual property protection.

We have already submitted funding applications to two other funders: a BlueDot Rapid Grant for $8,000 and Transformative AI Research Grants for $32,000. Both applications are currently waiting for a decision. If we receive funding from several sources, we will not duplicate the same expenses. The money will be distributed between different stages of the research depending on what has already been funded and what work still needs to be done.

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

Today, NadIntellect Lab is a small independent project. The main research is led by me, Evgeniy Galuschak.

I develop the concept and architecture, formulate hypotheses, think through experiments, and use strong LLMs for cross-checking and criticism. I try to break my own hypotheses and assumptions, remove weak directions, and move the ones that remain toward experimental testing.

I do not have an academic background in AI safety, and I am not trying to create the impression that I do. This project started for me simply as a hobby because I have always been interested in this topic. Over time, this hobby turned into more serious technical work. I moved from theory to building the architecture, then to specifications, and now to experiments.

On the technical side, I work with a programmer, Andriy. He is responsible for technical review, fixing implementation problems, and the engineering side of the experiments.

In our research, we use strong LLMs as tools for search, analysis, criticism, working with technical documentation, and checking hypotheses between different AI models. But we do not consider checks by AI models to be independent external review. That is why we want to move toward independent technical and academic review by human specialists.

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

The most likely reasons for failure are that our architecture may not limit information transmission as reliably as we expect; there may be channels or ways of interaction that we did not see or account for; or some of our assumptions may simply turn out to be wrong.

If the project or an experiment does not give the result we expect, this will not mean that the research was useless. Right now, at the very beginning, one of our main goals is to use experiments to find gaps in the architecture and see what does not work.

The experiments may show that we really cannot limit information transmission through a constrained channel reliably enough. It may also turn out that the model has other ways to transmit information outside the system that we did not see or account for.

If this happens, we will document the result, analyze the problem again, and look for the cause. Maybe we will need to change part of the architecture, or maybe we will need other approaches. This is exactly why we need experiments — to find things that we could not predict in theory.

But if repeated experiments show that the architecture or the system itself is built on the wrong basis, we are ready to abandon this architecture, change it substantially, or close this research direction.

The goal itself is to test what works and what does not. We are starting small and moving from theory to practice.

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

We have not received external funding for NadIntellect Lab yet.

We currently have two submitted applications: a BlueDot Rapid Grant application for $8,000 and a Transformative AI Research Grants application for $32,000. Both applications are waiting for a decision, so we do not count this money as funding already received.

Up to this point, the project has been developed independently and funded by us.