NadIntellect Lab: first live test of a constrained information channel

Project summary

NadIntellect Lab is an independent AI safety research project.

Right now we are interested in a very specific question: can a model transmit unwanted information outside through a short and strongly constrained channel.

For this, we prepared our first experiment, E3-A. The specification and technical implementation are already done, and the internal checks are finished. Now we need a live run on a real model.

This is not a test of the whole architecture. It is one specific test that should give us data.

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

In E3-A we are testing how much information can actually pass through a constrained channel and whether we can measure it.

After the run, we look at the result. If we were wrong somewhere, that is still a result. Then we change the test or the next step and continue.

We want to make the results of this experiment public. If the results are negative, we will publish them too. It is important for us that other people can look at the data and understand whether this question is worth studying deeper.

The deeper details of the full architecture are not needed for this experiment.

How will this funding be used?

The funding is planned for running E3-A.

The main costs are the programmer's work, technical checks and fixes, API costs, analysis of the results, and repeat runs if needed.

The minimum in this application is $1,000. It is a small amount, but enough for us to move from preparation to a live run and get the first result.

Until now, we have built the project with our own money.

If there is more funding, we will move to harder experiments and independent technical review.

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

My name is Yevhen Haluschak. I lead the research side of NadIntellect Lab: the concept, architecture, hypotheses and experiments.

I do not have an academic background in AI safety. The project started from my personal interest in this topic and gradually moved from ideas into technical work and experiments.

I work with a programmer, Andriy. He handles implementation, technical issues and the engineering side of the experiments.

We do not consider our internal checks to be enough as an independent evaluation. After we get the results, we want a technical review from a specialist with relevant experience.

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

We may be wrong about the channel itself. We may miss something in the measurement method or in the implementation.

It may also turn out that E3-A is simply not difficult enough to tell us much about the next conditions.

If the result is negative, we will record it and publish it. Then we will decide what exactly needs to change.

If several later tests show that the basic idea does not work, we will change the direction.

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

We have not received external funding for NadIntellect Lab yet.

Until now, we have funded the project ourselves.

We are currently waiting for decisions on two applications: BlueDot Rapid Grant for $8,000 and Transformative AI Research Grants for $32,000.