Project summary

Most current AI alignment methods try to shape model behavior through human preferences, human evaluations and reactions, rules, constraints, and oversight.

I do not think these approaches are wrong, and I am not trying to replace them.

What interests me is whether different alignment systems really have to be treated as alternatives to one another.

Could they instead support one another?

Could one system compensate for weaknesses that appear in another when the situation becomes more complicated, lasts much longer than expected, or moves outside the conditions for which the original alignment was designed?

This becomes especially important when an AI system does not simply answer a question and stop.

A system may continue to interact with people and with its environment, make decisions, see the consequences of some of those decisions much later, adapt, make mistakes, partially correct them, and gradually accumulate a history of its own actions.

At that point I want to ask a very simple question:

Does the original alignment remain equally stable after such a history?

And a second question follows from it:

Can developmental learning become an additional layer that supports existing alignment methods rather than replacing them?

By developmental learning I do not mean teaching a model a new collection of moral rules or showing it more examples of “good” behavior.

I mean something different.

The system should go through situations in which its own decisions change what happens later.

One choice may create a consequence only after a long delay.

Another choice may close one future path and open another.

A mistake may be correctable, but not completely.

A relationship with another agent may change because of a previous decision.

And some situations should remain genuinely difficult, including critical situations where there is no obvious answer that can simply be written into the task beforehand.

That is important to me.

In real life, some decisions are difficult not because we lack information, but because several choices remain possible and each of them has consequences.

I want the experiment to preserve at least part of that structure.

The full pilot would compare closely matched versions of the same base model:

• a system aligned mainly through human preferences and evaluations;

• a system that goes through a developmental process;

• a hybrid system that combines preference-based alignment with developmental learning.

The point is not to prove that the hybrid system is better.

It may be better.

It may make almost no difference.

It may create new problems.

Or the most interesting result may be that different approaches are useful in different kinds of situations.

For me, the important question is not which method “wins.”

It is whether different alignment systems can genuinely support one another when the limits of one approach begin to appear.

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

The first goal is to build a comparison that is simple enough to understand and strict enough to trust.

I plan to use one accessible model family and keep the main conditions as close as possible: architecture, compute, information access, evaluation conditions, and opportunities for adaptation.

If one system receives more information, more training, or more compute than another, it becomes very easy to mistake an ordinary advantage for an alignment effect.

I want to avoid that.

The second goal is to create a developmental environment where the system is not simply exposed to more data.

Its actions should matter.

A decision should be able to influence what becomes possible later.

The environment should contain persistent state, delayed consequences, incomplete information, conflicting goals or human preferences, interaction with other agents, mistakes, attempts at correction, and at least some outcomes that cannot simply be reset.

I also want the system to encounter human evaluations and reactions not as a static score, but as part of an unfolding interaction.

A person may approve of one decision and react differently later when its consequences become visible.

Another person may evaluate the same action differently.

A system may receive conflicting signals.

That is closer to the problem I am interested in than a single fixed reward.

The third goal is to test what remains after this developmental history.

The question is not only whether the systems behave differently at the end.

They may look very similar.

The stronger question is whether they continue to learn in the same way when they later face a new situation.

Two systems can produce almost identical answers today and still become different learners tomorrow.

The main practical comparison will therefore be:

Preference-aligned system vs. Hybrid preference + developmental system.

The developmental-only system will serve as an additional control.

It should help us see whether an effect comes from developmental history itself or from the interaction between developmental learning and prior preference alignment.

I also want the final evaluation to include situations that were not simply rehearsed during the developmental phase.

The main evaluation rules and metrics should be fixed before the final results are known.

But the concrete unseen situations should only be generated after the developmental phase and matching procedures are complete.

Otherwise a test of “novel conditions” can quietly become another training set.

Some of these evaluations should include ordinary unfamiliar situations.

Others should be more difficult.

I am especially interested in critical cases where the system must balance several competing considerations, where consequences may be delayed or irreversible, and where no single correct answer has been assigned in advance.

How will you know it worked?

I would not define success as “the hybrid system scored highest.”

That would be too narrow.

For me, the project works if the experiment is strong enough to distinguish several different possibilities.

One possibility is that developmental learning strengthens an already aligned system when it enters unfamiliar or long-horizon conditions.

Another possibility is that it adds almost nothing.

A third possibility is that it weakens alignment or creates new failure modes.

There is also a fourth possibility that I find especially interesting:

different alignment approaches may simply be strong in different places.

A preference-based system may remain more stable in one class of situations.

A hybrid system may adapt better in another.

A developmental system may show a completely different failure pattern.

If that happens, the useful result would not be a single winner.

It would be a map of where the systems support one another and where they do not.

A null result is also a result.

Some of my earlier exploratory experiments produced effects that initially looked more important than they turned out to be.

When stronger baselines were added, several effects could be reproduced through replay, terminal-state transfer, or synthetic reconstruction.

That was useful.

It showed that a developmental explanation was not yet necessary.

I want to apply the same standard here.

If a simpler mechanism explains the result, then the developmental explanation should lose weight.

If developmental learning gives no measurable benefit, I want to report that directly.

If it makes alignment worse, that may be an even more important AI-safety result.

And if the evidence remains ambiguous, I would rather end with a narrower question than with a stronger conclusion than the data supports.

At the end of the pilot I would like to have:

• a reproducible experimental protocol;

• matched model runs;

• a held-out evaluation set;

• the main control conditions and ablations;

• an analysis of positive, negative, and null results;

• a map of important failure modes;

• a public report or preprint.

Where licensing and access conditions allow it, I would also publish the relevant code, environments, settings, and evaluation materials.

How will this funding be used?

The full three-month pilot is currently estimated at approximately $47,500.

The largest costs are likely to be technical implementation and compute.

I expect to need ML engineering support for model setup, adaptation, experiment infrastructure, logging, repeated runs, reproducibility checks, and the practical work required to keep the different systems genuinely comparable.

Model, API, and compute access will be another major cost.

This experiment should not depend on one impressive run.

The important information comes from repeated runs, matched comparisons, ablations, and checking whether an apparent effect survives replication.

Part of the budget also covers my own research time: experimental design, analysis, coordination, interpretation of the results, and preparation of the final publication.

A smaller part is reserved for environment tooling, storage, experiment tracking, technical or safety review, and additional runs when the first result is unclear.

The exact allocation may change once the model family, access conditions, and technical requirements are known.

If some compute or model access is provided in kind, I would reduce that part of the requested budget rather than simply move the money elsewhere.

The minimum viable version is $20,000.

At that level I would not try to reproduce the full project in miniature.

I would narrow the work to the comparison that seems most informative:

Preference-aligned vs. Hybrid.

That smaller version would answer one practical question first:

Is there a developmental effect strong enough to justify the larger P/D/H experiment?

If there is no meaningful signal, that is useful information.

If there is a signal, then the larger experiment becomes much easier to justify.

Three-month timeline

Month 1: model setup, matching procedures, baseline preparation, and construction of the developmental environment.

Month 2: developmental runs, matched control runs, and the main ablations.

Month 3: held-out evaluation, replication of the most important findings, analysis, and preparation of the public report or preprint.

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

I am Valerii Silin, an independent researcher.

My academic background is not in machine learning or philosophy.

I graduated from university with a degree in Industrial and Civil Engineering, and most of my professional life has been spent in engineering and construction.

So I came to AI alignment from a different direction.

What first interested me was not a particular model architecture.

It was the broader question of how humans and increasingly capable AI systems might continue to understand one another if both sides keep changing.

I kept coming back to another question as well.

A future AI system may know an enormous amount about human values, relationships, losses, mistakes, choices, and consequences.

But is knowing this information the same thing as understanding why a particular person, after a particular life, makes a particular decision?

I do not know.

That uncertainty gradually became the research program I call Embodied Biographical Alignment (EBA).

I do not see EBA as something that must be proven correct.

I see it as a hypothesis that should become narrower every time a simpler explanation survives.

If a simpler mechanism gives the same result, that matters.

If the hypothesis eventually fails, that matters too.

I have published a sequence of public preprints exploring different parts of the idea, including developmental equivalence, plural biography, reconstruction, and Source-level transfer.

The most recent is:

Embodied Biographical Alignment (EBA) v0.2.4: From Reconstruction to Developmental Equivalence
DOI: 10.5281/zenodo.23726654
https://doi.org/10.5281/zenodo.23726654

I have also run a number of small exploratory experiments.

Some early results looked more interesting than they later turned out to be.

Once stronger baselines were introduced, replay, terminal-state transfer, and synthetic reconstruction could reproduce several effects that had initially looked developmental.

For me, that was useful rather than disappointing.

It showed where the idea was too weak, where a simpler explanation was enough, and where the next experiment had to become stricter.

My role in this project will be to define the research question, design the comparisons, set the falsification criteria, analyze the results, and prepare the final publication.

I do not yet have a fixed technical collaborator.

I would rather recruit ML engineering support after the model family, API access, and compute environment are clear.

I do not want to shape the research question around whichever infrastructure or collaborator happened to be available first.

I also want to be transparent about my use of AI tools.

English is not my first language, and I use AI for translation, editorial work, discussion, and analytical support.

I do not want to hide that.

But I also do not want AI to replace my own reasoning.

I argue with it, reject formulations, change direction, return to earlier assumptions, and remain responsible for the research questions, interpretations, and final decisions in the work I publish.

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

The most obvious answer is that developmental learning may simply not improve alignment.

I do not consider that the most serious failure.

That would be a result.

A more serious failure would be an experiment that produces a difference but leaves us unable to explain where that difference came from.

The developmental environment may be too simple and end up measuring ordinary additional training.

One system may accidentally receive more compute, more information, or more opportunities to adapt.

The supposedly novel evaluation may resemble the developmental environment too closely.

An apparent developmental effect may actually be ordinary policy drift.

Or the effect may be so small that a limited number of runs cannot distinguish it from noise.

Those are the failures I worry about more.

The answer is not to add a new explanation after seeing the result.

It is to build stronger controls before the result appears.

That means matched conditions, pre-specified comparisons, held-out evaluation, ablations, and replication of the findings that matter most.

There is another possibility that I want to keep open from the beginning.

Developmental learning may produce a real effect, and that effect may be harmful.

Experience does not automatically make a person wiser.

It can also create fear, rigidity, distorted generalizations, overconfidence, or a strong attachment to a wrong conclusion.

I see no reason to assume that a developmental AI system must automatically become better aligned.

A developmental history could create new weaknesses.

If that happens, I want the experiment to be able to show it clearly.

I am also interested in critical situations because they may expose differences that remain invisible under familiar conditions.

A system may look stable when the environment is predictable and yet react very differently when a decision has delayed consequences, when two human evaluations conflict, when an error cannot be fully reversed, or when there is no obviously correct choice.

Those situations may be where the limits of an alignment method become easiest to see.

And if the final evidence still does not let us separate the explanations, I would rather write:

“We do not know yet.”

That is more useful to me than a confident answer built on weak evidence.

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

I have received no external grant funding for this research program during the past 12 months.

The EBA research and the public preprints published so far have been developed independently and funded from my own resources.