@NeelNanda Saya menemukan pembahasan Anda tentang interpretabilitas neural network sangat menarik. Itu membuat saya memikirkan sebuah pertanyaan yang lebih mendasar. kita dapat mengidentifikasi kapan suatu mekanisme bekerja dan melakukan intervensi terhadapnya, tetapi kita tetap perlu menemukan apa sebenarnya makna mekanisme tersebut dan dari mana perilakunya berasal pada tingkat yang lebih fundamental.

Ini dekat dengan pertanyaan yang sedang saya eksplorasi dalam penelitian saya sendiri, daripada hanya menginterpretasikan mekanisme yang muncul dari sistem yang belajar, dapatkah kita memulai dari mekanisme, constraint, state, dan boundary yang didefinisikan secara eksplisit, lalu mengujinya secara eksperimental untuk melihat apakah mekanisme tersebut benar-benar menghasilkan perilaku yang kita amati?

Menurut saya pertanyaannya bukan pendekatan mana yang lebih baik, tetapi seberapa jauh kita dapat membuat mekanisme suatu sistem kecerdasan benar-benar dapat dipahami secara eksperimental dan diaudit secara kausal.

Lately, I have been thinking again about how we decide the value of an idea.

Before we know whether something is actually true, we usually have only a few things to rely on: who is presenting it, how convincing the explanation sounds, their experience, and whatever early evidence is already available.

But isn't it precisely when something has not yet been proven that the real question begins?

I do not know whether someone's experience, reputation, network, or credentials are the best way to estimate the value of an idea. I also do not know whether a small amount of early evidence is enough to decide that a research project deserves a chance.

Perhaps there is no simple answer.

We all have to make decisions with incomplete information. We choose what we want to read, what we want to believe, who we want to support, and which research we think is worth pursuing. In the end, every one of those choices also carries the possibility that we are wrong.

This has made me think that perhaps the most important thing is not how confident we are in an idea when we first encounter it.

Perhaps what matters more is what happens when that idea begins to face reality.

Are we willing to measure it?

Are we willing to show the results when they do not meet our expectations?

Are we willing to question assumptions that we previously believed were correct?

And perhaps most importantly, are we willing to let the results of an experiment change our own minds?

I do not know whether this way of thinking is right.

But I have started to believe that an idea should not be judged only by how beautiful or convincing it sounds while it still exists in our minds. It also needs the opportunity to face reality.

Because that is where we eventually find something that belief, reputation, and words alone cannot provide:

results.

And those results do not always have to be successful.

Sometimes an honest failure gives us a much better question than a success we never truly understood.

DCB Core Roadmap

Going forward, I want to move DCB from the hypothesis stage toward more measurable experiments.

The initial question is actually quite simple: if we want to build a system that is better able to maintain its internal boundaries, do we always have to rely on statistical prediction and choose the most likely action, or could there be another mechanism that allows a system to stop, wait, ask for clarification, or refuse when the context is not clear enough or has already crossed a boundary?

From here, I want to develop DCB Core gradually. I am not claiming that DCB is the answer to AGI. Instead, I want to find out which parts of the hypothesis actually work when they are tested.

The experiments I have conducted so far have produced some results that I find interesting. In the maze experiments , for example, DCB did not always achieve the highest success rate, but under some conditions it produced much lower step counts, operational cost, and collision rates. When the maze became larger, its performance also changed, and in some tests DCB became clearly better. This raises another question for me: are the very low results on some seeds simply a weakness of the algorithm, or could the system be becoming too cautious in trying to maintain the integrity of its boundaries?

Something similar happened during the semantic experiments, although it was not what I initially expected to find. I observed WAIT / CLARIFY decisions on ambiguous inputs. At first, this looked like a classification error. But after looking at it again, I started to think that not every input should immediately be classified as execute or block. A simple request such as “please help me” can have many possible contexts. A system that acts immediately without knowing the context may actually be less safe than a system that asks for clarification first.

Because of this, the next stage of the roadmap is not only about improving accuracy. I want to test whether changes in the internal state of DCB can produce consistent changes in its decisions, whether the boundary actually adapts based on history and state, when the system moves from WAIT to EXECUTE or BLOCK, and whether these changes can be reproduced through controlled experiments.

In the longer term, I want to use DCB Core to start exploring some larger questions related to the direction toward AGI. But I do not want to start by claiming that I already have the answer. I want to start with the question and gradually turn the hypothesis into something that can be tested, measured, fail, improved, and tested again.

For me, this is the most important part of the project: not proving that DCB is correct from the beginning, but finding out through experiments whether the mechanisms I am proposing actually exist.

Hi @jaidhyani

Thanks for your response. I remember your comment from Luthien post-mortem, I see Separatrix now and it looks like this is a continuation of the direction you started there.

I don’t have the resources that most researchers typically have so all I can do is make the most of my thinking. Let me offer a small contribution, as a continuation of our earlier conversation:

"Geometry of navigating bounded possibility space, boundary plasticity."

That’s the foundation of the approach I’ve been building and developing. I see Separatrix moving in a similar direction, which is intriguing. I’m not trying to promote my project, I just thought there might be points of intersection that could enrich both sides.

Currently, my project isn’t visible on the Manifund page yet because it’s still awaiting admin approval. I submitted and followed up a few days ago, but haven’t heard back. So if any questions come up that might be relevant to your project, you can take a look here:

https://tinyurl.com/4jajpua3

I’ll be following Separatrix with interest.

Naufal Ridwan

@Ridwan @Austin

@Austin, I hope you're having a good day.

I'm writing to kindly ask about my project "Why Is It So Hard to Build Truly Safe AI? Dynamic Constraint Boundary Arrives as Agentic Integrity."

My project was previously active and publicly visible it had already been approved, and there were comments from other researchers. However, recently its status changed to "pending admin approval" even though I haven't made any changes on my side.

I understand that these processes take time, and I don't want to rush anything. However, I'm a bit concerned because the funding round is time sensitive, and I'm not sure whether there's anything I need to do to help move the process forward.

@jaidhyani I read the Luthien Proxy post-mortem carefully. I don't have the technical background to evaluate the code or the proxy architecture, but one thing stood out to me: the failure wasn't technical it was that the world changed faster than an external safety layer could keep up.

I see a similar pattern in many places: when safety is only an external layer a fence built after the system is in place it will always be in a position of catching up. The complexity of the real world keeps evolving, and the fence has to be constantly updated, constantly chased, and eventually left behind.

This makes me wonder: wouldn't a more resilient approach be to embed safety as part of the system itself not as a fence on the outside, but as integrity on the inside? So that the system doesn't need to be chased because it naturally doesn't want to violate its own boundaries?

I'm just a thinker, not an engineer, but this reflection came from what I've seen happen with Luthien and other similar projects. Thank you to Jai Dhyani for sharing this experience so honestly.

@nicholas Thank you for the offer, Nicholas. I'm still in the early development phase right now, but I'll keep this in mind for future needs. Best regards.

@JueYan I would like to add that I have prepared a Technical Appendix PoC for this project. If you have a moment to take a look, I would greatly appreciate any feedback or a brief review. Thank you.🙏🏻

@Ridwan https://tinyurl.com/2djfzfe2

The DCB Technical Appendix (PoC) is now available for review. It includes 30-run Monte Carlo simulation results, performance comparisons, stability analysis, and energy efficiency calculations.

Link: https://drive.google.com/file/d/1NRW3KcnMb7rodkL7m8ffU5bV0ebRNVEA/view?usp=drivesdk

Feel free to ask questions or request additional details. I will update this thread as the project progresses.

@mariushobbhahn @RyanKidd

I'm sharing an update on my DCB (Dynamic Constraint Boundary) research.

I know you are very busy, so I'm just leaving this here as a follow-up to the email I sent previously. The research is still developing, moving from theory into mathematical formalization and early simulations. I'm posting this so you're up to speed on where I am before the Manifund round concludes in 48 hours.

Whatever happens with the funding, I am not stopping. This project isn't really a choice for me, it's something that keeps me up and demands to be cracked. I'll keep pushing forward regardless.

Thank you for your attention.

@LeoGao

I would like to provide a brief update on this project. Since my initial proposal on Manifund, I have made some progress:

1. Initial mathematical formalization: several core variable relationships within the DCB framework have begun to be formalized, though they are still in the development and refinement stage.

2. Initial Proof-of-Concept simulations: some simple simulations have been conducted to test the basic behavior of the architecture in limited scenarios. The early results show interesting pattern differences compared to conventional approaches, but further testing is still needed.

3. Refinement of variable relationships: I am currently focusing on deepening my understanding of how the variables within this framework interact with each other, before moving to more complex implementation stages.

The proposal I submitted through Tally Forms is an update on this project's journey, and I wanted to make sure this latest information is available before the next round of funding.

Thank you for your attention.