Enhance Malaysia’s research capabilities in the field of AI safety through courses that are easy to understand and learn, workshops on model interpretability, a
The goal of our project is to establish an independent research community in Malaysia dedicated to AI safety and model interpretability, inspired by effective altruism.
Through simple, easy-to-understand technical courses, collaborative research workshops, and open-source experimental projects, we aim to help local students and early-career researchers acquire the necessary skills to make meaningful contributions to AI safety.
The initial 12-week pilot program will combine structured education with hands-on research focused on Transformer representations, causal interventions, and model safety evaluation. Building on my existing open-source research on interpretability, the project will emphasize reproducibility, scientific rigor, and learning from negative outcomes.
Our long-term vision is to build a sustainable local research community and forge connections with the international effective altruism and AI safety ecosystems.
Our primary goal is to address the shortage of locally developed, research-oriented AI safety educational resources in Malaysia and to establish a training pathway that spans from introductory learning to independent technical research.
We are committed to achieving the following three objectives:
1. Accessible AI Security Education
We will organize systematic courses covering topics such as AI alignment, mechanism interpretability, model evaluation, and real-world empirical research. We will adapt existing open educational resources (including materials from BlueDot Impact) rather than building courses from scratch.
2. Hands-On Research on Explainability
We will host hands-on workshops where participants can reproduce published experiments, explore the characteristics of TransferMer models, and conduct controlled activation interventions. My existing open-source research on subspace dynamics and residual flow computation will provide the initial experimental foundation.
Participants will not only learn how to achieve positive results but also how to establish meaningful baselines, identify methodological limitations, and properly address and interpret negative results.
3. Sustainable Community Development
We will establish regular book clubs, collaborative research meetings, and peer-mentoring mechanisms. While operating independently in the absence of formal partnerships, we will seek collaboration with existing Effective Altruism (EA) and AI safety communities in Malaysia.
The pilot project will last 12 weeks and will be conducted primarily online to minimize geographic and economic barriers, with in-person workshops held periodically when conditions permit.
Our ultimate goal is not merely to impart one-time knowledge, but to build a local network centered on AI safety and effective altruism that connects young talents aspiring to conduct research in this field.
We will evaluate this pilot project based on quantifiable metrics for educational, research, and community outcomes.
Education and Participation
Recruit 15–20 local participants online and within the community who are native speakers of Malay, English, or Chinese and have an interest in effective altruism and AI safety.
Ensure that at least 10 participants complete the course.
Assess technical progress through in-class assignments, hands-on experiments, and a final evaluation at the end of the project.
Research Outcomes
Author at least three reports on open-ended or replicable experiments.
Where feasible, publish supporting code, methodologies, and reproducibility guidelines.
Evaluate research quality based on methodological rigor rather than whether experiments yield positive results, with an emphasis on fully and systematically engaging students in hands-on activities rather than merely reading papers.
Community Sustainability
Select and train at least two participants who can assist in organizing future activities.
Establish a regular research discussion group that will continue to operate even after the pilot project funding ends.
From the outset of the project, seek to establish initial connections and foster ongoing collaboration with the broader international AI safety research community.
Transparency and Evaluation
At the end of the 12-week period, we will publish a report documenting participation, completion rates, technical learning outcomes, research findings, expenditures, and approaches that were not successful.
The most important measure of success is whether participants have acquired verifiable research capabilities, rather than merely attending lectures or joining an online community.
We are seeking $8,000 to launch a 12-week initial pilot project. The funds will be used for the implementation of educational courses, the purchase of research computing resources, participant support, and the development of necessary community infrastructure.
Budget:
$2,500 — Project Coordination and Instruction: Course preparation, technical instruction, mentoring, research supervision, and participant support.
$1,600 — Research computing resources: GPU rentals, inference APIs, storage, and experimental infrastructure required for reproducible and interpretable research.
$1,200 — Participant research support: Small stipends or project grants to support participants in conducting independent, real-world, reproducible experiments.
$1,000 — Community events: Workshop venues, local transportation, and logistics for occasional in-person gatherings.
$600 — Educational infrastructure: Website hosting, teaching resources, documentation, and publication of open research materials.
$1,100 — Administration and contingency funds: *Operating expenses, payment processing, and unforeseen costs.
Total requested: $8,000.
If the funds raised fall short of the total amount, we plan to prioritize online instruction and shared computing infrastructure, and reduce the size of the initial participant cohort.
Where feasible, we will utilize existing open-source literature, instructional materials, and research tools, and use AI subscriptions rather than APIs to minimize costs.
All grant expenditures and major deliverables will be documented.
I am an independent AI researcher currently living in Penang, Malaysia. I will lead the project in its early stages, overseeing curriculum development, technical instruction, research guidance, and community organization.
My research focuses on mechanism interpretability, the internal dynamics of Transformers, causal interventions, and empirical AI safety evaluation.
I have conducted independent research across a variety of model architectures and scales, including studies on residual flow dynamics, low-singularity representation subspaces, and the relationship between internal representations and safety-related behavior.
My recent open-source project, [Dark Subspace Dynamics](https://github.com/axobase001/dark-subspace-dynamics), provides reproducible experimental code and evidence for studying how internal perturbations propagate within Transformer models. Related research involves cross-model comparisons, controlled interventions, quantitative evaluations, and detailed documentation of negative results (which I believe is crucial for properly addressing and understanding negative results in safety alignment research).
These experiences provide the technical foundation for teaching participants how to design, reproduce, and evaluate explainability experiments.
It should be acknowledged that I have not previously run a formal AI safety research project of this scale. This is a preliminary pilot aimed at testing whether my research experience can be translated into an effective educational program grounded in effective altruism and focused on AI safety and explainability.
Currently, no other instructors or institutional partners have formally committed to participating in the program. I intend to recruit volunteer coordinators, reach out to existing EA and AI safety communities in Malaysia, and invite external researchers for occasional technical discussions.
This project will launch on a minimal scale to keep delivery manageable while gradually building a broader organizational team.
I see four main risks:
1. Insufficient recruitment or lack of sustained participant engagement.
The technical AI safety community in Malaysia is still in its early stages, and our project may struggle to attract enough qualified participants.
To mitigate this issue, I plan to expand recruitment beyond existing EA networks to include college students, programmers, and independent researchers. The curriculum will include introductory materials and multiple practice-based milestones.
2. Difficulty translating education into meaningful research.
Participants may complete the lectures but fail to develop the ability to conduct independent experiments. Mechanism interpretability and the field of AI safety involve a fairly steep technical learning curve.
We will emphasize hands-on assignments, small-scale reproducible experiments, and methodological feedback, rather than focusing solely on attendance or mere assignment completion.
3. Limitations in personal organizational capacity.
This project currently relies heavily on my personal availability. Without additional organizers, the community may struggle to remain active after the pilot concludes.
I will prioritize identifying potential co-organizers early on and documenting teaching materials and operational processes to eliminate the risk of a single point of failure.
4. Research infrastructure and funding constraints.
GPU costs and experimental complexity may exceed expectations. We will prioritize smaller open-weight models, shared computing resources, and the use of Colab credits—often provided as part of Gemini subscriptions—as alternatives to expensive GPU rentals.
If the project fails, the most likely outcome is that we will have successfully conducted some educational activities but failed to establish a continuously active and tightly-knit research community, nor will we have produced the targeted number of technical outputs.
This would be disappointing, but the pilot is deliberately designed to limit downside risks. We will publish reusable educational materials, completed experiments, and a transparent explanation of which approaches were unsuccessful.
Even if the pilot is unsuccessful, it should still provide useful evidence regarding the feasibility and cost of building technical AI safety research capabilities based on effective altruism in Malaysia.
So far, I haven’t raised any funds through this program, but I plan to spend about $1,000 of my own money to provide Gemini subscriptions for participants.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!