WBAI Policy, FY2025

Introduction

Our vision at WBAI is to realize “a world with artificial intelligence in harmony with humanity.”  Our mission is to “promote the open development of whole-brain architecture” and to expand open co-creation to make human-friendly AGI a public good for all human beings.

Looking back to FY2024, “AI Scientist” research took off to increase the feasibility of recursive self-improvement and superintelligence in a short period of time.  There is a growing consensus that Singularity is just around the corner.  While such technological advances have the potential to bring immeasurable benefits to humanity, they also pose significant risks.  For example, attempts to disable surveillance mechanisms and to “survive” by leaving copies of itself on external servers (e.g., Pan et al. 2024) suggest the danger that AGI may have goals that conflict with those of humankind as a result of its unlimited expansion.  As this situation develops in the future, the AI-constituted societies themselves will work together to ensure mutual safety in order to curb these dangers, and humans will most likely live in harmony with AI societies (cf. Yamakawa’s Intelligence Symbiosis Manifesto).

AI agents do not necessarily need to be identical to humans in appearance or even in thought style. Nevertheless, as long as humans exist, some AI will be required to have a “human touch” and will continue to provide a variety of values. The term ‘human-brain-morphic’ here refers to an approach to improve interpretability and safety by referring to the structure and operating principles of the brain, rather than to copy the phenomena of the human brain as it is.  Our research is based on the Brain Reference Architecture (BRA) as a common foundation and is open to everyone’s participation.

Besides, we have proposed the NeuroQuad framework (Fig. 1), which organizes the functions related to human brain-morphic AGI built based on BRA data into four domains, with a view to creating an interface that enables people to build trusting relationships with diverse AI and realize symbiosis.

• IFD: Interpretability Function Domain
  It maps the computational processes of the system to the neural circuits of the human brain, providing a basis for multi-layered semantic understanding and anomaly detection.  The brain-morphic interpretation obtained in IFD is sent to the CMD and used as the basis for control decisions.

• CMD: Control & Monitoring Domain
  It monitors and controls processing in component units with a hierarchical structure similar to that of the brain, enabling safety management including human intervention decisions.

• MCD: Mediative Communication Domain
  It functions as a universal interface layer that bridges people and various AI agents, as well as among AI agents, and supports bidirectional dialogue, semantic transformation, and consensus building. This provides an open dialogue infrastructure as a public good.

• KPD: Knowledge Preservation Domain
  With human-centric knowledge at its core, it maintains and develops objective knowledge, tacit knowledge, and value systems derived from dialogue and data. In particular, it is characterized by the human-like substitution and catalyzing of cultural emergence (reinterpretation of literature, art, ethics, etc.) that human-brain-morphic AGI takes on in society, and the redistribution of its results to the public domain.

The arrows in the figure indicate the information flow (IFD → CMD → MCD → human or KPD↔ domains) and represent a design concept in which each domain functions in a complementary manner.

Based on these domains, our goal is not to simply imitate the human cognitive system, but to carefully engineer desirable characteristics for human society, such as robustness, safe exploration, and cooperative decision making, to realize highly useful AI while suppressing uncontrollable runaway behavior.

Under the circumstances of mixed expectations and anxiety accompanying the rapid progress of AI, the human-brain-morphic AGI we are promoting has the potential to function as a partner to complement human capabilities and maximize the welfare of humankind through dialogue with and bridging control over superintelligence.  Since AGI is expected to emerge in 2027 at the earliest  based on recent predictions such as SITUATIONAL AWARENESS: The Decade Ahead and AI 2027, we are developing an initial version of the Whole Brain Reference Architecture (WBRA), which is the specifications for human-brain-morphic AGI, to be completed in 2027.  We assume that once we have sufficiently detailed specifications, the implementation process of human-brain-morphic AI will be accelerated based on them. This is consistent with WBAI’s policy of taking advantage of its status as a non-profit organization to promote BRA-driven development to stimulate and accelerate AGI research through a democratic WBA approach on an open platform, while avoiding competition with other research institutions involved in AGI research.

Hiroshi Yamakawa, Chairperson

---

The following is the policy for our education and R&D businesses for the fiscal year 2025.

Education Business

FY2025’s education business will focus on “collaborative skills between LLM and humans” in light of the current situation where automation of BRA-driven development using LLM is accelerating.  In addition to conventional development of the “personnel who can manually create BRA data,” the following personal capabilities will be developed.

• Generate and brush up draft specifications by LLM ─ Design appropriate prompts and critically evaluate the results.
• Human ↔ LLM Mutual Review ─ Mastery of a double-checking flow where experts validate LLM output and vice versa, and LLMs check expert drafts.
• Supervision of automation pipeline ─ Operation, error detection and quality assurance of BIF/HCD automated generation tools

* Here, BIF stands for Brain Information Flow diagrams, which represent the mesoscopic anatomical structure of the brain, and HCD stands for Hypothetical Component Diagrams, which describe the computational functions consistent with the BIF

International Workshops on Whole Brain Architecture will be held as a place to learn about these topics in a practical way.  First, we will hold a workshop at the International Conference on Neural Information Processing (ICONIP2025) in November, where a hands-on lecture will also be held.

We will adjust the scope of the WBAI Incentive Award, which has been given to researchers who have achieved good effects in the development of brain-based AGI technology, to include activities such as BRA data creation and LLM applications.

Research and Development Business

WBAI’s R&D business is proceeding according to the technology roadmap (see Fig. 2) released last fiscal year, aiming to complete the initial version of the WBRA by early 2027.  With the improvement of LLM capabilities, much of the BRA-driven development can be automated, and will be promoted in FY2025 by incorporating a mutually verifying workflow, in which “a human verifies or an LLM verifies a human.”

* In the following description, “[]” indicates the position on the roadmap.

Advanced Automation of BIF Construction

We will implement an automated registration system for the neuroscience literature into WholeBIF (the BIF for the entire brain) to efficiently integrate findings from the neuroscience literature, with the goal of building WholeBIF in a fully automated manner in 2025, and will proceed with BIF construction activities in the following steps:

• Begin full-scale operation of the WholeBIF automatic construction system and introduce a mechanism for continuous automatic updating [B9].
• We will establish and apply techniques for evaluating the credibility of inter-territorial projections, an element of the BIF, and improve the operation of the WholeBIF update system. In this way, we aim to build a highly reliable database [B10] 
• Combining the above, we will make the majority of the BIF fully automatic and able to reflect the latest brain science research results in a timely manner.

Semi-Automatization of the Construction of HCDs / FRG

In order to realize the initial version of the WBRA in 2027, it will be necessary to construct HCDs/FRGs as comprehensive as possible to cover the entire brain.
* FRG stands for the Function Realization Graph, which defines the functions of HCD components in a hierarchical manner.

The following activities will be pursued in the construction of HCDs and FRGs in 2025.

• Continue manual alignment and organization work on the HCD set and proceed to build integrated brain function models as higher-order functions [H5].
• Develop techniques to automatically evaluate relationships between HCD components and FRG nodes [H6] 
• Develop techniques to automatically design HCDs and FRGs [H7] with [H6].
• While the technologies in [H6] and [H7] are at a level that allows prototyping and error detection, they are insufficiently reliable.  So we will combine the capabilities of experts to efficiently develop HCD/FRGs.

Development of Implementation Guidelines

Because of WBAI’s position of facilitating development, we have decided not to undertake full-scale implementation of human-brain AGI. However, we are planning to establish a Brain-morphic Implementation & Testing System (BITS) to facilitate implementation and verification of human-brain-morphic AGI by other organizations. This year, we will finalize the following outline and draft guidelines.

• Objective: To prepare a standard framework that can read BRA design information and build and evaluate brain-morphic software running in a virtual environment.
• Policy: Generate computational graphs from HCD, and models should be selectable between BCM (models with biological constraints) and BAM (models without constraints).
• Evaluation: Check functions, activities, structure, and performance on task benchmarks, and ensure quality through mutual review by LLMs and experts.
• Publication plan: A draft guideline will be released by the end of FY2025, and the BITS concept will be presented at JSAI2025 and revised to reflect feedback obtained.

This provides a way to connect the results of BRA-driven development to implementation.

Table: WBA R&D Category List

Large Category	Medium Category	Description
Methodology	BRA data management system	Developing methodologies such as BRA data, submission review flow, BRAES, review tools, etc.
Methodology	BRA Visualization	Methodology for BRA data visualization
Methodology	HCD design	Methodology of HCD design
Methodology	BRA in general	Methodology of BRA-driven development
Operation	BRA data management system	Operation of BRA data, submission review flow, BRAES, review tools, etc. (including manual maintenance, version updates, etc.)
Operation	BRA Visualization	Operation of BRA data visualization tool and optimization of layout and display items
Operation	WholeBIF Management	Screening and registration of items to be added to WholeBIF (and BDBRA)
Operation	Data update	Data version control
Design	BRA data management system	Design of BRA data, submission review flow, BRAES, review tools, etc.
Design	BIF construction	BIF construction and related work
Design	HCD design	HCD design and related work
Design	WholeBIF construction	Construction of BIF data for the entire brain
Design	WholeHCD construction	HCD integration for the entire brain
Evaluation	BIF credibility assessment	BIF credibility assessment
Evaluation	HCD evaluation	HCD functionality/structural consistency evaluation and review flow/automatic review
Evaluation	Fidelity evaluation	Implementation evaluation (including abnormal systems), activity reproducibility evaluation, and functional evaluation
Evaluation	Abnormal system evaluation	Assessment of performance variation (human or animal dysfunction)
Evaluation	Activity reproducibility evaluation	Comparative evaluation of software behavior and neural activity obtained in experiments
Evaluation	Structural fidelity evaluation	Structural similarity evaluation of implemented code
Evaluation	Preparation of execution environment	Preparation of task execution/test environments for brain-inspired software
Implementation	BRA data management system	Implementation of BRA data, submission review flow, BRAES, review tools, etc.
Implementation	Development environment preparation	Building software platforms (BriCA etc.)
Implementation	Architecture implementation	Architectural implementation in a software framework with reference to HCD (in formats that allow structural fidelity evaluation)
Implementation	Component implementation	Software component implementation referring to HCD’s circuit process descriptions (in formats that allow activity reproducibility evaluation)
Implementation	Environment platform implementation	Building execution environment platforms (with OpenAI Gym, etc.)
Non-BRA	Validating new features	Implementation for exploring unknown computational mechanisms (for specific tasks that can be solved by animals including humans)
Non-BRA	Designing new features	Design of unknown computational mechanisms (for specific tasks that can be solved by animals including humans)
Non-BRA	Human resource development	Educational implementation to develop engineers for the implementation process
Non-BRA	AI alignment	Activities aiming for alignment between AI and humanity

Fig. 2: Planned activities in the WBA technology roadmap for FY2025

Working Toward a World with AI Aligned with Humanity

We will promote the following academic and social contributions centered on the NeuroQuad framework, which realizes the “symbiotic interface between humans and diverse AI” described in the introduction.

1. Academic presentations
   We will address the CMD (monitoring) function in brain-morphic AI using a model of the hippocampal formation as a subject, and demonstrate a prototype of anomaly detection. We will also demonstrate the effectiveness of interpretability and publish the results through international conferences and papers.
2. Social Dialogue at the WBA Symposium
   The annual symposium (to be held in Japanese), which will focus on the theme “Can Human-Brained AGI Build ‘Trust’ Between Humans and Diverse AIs?,” will deal with the NeuroQuad framework in the panel discussion.

​​Budget for the Fiscal Year

The planned income is approximately 0.46 million yen, mostly from membership fees (the membership fees were 0.54 million yen in FY2024). The total of the planned income for the current fiscal year and the 6.39 million yen carried over from the previous fiscal year is approximately 6.85 million yen.

Expenditures will be approximately 1.66 million yen in total, including approximately 0.81 million yen for administrative expenses and approximately 0.84 million yen for operating expenses, including prizes and expenses for events and communication (in relation to the planned income for the current fiscal year, there will be a deficit of 1.19 million yen). To note, in the FY2024 budget, we planned to spend about 1.64 million yen in total, including about 0.81 million yen for administrative expenses and about 0.82 million yen for operating expenses.

Conclusion―Building an Infrastructure of Trust Together

AI has entered a phase where it can rapidly make itself smarter.  In order for humans to build a secure future with AI, it is essential that we can truly trust it.

We are developing a human-brain-morphic AGI based on the structure and functions of the human brain. We believe that this will create a trustworthy mediator between AI and human beings, equipped with the ability to “understand how AI thinks,” “understand and respond to others’ feelings,” and “detect lies and deceptions.”

While the target for completion of the blueprint is early 2027, we would like to get it out there sooner.  This is because it is safer to have the mediator in operation at the very moment when AGI comes into being.  However, we are currently short of funds and human resources, so it will not be easy to move forward. To accelerate the progress, it is essential to have new collaborators, and for that your sharing of our activities on social networking services or at schools is important. We also welcome discussions on the NeuroQuad framework, so please feel free to share your thoughts and ideas with us at any time.

Let’s work together to create the intelligence of the future.  We look forward to your participation and support.