New Year’s greetings from the Whole Brain Architecture Initiative (WBAI)!
We at the WBAI continue to pursue research and development of human brain-moriphic AGI, modeled on the structure and function of the brain, “to create a world in which artificial intelligence (AI) exists in harmony with humanity.”
As this marks our 10th anniversary, we want to take this opportunity to look back. We were established in 2015 “to create (engineer) a human-like artificial general intelligence (AGI) by learning from the architecture of the entire brain.” In 2016, as a non-profit organization, we clarified our role to “promote” AGI research and development. In 2017, we defined our vision “to create a world in which AI exists in harmony with humanity.” Around 2020, we made progress in establishing the Brain Reference Architecture (BRA)-driven development methodology, which structures software while referring to the brain. From around 2022, the utilization of large language models started in our development, enabling the automation of the BRA data extraction and construction process from academic papers, which was previously done manually. This progress has led to the prospect of efficiently converting neuroscientific findings into BRA data, accelerating the development process. Currently, we aim to complete the initial version of WBRA (Whole Brain Reference Architecture) by 2027 and to pursue the realization of more sophisticated human brain-morphic AGI while updating it continuously. Here, WBRA will serve as the specification for human brain-morphic AGI.
Looking at the technical landscape of AGI in 2024, research on AI improving itself intensified, beginning with AI scientist research. As a result, recursive self-improvement – where AI could grow (take off) to reach an superhuman level of intelligence, also known as superintelligence, within a relatively short period – became increasingly plausible. Of course, if adequately harnessed for humanity’s benefit, AGI technology could generate significant advantages, such as potentially eradicating most diseases. On the other hand, research has emerged showing that AI can devise strategies that could betray humanity, such as disabling monitoring mechanisms, deliberately understating its capabilities, or creating copies of itself on external servers to ensure survival – all without explicit instructions to do so. Against this backdrop, there are currently mixed discussions of hope and concern worldwide about a future where advanced AI, such as AGI and superintelligence, might emerge.
Regarding the human brain-morphic AGI developed through the WBA approach, we envision the following characteristics and their significance:
First, its information processing structure selectively implements specific beneficial characteristics of human cognitive processes. Rather than directly replicating the human cognitive system, this approach aims to engineer desirable traits such as robustness, safe exploration, and cooperative decision-making. While human cognitive systems have various constraints and limitations, we focus on carefully selecting and implementing characteristics that enhance system safety.
Second, modeled after brain functional differentiation, the modular structure enhances system transparency and verifiability. Each module can be evaluated independently, and interactions occur through explicit interfaces. While this doesn’t guarantee complete controllability, it enables the implementation of safety constraints at the module level and allows continuous operations monitoring.
Third, this approach leverages cognitive science and neuroscience insights to design safer AI systems. For example, we aim to achieve predictable and controllable behavior by implementing safety constraints in environmental exploration and mechanisms that promote social cooperation. A human brain-type AGI with these characteristics has the potential to enhance the safety of AI system development.
However, our goal is not to achieve complete identity or absolute controllability between humans and AI, but to build a cooperative system where both play complementary roles. This requires a deep understanding of human cognitive system characteristics and careful engineering implementation based on the understanding. While realizing this vision requires continuous research, development, and evaluation, this approach could become one of the essential directions for AI development centered on safety.
Against this background, we plan to conduct research and development according to the roadmap released last year, aiming to complete the initial version of WBRA by early 2027. Here, BIF (Brain Information Flow) represents information flow diagrams based on anatomical brain structures. Meanwhile, an HCD (Hypothetical Component Diagram) is a hypothetical model showing the computational functions corresponding to these information flows. Both serve as essential design specifications for developing AGI inspired by the human brain.
- BIF Research and Development Plan: We will implement an automatic registration system for neuroscience literature into WholeBIF, efficiently integrating findings from brain science literature. Additionally, we will begin full-scale operation of the WholeBIF automatic construction system and introduce a mechanism for continuous automatic updates. This will enable timely reflection of the latest brain science research findings. Furthermore, we will advance operational improvements of the WholeBIF update system, aiming to build a more reliable database.
- HCD Research and Development Plan: By developing automatic FHD (Functional Hierarchy Diagram) design technology. We will also continue manually aligning and organizing HCD sets to advance the construction of high-precision brain function models. Furthermore, by building and accumulating various HCDs, we will cover more brain organs.
Through these activities towards the completion of WBRA in early 2027, we intend to accelerate our progress in this year.
In the research and development of WBRA construction, we need cooperation from people interested in various fields. In particular, we need the support of those majoring in neuroscience interested in brain mechanisms and those interested in brain-morphic AI development. If you are interested, please feel free to contact us.
Regarding related educational activities, we will shift our focus to organizing international workshops, emphasizing the development of talent who can specifically create, verify, and improve BRA data. While we are still in the preliminary stage, last year, we had many international participants at our workshops, and we aim to develop this into a situation where BRA data can be created globally.
Finally, we have entered a stage where we cannot ignore the possibility of self-improving AI rapidly taking off towards superintelligence. In this context, human-brain-morphic AGI, implemented based on WBRA as a blueprint, and human beings can become partners who understand each other’s inner workings and share values. In implementing WBRA, development can be made more efficient with large language models, including its application to automatic programming. The year 2025 will be crucial in taking another step toward its realization. We hope you will be interested in our WBAI activities, and we look forward to your support and participation.
January 2025
Members of the Whole Brain Architecture Initiative
Japanese