An Open Letter to the Board of Directors at OpenAI
On November 1st, 2023, at the Hawking Fellowship Award event at Cambridge University, in response to a question regarding whether Large Language Models (LLM) provide the right approach to achieve Artificial General Intelligence (“AGI”) or whether “another breakthrough is needed,” OpenAI CEO Sam Altman replied:
“We need another breakthrough.” Further clarifying, “teaching it (a LLM based AI system, such as ChatGPT) to clone the behavior of humans and human text - I don't think that's going to get us there.”
The OpenAI Charter states:
“…if a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project.”
We believe VERSES qualifies for your assistance. In return, we offer our assistance to help ensure that general and superintelligence are developed and deployed in a safe and beneficial manner for all of humanity.
While the dominant deep learning methods that underpin today’s foundational models, such as GPT-4, LLaMa, BARD and Claude, among others, have demonstrated truly amazing progress in the field, their lack of generalizability, explainability and governability suggest that they will not lead to AGI that is adaptable, safe or sustainable. The fading confidence in their viability as the foundation for AGI, as expressed by Mr. Altman above, is further supported by recent research highlighting the limitations of LLMs — and by a growing consensus in the AI Industry.
Deep learning is not enough.
While some promising alternatives are being researched, we believe there is one particular approach that offers a smarter, safer and more sustainable foundation and, thus, a better path to general intelligence — based on our best scientific understanding of intelligence as it occurs in nature.
Over the past several years, our team of computer scientists, neuroscientists and engineers at VERSES AI Research Lab, led by our Chief Science Officer, Dr. Karl Friston, has developed this alternative approach, known as Active Inference, founded on first principles and guided by nature's blueprint for intelligence in biological systems. Active Inference codifies mathematical principles that explicitly link intelligence, cognition and rational behavior to physical processes at all scales: from the macroscopic scale of humans to the microscopic scale of cells and even down to the scale of quantum particles. Established by the Free Energy Principle, this link between physics, cognition and behavior provides an explicit recipe for modeling any physical system as an Active Inference Agent that can reason, learn, plan, predict and act in real-time.
In contrast to the inflexible and monolithic nature of today’s “black box” AI models, these agents are flexible, specialized and composable. They can interoperate, share and learn together, enabling the kinds of collective intelligence we find in nature. With Active Inference, agents automatically seek out and learn from the most relevant data and then form beliefs in the most computationally efficient way. Critically, this approach results in a transparent and human-interpretable decision-making process. And perhaps most importantly, these agents have a natural drive toward homeostasis, seeking a dynamic equilibrium to harmonize with others and their environment to form healthy, stable ecosystems.
VERSES recently achieved a significant internal breakthrough in Active Inference that we believe addresses the tractability problem of probabilistic AI. This advancement enables the design and deployment of adaptive, real-time Active Inference agents at scale, matching and often surpassing the performance of state-of-the-art deep learning. These agents achieve superior performance using orders of magnitude less input data and are optimized for energy efficiency, specifically designed for intelligent computing on the edge, not just in the cloud. Building on this breakthrough, we developed a novel framework to facilitate the scalable generation of agents with radically improved generalization, adaptability and computational efficiency. This framework also features superior alignability, interoperability and governability in accordance with and complemented by the P2874 Spatial Web standards being developed by the Institute of Electrical and Electronics Engineers (IEEE).
Although we are still in the early stages of our development, we believe that recent successes indicate a promising alternative path forward that, per your charter, we wanted to share with you and — given its potential importance — we wanted to share with the world.
At VERSES, we believe that following nature's blueprint for intelligence offers us the smartest, safest and most sustainable foundation to scale systems to general and superintelligence. Importantly, we believe that our approach is complementary to LLMs and other AI models, both filling in the gaps in their abilities and enabling their interoperability and composability, thus offering a more scientifically grounded foundation for the kind of dynamic, real-world cognitive, emotional, social and embodied intelligence associated with AGI. We are building on this foundation and want to support and encourage others to build on it as well.
In the spirit of cooperation and in accordance with your charter, we invite you to connect and collaborate with us and evaluate our work, near-term goals and timelines with the hope that, upon validation, you will find our approach to building safe, intelligent systems to be the most promising path to benefiting humanity. We request your assistance in supporting our research and encourage you and the industry to consider evaluating and incorporating it into your own research agenda and joining us in committing to a human-centered, standards-based approach that seeks to guide us along the natural path to developing general intelligence and beyond.
Let’s build not only smarter technology, let’s build a smarter world.