Toward Next-Generation AI

1. Introduction

Modern artificial intelligence, driven largely by deep neural networks and large language models, has achieved great success in areas like language understanding, image generation, and code generation. Yet these systems cannot learn continuously without forgetting, they lack genuine world modeling, and they mostly have every modality handled by language backbone. This document outlines a set of proposals for moving AI toward systems that are more brain-inspired, modular, with pre-encoded knowledge and capable of recursive, continuous thought processes, moving most of acquired knowledge out from parameters of the model into tightly integrated memory modules (increasing generalization and computational capability of models) and continual learning. These ideas draw from neuroscience, cognitive science, and emerging research trends to sketch a roadmap for more powerful agentic AI solutions.

A particularly important application of these ideas is the development of AI Co-Scientist systems: collaborative agentic AI platforms where AI agents work alongside human scientists to conduct autonomous research, generate novel hypotheses, design and execute experiments. Current AI research assistants operate in a reactive mode, responding to specific instructions rather than autonomously formulating novel research directions. They lack mechanisms for systematically learning from existing literature, identifying research gaps, and proposing genuinely novel solutions. They cannot accumulate knowledge across research projects without catastrophic forgetting, and they have no principled way to measure the novelty of their proposed solutions.

The vision outlined in this document directly addresses these limitations. The principles of domain knowledge injection, world models, modular architectures, continual learning, memory consolidation, creativity, and multi-agent collaboration together form the foundational building blocks for next-generation AI Co-Scientist systems that can function as genuine scientific collaborators.

2. Encoding Domain Knowledge into Neural Networks

Every living brain begins life with pre-encoded knowledge. Insects fly almost immediately after birth. A lot of mammals walk within minutes. Human infants are born with innate reflexes and more. All of this knowledge is encoded in DNA and the biological processes before birth, the product of billions of years of evolutionary optimization. By contrast, today's neural networks are initialized with random weights, discarding domain knowledge and forcing the network to learn everything from scratch. We need systematic methods to inject domain knowledge into neural networks, either before or during training.

The concept of scaffolding—using domain knowledge to pre-structure the network's topology or initial weights, much like DNA scaffolds the brain—represents a promising but underexplored direction. Imagine initializing a physics simulation network with known conservation laws encoded in its architecture, or bootstrapping a chemistry model with molecular symmetry constraints built into its weight structure rather than hoping the model discovers them from data alone. The reliance on random initialization is a bottleneck for Agentic AI, particularly in scientific domains where data is sparse and the cost of "learning from scratch" is prohibitive. Random initialization wastes significant training time and fails to leverage the accumulated domain knowledge. To create agents capable of scientific discovery, we must transition to systems that begin their existence with a functional and clear understanding of their domain.

In the context of AI for Science, domain knowledge can be represented by Knowledge Graphs (KGs) which will act like scaffolding for the model. Unlike general-purpose LLMs that can hallucinate physical laws, a model initialized with neuro-symbolic scaffolding using KGs begins with a latent space aligned to known scientific truths. By embedding domain ontologies (e.g., the periodic table, metabolic pathways, or causal physical laws) directly into the network's initial weights, we create a "prior" that constrains the search space. This prevents the model from exploring physically impossible solutions, such as negative density or violations of conservation laws.

3. World Models as the Foundation for Learning

Children spend roughly five to six years building world models before they learn to read and write. They observe their environment, imitate adults, play games, role-play scenarios, and dream about the future. Play is not idle time—it is structured exploration through which children learn physics, social dynamics, causality, and planning. A critical limitation of current LLMs is their reliance on autoregressive next-token prediction, which models the statistical correlations of language rather than the causal dynamics of reality ("World Models"). Text is a low-bandwidth, highly compressed abstraction of the world. As Yann LeCun argues, to achieve human-level intelligence, AI must learn from high-bandwidth sensory data (vision, audio) to understand the physical constraints of the environment.

Neural network training should begin with rich world models built from multimodal observation—video, audio, and interaction with the environment—before specializing in language and symbolic reasoning. The brain constructs its world model through constant sensory integration, even during sleep and idle moments.

The brain can also generate predictions and simulations without any external input: imagining scenarios, rehearsing plans, consolidating memories. World Models serve as the agent's internal simulator. They allow the AI to "imagine" the consequences of a sequence of actions without physically executing them—a capability essential for planning, safety, and scientific hypothesis generation.

World models are perhaps the most directly relevant concept for AI Co-Scientist systems. Scientific research (especially in physical sciences) is fundamentally about building, testing, and refining models of reality. An AI Co-Scientist with genuine world modeling capabilities could simulate the likely outcomes of proposed experiments before committing resources to executing them—a form of "imagination" that dramatically accelerates the research cycle.

4. Modular, Brain-Inspired Architectures

Most of the current multimodal AI systems typically funnel all inputs through a single large language model. In the human brain, distinct cortical regions handle different modalities: the visual cortex processes sight, the auditory cortex processes sound, Broca's and Wernicke's areas handle language production and comprehension. Importantly, the vision system is the largest of these (an estimated 30% to 50% of the cerebral cortex is dedicated to visual processing), reflecting the fact that most information we process is visual. The language system, while critical, is comparatively compact. "Most of the information we get goes through our vision." Therefore, Vision must be the anchor of the World Model, not Language.

We need modular neural network architectures in which specialized modules handle vision, audio, language, motor control, memory, and other modalities independently, and communicate through efficient inter-module channels. Cross-attention, the current dominant mechanism for module interaction, may not be the most effective and efficient approach. A more promising direction involves direct exchange of learned embeddings between modules, analogous to how dense neuronal fiber tracts connect cortical regions in the brain. Each module should be smaller than today's LLMs, but collectively they should be more capable because they can develop deep specialization.

Furthermore, neural networks should adopt localized connectivity rather than full connectivity. The biological brain is not fully connected neurons. Neurons are connected primarily to their neighbors, with long-range connections reserved for inter-region communication. Fully connected layers are computationally expensive and biologically implausible. Sparse, locally connected architectures would reduce training and inference costs while potentially improving generalization.

5. Continual Learning and Alternatives to Backpropagation

Continual learning—the ability to learn new tasks without forgetting old ones—is arguably the single biggest blocker for truly powerful agentic AI systems and robotics. The root cause is backpropagation itself: global gradient-based updates are inherently destabilizing to previously learned weights, causing catastrophic forgetting. The biological brain does not perform backpropagation. It uses local learning rules, neuromodulatory signals, and synaptic consolidation processes that allow ongoing learning throughout a lifetime.

Neuromodulatory signals use environmental cues such as rewards, risks, novelty, and effort. They alter brain states, enabling flexible behavior and learning by modulating the sensitivity of large, interconnected brain regions. Synaptic consolidation is the rapid, initial, and local stabilization of neural connections (synapses) within minutes to hours of learning, transforming fragile, short-term memories into stable, long-term forms. Emerging research is making alternatives viable. These approaches open the door to AI systems that can learn on the job, personalize to individual users, and adapt to new domains without retraining from scratch.

Continual learning is also essential for personalization. An AI assistant—a digital butler, for example—should learn everything about the person it serves: preferences, routines, communication style, and context. This requires both model parameter updates (the agent gets better at its tasks) and memory updates (the agent accumulates and organizes facts). Repetition should strengthen knowledge in both systems, mirroring how the brain consolidates through rehearsal.

Continual learning is perhaps the most critical unsolved challenge for AI Co-Scientist systems. A truly effective scientific collaborator must accumulate knowledge and methodological expertise across multiple research projects and from new literature and research without catastrophic forgetting. For the AI Co-Scientist, the ideal approach should combine strategies for parametric continual learning (preserve learned reasoning capabilities) and externalized knowledge accumulation (using dynamic knowledge graphs to maintain structured scientific facts). A personalized AI Co-Scientist should develop an increasingly sophisticated understanding of its domain over time through collaboration with a human scientist and through learning from new literature and research, building both methodological expertise and domain knowledge that makes it an increasingly valuable research partner over time.

6. Memory Architecture and Consolidation

Current LLMs store far too much information in their parameters. They blend knowledge storage with reasoning capability, resulting in very large models that are expensive to train and prone to hallucination. A better division of labor would position models as generalization and reasoning engines while offloading factual information to tightly integrated memory systems. This mirrors the brain's separation between the hippocampus (episodic memory, facts) and the neocortex (generalized skills and patterns).

Memory should include both short-term (working memory for current tasks) and long-term components and should be trained jointly with the model rather than bolted on as a retrieval-augmented afterthought. Critically, we also need a process for memory consolidation: an offline phase in which the system reviews, reorganizes, strengthens important memories, weakens irrelevant ones, and cleans out stale information. Memory consolidation is one of the brain's most powerful mechanisms, and its absence in current AI is a significant gap. Memory consolidation should also mimic Hippocampal-Cortical Consolidation, the biological process where the hippocampus (short-term memory) "teaches" the cortex (long-term memory) during REM sleep. In AI, this can be implemented as a fast-learning "episodic" module (hippocampus) training a slow-learning "semantic" module (cortex) during offline periods.

Agents should also maintain experiential memory: when faced with a new task, they should be able to retrieve and draw on similar past experiences, much like a senior scientist finds solutions to problems more quickly using their past experiences.

7. Intuition, Imagination and Creativity

Intuition, imagination and creativity are interconnected cognitive and emotional processes that drive innovation. They represent a shift away from pure, logical, linear thinking toward a more holistic, "right-brain" approach that allows for new ideas to emerge. They are a very important part of the scientific process for creating truly novel hypotheses and solutions for problems.

Intuition allows an expert to rapidly narrow down a large solution space based on experience and pattern recognition—selecting which approaches are worth trying and which are dead ends without exhaustive search. It is a crucial component of the creative process, guiding and accelerating decisions when analytical information is insufficient.

Imagination is the ability to form mental images or scenarios that are not currently present. It is the foundation of creativity, allowing individuals to explore alternative realities, simulate future possibilities, and break away from logical constraints. Current AI models are not capable of imagination. Hallucinations can be in some way turned into imagination but the best way to have AI models imagine is to include some kind of recursive process and evaluating self thought for creating future plans and scenarios.

Creativity involves generating genuinely novel solutions—not novel relative to the model's experience, but novel in an absolute sense. Measuring novelty itself is a deep problem. Measuring novelty involves quantifying how new, unexpected, or different an idea, product, or research finding is compared to existing knowledge. Novel does not just mean new and untried. Novel ideas should bring solutions to unsolved problems or at least should be promising solutions even if they are not proven yet.

8. Swarm Intelligence and Multi-Agent Collaboration

An intriguing alternative to building ever-larger monolithic models is to create large numbers of small, specialized agents that achieve intelligent behavior through communication and coordination—a form of swarm intelligence. Ant colonies and bee hives demonstrate that individually simple agents can produce sophisticated collective behavior. Could we create thousands of small AI agents that collaborate to solve complex problems?

For science applications, this maps naturally to having specialist agents in different fields—chemistry agents, physics agents, biology agents—that communicate with each other to tackle interdisciplinary research problems, much like a well-functioning research team. Key problems are how to make agents communicate efficiently and effectively and how to train a collective of agents (especially agents with different expertise) to work together and solve problems. For communication, standard cross-attention scales quadratically and is prohibitively expensive. Communication can probably be done just by exchanging embeddings, filtering out noise and transmitting only the "semantic essence" of the agent's state as one way of exchanging embeddings for communication between agents.

9. Self-Evaluation, Reflection, and Recursive Processing

Models should be able to evaluate their own outputs, reflect on their reasoning, and iteratively improve their solutions. This requires some form of recursion: the ability to take one's own output as input and refine it. Current autoregressive models generate tokens left to right without looking back; what we need is a system that can pause, assess, backtrack, and revise. The brain does this constantly—we re-read our own writing, reconsider our plans, and revisit our assumptions. Additionally, models should be able to run continuously without external input, engaging in what might be called autonomous contemplation: background processing that generates insights, reorganizes knowledge, and prepares for future tasks.

Recursive processing is essential for the scientific method itself. Real scientific discovery is inherently iterative: hypotheses are proposed, tested, refined, and sometimes discarded entirely. An AI Co-Scientist must be capable of this full cycle of scientific reasoning, including the ability to recognize when a research direction is unproductive and pivot to alternative approaches.

10. The Case for Neuromorphic AI

Many of the ideas outlined above converge on the need for neuromorphic computing hardware: chips designed to emulate the structure and dynamics of biological neural circuits rather than the von Neumann architecture that underlies conventional CPUs and GPUs. Newest developments in neuromorphic platforms have already demonstrated improvements in energy efficiency and real-time processing for specific tasks. Research published in Nature Communications (2025) shows hybrid neural networks inspired by corticohippocampal circuits can achieve continual learning on neuromorphic substrates. There should be significantly more research and investment in this direction. As Isaac Asimov imagined with his positronic brains, the future of AI may depend not just on better algorithms but on fundamentally different computing substrates that are co-designed with brain-inspired architectures.

For AI Co-Scientist systems specifically, neuromorphic computing could enable the always-on, continuous learning and background consolidation processes described throughout this document. The energy efficiency of neuromorphic platforms would make it practical for AI Co-Scientist agents to run continuously—monitoring literature, consolidating memories, and generating hypotheses even when not actively engaged by a human scientist—without the prohibitive energy costs of maintaining large GPU clusters in an always-active state.

11. Entropy-Based Real-Time Weight Updates

If we could measure the entropy of an entire network or subnetwork in real time, we could update weights without backpropagation by directly minimizing local entropy measures. The brain appears to perform something analogous—neurons adjust their synaptic strengths based on local activity patterns, not global error signals. Developing practical, scalable methods for entropy-based real-time weight updates could unlock truly online, always-learning AI systems that adapt to their environment moment by moment.

Entropy-based weight updates would enable the ultimate form of continual learning for AI Co-Scientist applications: an agent that refines its scientific reasoning capabilities in real time as it reads papers, runs experiments, and receives feedback from human collaborators. Rather than requiring periodic retraining cycles, the agent would continuously improve its capabilities through local, efficient parameter updates—becoming a better scientific partner with every interaction.

12. Conclusion

The ideas presented here share a common thread: AI needs to become more brain-like in concrete architectural and algorithmic choices. This means modular systems where specialized components communicate through shared embeddings; continual learning powered by local learning rules rather than global backpropagation; tight integration of memory and reasoning; offline memory consolidation; curriculum-based learning from imitation to mastery; and neuromorphic hardware co-designed with these algorithms. No single idea is sufficient, but together they sketch a path toward AI systems that can learn, adapt, create, and collaborate in ways that today's models cannot. The most powerful agentic AI—including AI for science—will emerge not from scaling current architectures but from rethinking the foundations of how neural networks are built, trained, and deployed.

The longer-term vision envisions AI Co-Scientist agents that operate continuously, consolidating knowledge during idle periods, generating hypotheses autonomously, and becoming increasingly capable research partners over years of collaboration with human scientists. Building such a system requires the full complement of brain-inspired principles outlined in this document, and its success will demonstrate that the path to more powerful AI lies through deeper understanding of intelligence itself.