Artificial General Intelligence — Ch.09: Chapter 09. Path to AGI [EN]

STATERA Library · Volume 47 · Chapter 09 · EN
Source: https://github.com/Statera-Library/01_Vol_47_Artificial_General_Intelligence/blob/main/Chapter_09_Path_to_AGI/Chapter_09_Path_to_AGI_EN.md

Volume 47. Artificial General Intelligence

Chapter 09. Path to AGI

09.00. Current State of AI

Artificial intelligence has entered a stage in which systems can perform an extraordinary range of cognitive tasks, yet their capabilities remain uneven and fundamentally different from general human intelligence. Modern AI can generate language, interpret images, synthesize audio and video, write software, solve mathematical problems, retrieve knowledge, and interact with external tools. These abilities represent major progress toward increasingly general computational intelligence, but they should not be confused with the existence of fully realized artificial general intelligence.

The dominant foundation of contemporary AI is the large-scale pretrained model. Instead of constructing a separate model for every task, modern systems learn broad statistical representations from enormous collections of text, images, video, audio, code, and increasingly multimodal data. A single foundation model can subsequently support many applications through prompting, instruction tuning, fine-tuning, retrieval, or tool integration. This transition has changed AI from collections of specialized algorithms into increasingly reusable general-purpose computational platforms.

Large language models demonstrate this transformation most clearly. Transformer-based systems learn rich representations of language and knowledge through large-scale prediction objectives and can perform tasks that were not explicitly represented during training. In-context learning allows models to adapt their behavior from instructions and examples supplied at inference time. More recent reasoning-oriented approaches additionally allocate computation during inference, enabling models to explore intermediate possibilities, revise candidate solutions, and improve performance on difficult mathematical, scientific, programming, and planning problems.

Multimodal AI is expanding these capabilities beyond language. Modern models can jointly process combinations of text, images, speech, video, spatial information, and other sensor signals. Shared or aligned representation spaces allow information from one modality to influence interpretation in another. This development is important for the path toward AGI because intelligence operating in the physical world cannot depend on language alone. General systems must connect linguistic concepts with perception, spatial relationships, temporal change, physical objects, actions, and consequences.

Another major transition is the emergence of agentic AI. Traditional models primarily transformed inputs into outputs, whereas AI agents can operate through repeated cycles of observation, reasoning, planning, action, and evaluation. They may call software tools, search databases, execute code, retrieve external information, manipulate digital environments, and coordinate multiple specialized models. Memory systems and retrieval-augmented generation further extend agents beyond the immediate context window by allowing relevant information from previous interactions or external knowledge stores to be recovered when needed.

Despite these advances, present AI still exhibits substantial limitations in reasoning reliability. Models can solve difficult problems while failing unexpectedly on simpler variants, and successful reasoning does not necessarily imply a stable internal understanding of the underlying problem. Hallucination, brittle generalization, sensitivity to prompting, incorrect assumptions, and inconsistent multi-step reasoning remain significant weaknesses. Increasing inference-time computation can improve performance, but additional computation alone does not guarantee correct reasoning or eliminate errors in the model\'s underlying representations.

Long-horizon autonomy remains another major gap between current AI and AGI. An intelligent system may need to pursue objectives over hours, days, months, or longer while maintaining goals, updating plans, remembering relevant experiences, recovering from failures, and responding to unexpected changes. Current agents can perform increasingly complex sequences of operations, but reliability usually decreases as tasks become longer and dependencies accumulate. Small errors in perception, reasoning, memory, planning, or tool use can propagate through an extended action sequence and eventually cause the entire task to fail.

Memory also remains fundamentally incomplete. Transformer context windows provide powerful temporary working spaces, while retrieval systems can approximate forms of external long-term memory. However, human-like memory involves more than storing and retrieving documents. Episodic experiences, semantic knowledge, procedures, priorities, contextual associations, and continual consolidation interact dynamically. An AGI system would likely require integrated memory mechanisms that determine what should be remembered, forgotten, generalized, retrieved, revised, and converted into reusable knowledge without requiring complete retraining.

World modeling represents another important frontier. Intelligent agents need internal representations that describe not merely what exists but how states evolve through time and how actions change future possibilities. Current generative and predictive models can learn substantial statistical structure from video, robotic trajectories, simulations, and multimodal observations. Nevertheless, robust prediction of complex physical and social environments remains difficult. Long-term uncertainty, hidden variables, causal structure, rare events, and interactions among multiple agents make realistic world modeling substantially harder than short-horizon pattern prediction.

Embodied AI exposes these limitations particularly clearly. Robots must transform noisy sensor measurements into estimates of objects, geometry, motion, affordances, risks, and possible actions while operating under real-time constraints. They must continuously connect perception with planning and control rather than generate isolated answers. Advances in robot foundation models, vision-language-action models, reinforcement learning, imitation learning, and world models are reducing the separation between digital intelligence and physical autonomy, but flexible manipulation and general-purpose real-world adaptation remain far behind human competence.

Current AI learning is also far less efficient than biological learning in many settings. Foundation models often require enormous datasets and computational resources before they acquire broad capabilities. Humans can frequently learn new concepts or behaviors from a few demonstrations, explanations, or direct experiences and can transfer knowledge across substantially different situations. Self-supervised learning, few-shot adaptation, meta-learning, reinforcement learning, continual learning, and synthetic data generation address parts of this problem, but no existing approach provides the flexible lifelong learning expected from general intelligence.

Knowledge and causality provide another dividing line. Large models encode extensive statistical regularities and can often produce explanations that resemble causal reasoning, yet correlation learned from observational data is not equivalent to discovering causal structure. General intelligence requires understanding interventions, counterfactual alternatives, mechanisms, and consequences. Integrating neural representation learning with causal models, symbolic reasoning, probabilistic inference, simulation, and experimentation may therefore become increasingly important as AI moves from recognizing patterns toward discovering how environments actually work.

The architecture of modern AI is consequently becoming more heterogeneous. Progress is no longer defined only by increasing the parameter count of a single neural network. Practical systems combine foundation models with retrieval, external memory, planning algorithms, code execution, search, specialized perception models, databases, knowledge representations, verification systems, simulators, and external tools. This suggests that increasingly general intelligence may emerge from coordinated cognitive architectures rather than from one monolithic model performing every computational function internally.

Scaling nevertheless remains one of the strongest drivers of capability. Larger datasets, greater compute, improved training algorithms, better model architectures, higher-quality data, and increased inference-time computation have repeatedly produced capabilities that were difficult to anticipate from smaller systems. At the same time, scaling faces economic, energy, hardware, memory, communication, and data-quality constraints. The central question is therefore shifting from whether scaling works to how far scaling can proceed and which missing capabilities require architectural or algorithmic breakthroughs beyond conventional scaling.

Evaluation has become correspondingly difficult. Benchmark scores in mathematics, coding, language understanding, science, or multimodal reasoning reveal important capabilities but do not by themselves establish general intelligence. Models may exploit patterns contained in training distributions, and benchmark performance can saturate without demonstrating robust transfer to unfamiliar environments. AGI evaluation therefore increasingly requires tests of generalization, adaptation, long-horizon planning, multimodal reasoning, autonomous task completion, uncertainty management, memory, tool use, and performance under genuinely novel conditions.

Safety and alignment are equally central to the current state of AI. More capable systems can produce greater benefits, but autonomy, tool access, persistent memory, and real-world control also increase the consequences of mistakes. Alignment methods, human feedback, constitutional constraints, interpretability, monitoring, sandboxing, permission systems, and human oversight are therefore becoming architectural requirements rather than optional additions. The challenge becomes more difficult as systems gain the ability to formulate plans and execute actions across increasingly complex environments.

The present state of AI can ultimately be understood as a transition from specialized machine intelligence toward integrated general-purpose intelligent systems. Foundation models have supplied broad representations, multimodal learning is connecting different forms of perception, reasoning methods are expanding deliberate computation, memory systems are extending temporal continuity, and agents are connecting cognition with action. Embodied systems and world models are beginning to connect these capabilities to physical environments, while evaluation and safety research attempt to determine how reliably such systems can operate.

The remaining distance to AGI cannot be measured by a single benchmark or model size. The critical challenge is integration: perception must connect with memory, memory with reasoning, reasoning with planning, planning with action, action with learning, and learning with persistent world knowledge. Meta-cognition must evaluate these processes, while safety mechanisms constrain their operation. The current generation of AI contains increasingly powerful pieces of this architecture, but the pieces are not yet unified into a consistently adaptive, autonomous, reliable, and generally intelligent system. This integration problem defines the starting point for the path from today\'s AI toward AGI.

09.01. From LLM to AGI

Large language models represent one of the strongest technological foundations for the transition toward artificial general intelligence, but an LLM alone should not be considered an AGI. LLMs provide broad linguistic knowledge, flexible representations, in-context adaptation, and increasingly powerful reasoning capabilities. The path toward AGI therefore involves transforming the language model from a primarily predictive model into a persistent intelligent system that can perceive, remember, reason, plan, act, learn, and evaluate itself across diverse environments.

The first major contribution of LLMs is their ability to construct highly reusable representations from large-scale data. Through pretraining, a model compresses statistical structures involving language, concepts, relationships, procedures, and fragments of world knowledge into distributed neural representations. These representations allow knowledge acquired in one context to influence performance in another. This broad transfer capability is important for AGI because general intelligence cannot depend on separately engineered models for every possible task or environment.

However, next-token prediction by itself does not provide all mechanisms required for general intelligence. Prediction can produce remarkable linguistic competence and implicitly capture substantial structure about the world, yet an autonomous intelligence must do more than generate plausible continuations. It must distinguish observation from assumption, maintain persistent goals, represent uncertainty, compare alternative futures, detect failures, revise beliefs, and choose actions according to their expected consequences. AGI therefore requires additional cognitive processes surrounding or extending the foundation model.

Reasoning is one of the most important extensions. Conventional LLM inference can generate solutions rapidly, but difficult problems often require deliberate computation over multiple intermediate states. Reasoning-oriented systems introduce mechanisms for decomposition, iterative inference, search, verification, reflection, and increased inference-time computation. Instead of treating the first generated answer as the final result, a system can construct candidate solutions, inspect inconsistencies, explore alternatives, and select stronger conclusions. This moves AI from immediate generation toward computational deliberation.

Planning extends reasoning from understanding problems to organizing future behavior. An AGI must convert abstract objectives into achievable sequences of actions while accounting for dependencies, resources, constraints, uncertainty, and changing conditions. LLMs can already generate plans, but reliable autonomy requires plans to remain connected to actual environment states. Execution results must therefore return to the planning system, allowing plans to be revised when assumptions fail. The result is a continuous loop of goal formation, planning, execution, observation, and replanning.

Tool use provides another bridge from LLMs toward general-purpose agents. A language model is limited by the information and computation represented within its parameters and current context, whereas an agent can access databases, search systems, software APIs, calculators, code interpreters, simulators, sensors, and external applications. Tools transform language-based reasoning into operational capability. The model no longer needs to internally reproduce every specialized function; instead, it can determine which resource is appropriate, invoke it, interpret the result, and incorporate that information into subsequent reasoning.

Memory changes an LLM interaction from an isolated inference episode into a temporally extended cognitive process. Context windows provide temporary working memory, while retrieval systems can supply information from external stores. More advanced architectures require episodic memory for previous experiences, semantic memory for accumulated knowledge, procedural memory for reusable skills, and working memory for current reasoning. Memory must also support consolidation, relevance estimation, updating, and forgetting so that experience becomes structured knowledge rather than an indefinitely growing collection of records.

Retrieval-augmented generation is therefore more than a method for improving factual answers. Within an AGI architecture, retrieval can connect the reasoning system to dynamically changing knowledge that does not need to reside permanently inside model parameters. An intelligent agent can retrieve documents, prior experiences, plans, environmental states, or specialized knowledge when required. This separation between parametric knowledge and external memory enables systems to update information more efficiently and provides a foundation for persistent knowledge management across long operational periods.

Multimodal perception expands the LLM beyond the symbolic world of text. General intelligence must interpret images, speech, video, spatial structure, motion, and potentially streams from cameras, LiDAR, radar, tactile sensors, proprioception, and other physical sensors. Multimodal models provide a mechanism for connecting linguistic concepts with perceptual experience. When these representations are aligned, an agent can reason not only about descriptions of the world but also about observations obtained directly from the environment, creating an essential bridge toward embodied intelligence.

World models extend perception by representing how environments change over time. Perception describes the current state, while a world model attempts to predict possible future states and the consequences of actions. This capability enables mental simulation, counterfactual reasoning, risk estimation, and model-based planning. An AGI cannot evaluate every possible action through real-world trial and error, so internal predictive models become important computational mechanisms for exploring alternatives before acting. LLM knowledge may contribute to such models, but language representations alone are insufficient for accurate physical dynamics.

Embodiment introduces an even stronger requirement for closed-loop intelligence. A robot or autonomous system receives continuous observations, chooses actions, changes its environment, and then observes the consequences. Intelligence therefore becomes a perception-action loop rather than a sequence of disconnected prompts. Vision-language-action models and robot foundation models are early examples of this transition. They attempt to connect semantic understanding with physical actions, but robust AGI will require substantially stronger spatial reasoning, temporal prediction, manipulation, control, adaptation, and safety mechanisms.

Learning must also continue after deployment. Most LLM knowledge is acquired during expensive offline training, while humans continually update their understanding through interaction and experience. AGI requires some form of continual learning that can acquire new concepts, procedures, preferences, environmental models, and skills without repeatedly retraining the entire foundation model. Memory, parameter-efficient adaptation, reinforcement learning, self-supervised learning, skill libraries, and meta-learning may provide complementary mechanisms for converting operational experience into progressively improved behavior.

Meta-cognition adds another layer by allowing the system to reason about its own cognitive processes. Instead of merely producing an answer, an intelligent agent should estimate confidence, recognize missing information, detect contradictions, determine whether additional computation is worthwhile, and decide when external verification or human assistance is necessary. Reflection and self-evaluation can improve performance, but reliable meta-cognition requires mechanisms that distinguish genuine error detection from another potentially incorrect generation produced by the same underlying model.

The transition toward AGI may therefore involve a shift from the concept of a large language model to that of a unified agent architecture. The LLM can function as an important reasoning, language, and knowledge component, while perception systems, memory, retrieval, planners, world models, learning mechanisms, tools, controllers, and verification modules provide complementary capabilities. These components must exchange information through coherent representations and operate within a continuous cognitive loop rather than as unrelated software modules.

Such integration also changes the role of scaling. Increasing model size, training data, and compute may continue to improve representations and reasoning, while inference-time computation can provide additional capability for difficult tasks. Yet architectural integration determines whether those capabilities can persist across time and influence real actions. Scaling and new mechanisms are therefore not necessarily competing paths. Larger foundation models may provide stronger cognitive primitives, while memory, world modeling, planning, embodiment, and continual learning organize those primitives into increasingly general intelligent systems.

Long-horizon reliability becomes the decisive engineering challenge as these components are integrated. An agent that succeeds with high probability on a single reasoning step can still fail frequently when thousands of dependent decisions are required. AGI therefore requires error detection, uncertainty estimation, redundant verification, checkpointing, recovery strategies, hierarchical planning, and mechanisms for escalating difficult situations. Intelligence must be evaluated not only by the quality of individual answers but by whether a system can maintain coherent and effective behavior throughout extended interactions with changing environments.

Safety and alignment must develop simultaneously with autonomy. A system capable of planning, tool use, persistent memory, learning, and physical action has a much larger operational impact than a model that only generates text. Permissions, behavioral constraints, verification, monitoring, interpretability, secure execution environments, human oversight, and controllable objectives must therefore become integrated parts of the architecture. The transition from LLM to AGI is consequently not simply a process of increasing capability; it is also a process of constructing reliable control around increasingly autonomous intelligence.

The path from LLMs toward AGI can ultimately be viewed as progressive integration around a powerful foundation model. Language modeling supplies broad representations and knowledge; reasoning supplies deliberate computation; memory supplies continuity; retrieval supplies dynamic knowledge; planning supplies goal-directed structure; tools supply external capabilities; multimodal perception connects the system to observations; world models provide prediction; embodiment connects cognition with action; continual learning provides adaptation; and meta-cognition provides internal evaluation and control.

No single component is sufficient by itself. General intelligence emerges only when these capabilities operate together across tasks, domains, environments, and time scales. The central transition is therefore from a model that answers questions to an agent that maintains an evolving internal state, understands its environment, formulates and pursues goals, learns from consequences, and adapts its behavior while remaining controllable. LLMs may provide the cognitive core of this transition, but AGI requires the surrounding architecture that turns broad predictive intelligence into persistent, grounded, learning, and autonomous intelligence.

09.02. Required Breakthroughs

Reaching artificial general intelligence requires more than incremental improvements in benchmark performance. Current AI already combines powerful foundation models, multimodal perception, reasoning, retrieval, tool use, and increasingly capable agents, yet these components remain incomplete and unreliable when operating across unfamiliar domains for long periods. The breakthroughs required for AGI therefore concern not only greater capability, but the integration, efficiency, adaptability, reliability, and control of intelligence as a persistent system.

A fundamental breakthrough is stronger generalization beyond training distributions. Modern models can transfer knowledge across many tasks, but their performance can deteriorate when environments, assumptions, representations, or problem structures differ substantially from familiar examples. AGI must infer reusable principles rather than depend primarily on statistical similarity. It should recognize structural relationships across domains and apply previously acquired knowledge to genuinely novel situations without requiring extensive retraining or carefully constructed demonstrations.

Reasoning reliability represents a second critical challenge. Current models can perform impressive mathematical, scientific, linguistic, and programming reasoning, yet their success remains inconsistent. More capable systems require mechanisms for decomposing problems, maintaining intermediate states, exploring alternatives, checking assumptions, verifying conclusions, and recovering from errors. Reasoning must become a controllable computational process rather than merely a plausible sequence of generated tokens, especially when decisions influence subsequent actions or real-world systems.

Causal understanding must advance alongside reasoning. Statistical prediction can reveal regularities, but general intelligence must distinguish correlation from mechanisms that actually produce events. An AGI should reason about interventions, understand how actions modify environments, and evaluate counterfactual possibilities such as what would happen under alternative decisions. Combining representation learning with causal inference, experimentation, simulation, and structured models may allow intelligent systems to move from recognizing recurring patterns toward understanding why changes occur.

World models require a related breakthrough in temporal and physical understanding. General agents need internal models that represent objects, agents, spatial relationships, dynamics, uncertainty, and possible future states. Such models should predict not only immediate observations but longer trajectories and consequences of actions. Accurate world models could support mental simulation, planning, risk assessment, and counterfactual reasoning, allowing an agent to evaluate many possibilities internally before committing expensive or dangerous actions in the external world.

Persistent memory is another missing capability. Current context windows and retrieval systems provide useful approximations, but AGI requires memory that remains coherent across long periods of interaction. Episodic experiences must be connected with semantic knowledge, reusable procedures, goals, and contextual relationships. The system must decide what information deserves storage, what can be compressed or generalized, when memories should be retrieved, and when outdated knowledge should be modified or forgotten. Memory must become an active learning mechanism rather than passive storage.

Continual learning must transform this accumulated experience into improved competence. Most foundation models acquire their fundamental capabilities through large offline training processes and change relatively little during deployment. AGI should learn continuously from interaction while avoiding catastrophic forgetting, uncontrolled behavioral drift, and corruption from unreliable observations. Combining external memory, self-supervised learning, reinforcement learning, parameter-efficient adaptation, skill acquisition, and periodic consolidation may provide a path toward stable lifelong learning.

Learning efficiency itself requires substantial improvement. Humans can acquire useful concepts from a few examples and frequently generalize from explanations, analogies, observation, and experimentation. Modern AI often depends on enormous datasets and computational budgets. AGI will require stronger few-shot learning, active learning, transfer, compositional generalization, meta-learning, and autonomous data selection. An intelligent system should increasingly determine what it does not know and seek experiences or information that most efficiently reduce its uncertainty.

Meta-learning could provide an important breakthrough by shifting intelligence from learning particular tasks toward learning how to learn. Instead of requiring a fixed training procedure for every new domain, a meta-learning system could adapt its representations, strategies, memory usage, and learning processes according to the problem encountered. This capability is closely related to cognitive flexibility because a general agent must recognize when previously successful strategies are inappropriate and construct new approaches without complete redesign by human engineers.

Meta-cognition is equally important for reliable autonomy. AGI should possess mechanisms for evaluating its own reasoning, knowledge, uncertainty, and limitations. It should recognize when evidence is insufficient, when a plan is failing, when additional computation is justified, and when external assistance should be requested. Current reflection methods provide early versions of these abilities, but genuine meta-cognitive control requires reliable confidence estimation, error detection, strategy selection, computational resource allocation, and verification mechanisms that are not simply repetitions of the original reasoning process.

Long-horizon planning represents another major technical barrier. Real objectives often contain hierarchical dependencies extending across thousands of decisions. AGI must maintain high-level goals while decomposing them into intermediate objectives and executable actions, monitoring progress and replanning when circumstances change. Hierarchical planning, memory, world models, verification, and feedback control must therefore operate together. The system must remain coherent even when plans extend far beyond a single inference context or encounter unexpected environmental changes.

Embodied intelligence requires breakthroughs in grounding these cognitive capabilities in physical interaction. A general robot must combine vision, language, spatial understanding, proprioception, manipulation, navigation, planning, and control while responding continuously to uncertain environments. Physical actions cannot simply be regenerated when incorrect because mistakes may damage objects, robots, or people. Embodied AGI therefore requires reliable perception-action loops, predictive world models, uncertainty-aware planning, rapid adaptation, and real-time control operating under strict safety constraints.

Another breakthrough concerns unified representations across cognitive functions. Current AI architectures often connect separately developed modules for perception, language, memory, planning, control, and retrieval. AGI requires these components to exchange information without losing critical semantic, spatial, temporal, or causal structure. Shared latent representations, structured world states, multimodal embeddings, knowledge graphs, or hybrid neural-symbolic representations may provide different parts of this bridge. The objective is not necessarily one universal representation, but coherent communication among specialized cognitive processes.

Neural-symbolic integration may become particularly valuable where flexible pattern recognition must coexist with explicit reasoning. Neural networks excel at learning distributed representations from complex data, whereas symbolic systems can provide compositional structures, variables, rules, and interpretable constraints. AGI may require architectures capable of moving between these forms rather than choosing one exclusively. Such systems could ground symbols in learned representations while using structured reasoning to manipulate relationships that pure statistical generation handles unreliably.

Autonomous skill acquisition is another necessary step. Instead of learning every behavior directly from human demonstrations, a general agent should discover useful reusable skills through interaction. Complex tasks could then be constructed by composing simpler learned behaviors. Reinforcement learning, imitation learning, self-supervised prediction, exploration, simulation, and automated curriculum generation may contribute to this capability. Skill libraries could gradually transform repeated experiences into procedural knowledge that planning systems can retrieve and combine for new objectives.

Self-improvement extends this principle from individual skills to the intelligent system itself. Future agents may evaluate their failures, identify missing capabilities, generate training experiences, improve prompts or strategies, reorganize memories, and optimize parts of their computational workflow. More advanced systems might modify models or architectures under controlled conditions. However, self-improvement becomes useful for AGI only when improvements are measurable, stable, reversible when necessary, and constrained by safety mechanisms that prevent optimization from producing unintended behavior.

Evaluation must also undergo a conceptual breakthrough. Static benchmarks cannot adequately measure systems capable of learning, using tools, interacting with environments, and adapting over time. AGI evaluation must examine generalization to unseen problems, long-horizon reliability, continual learning, multimodal understanding, causal reasoning, autonomous skill acquisition, uncertainty management, and recovery from failure. Interactive environments and open-ended tasks are therefore likely to become increasingly important alongside conventional tests.

Efficiency and computational architecture form another practical barrier. General intelligence cannot depend indefinitely on unrestricted increases in model size and inference cost. Better algorithms, sparse computation, mixture-of-experts architectures, efficient memory systems, specialized accelerators, model compression, distributed computing, and adaptive inference can reduce the resources required for intelligent behavior. Future systems may dynamically allocate computation, using inexpensive processing for predictable situations while activating deeper reasoning only when uncertainty, novelty, or risk demands it.

Safety and alignment constitute breakthroughs that must occur together with capability development rather than after AGI has been achieved. Increasing autonomy magnifies the consequences of incorrect objectives, deceptive information, unsafe exploration, or uncontrolled tool use. Reliable systems require interpretable goals, permission boundaries, monitoring, verification, robustness, secure execution, human oversight, and mechanisms for interrupting or correcting behavior. Alignment must remain effective even as systems learn new skills, modify strategies, and encounter situations not anticipated during development.

The required breakthroughs are therefore deeply interconnected. Better world models improve planning; better memory enables continual learning; causal reasoning strengthens prediction; meta-cognition improves verification; embodiment supplies grounded experience; and efficient learning accelerates adaptation. Progress in one component changes the requirements of others. AGI is unlikely to emerge from simply assembling independent modules unless their interactions form a coherent cognitive architecture capable of maintaining information, goals, learning, and control across time.

The path to AGI can consequently be understood as a transition from powerful but fragmented capabilities toward integrated adaptive intelligence. Scaling may continue to strengthen foundation models, but major progress must also occur in generalization, reasoning, causality, memory, world modeling, continual learning, meta-cognition, planning, embodiment, representation, efficiency, and safety. The decisive breakthrough may ultimately be architectural: creating a system in which these capabilities reinforce one another within a persistent perception-reasoning-action-learning loop while remaining reliable and controllable across unfamiliar environments.

09.03. Scaling vs New Paradigms

The path toward artificial general intelligence is often framed as a competition between scaling existing models and discovering fundamentally new paradigms. Scaling argues that larger models, richer datasets, greater compute, and longer inference can continue producing increasingly general capabilities. New-paradigm approaches argue that important properties of intelligence---such as persistent memory, causal understanding, world modeling, continual learning, embodiment, and meta-cognition---may require mechanisms that cannot emerge reliably from scale alone. The likely path to AGI may involve both directions rather than a simple choice between them.

Scaling has already demonstrated extraordinary empirical effectiveness. Increasing model parameters, training data, computational resources, and optimization quality has repeatedly improved language understanding, knowledge representation, coding, multimodal perception, and reasoning. Capabilities that once required specialized systems can now appear within a common foundation model. This suggests that sufficiently large models can discover representations and computational patterns that engineers did not explicitly program, making scaling one of the strongest known mechanisms for producing broad machine intelligence.

Scaling laws provide a useful framework for understanding this progress. Model performance often changes predictably as training compute, dataset size, and model capacity increase, allowing researchers to estimate how additional resources may improve loss or downstream capability. These relationships have encouraged the construction of increasingly large training systems and datasets. However, scaling laws describe empirical trends within particular architectures and training regimes; they do not establish that every capability required for AGI will improve indefinitely through the same process.

Data scaling presents an especially important limitation. Increasing model size is useful only when training data provide sufficient quality, diversity, and informational value. Repeating low-quality or redundant information cannot indefinitely produce stronger intelligence. Future scaling therefore increasingly depends on data curation, synthetic data, interactive experience, multimodal observations, simulations, and automatically generated training environments. The problem shifts from simply collecting more data toward generating the right experiences from which increasingly capable systems can learn.

Compute scaling faces similar constraints. Larger training runs require enormous accelerator clusters, memory capacity, communication bandwidth, energy, cooling, and financial investment. Even if hardware continues improving, unrestricted growth becomes increasingly expensive. This creates pressure for algorithmic efficiency, sparse computation, mixture-of-experts models, quantization, specialized accelerators, improved optimization, and hardware-software co-design. The future of scaling is therefore likely to involve obtaining more useful intelligence per unit of computation rather than simply maximizing raw compute.

Inference-time scaling provides another dimension. Instead of allocating all computation during training, systems can spend additional resources when solving difficult problems. Search, multiple candidate generation, verification, reflection, simulation, and iterative reasoning allow a model to devote more computation to situations requiring deeper analysis. This resembles adaptive cognition: routine problems may receive inexpensive processing, while uncertain or high-risk problems receive more extensive reasoning. Such dynamic computation could become an important bridge between conventional foundation models and more flexible cognitive systems.

Nevertheless, scale does not automatically solve persistent memory. A model with more parameters may encode more knowledge, and a larger context window may process longer sequences, but neither mechanism necessarily provides lifelong memory management. General intelligence requires deciding what should be remembered, consolidated, generalized, updated, or forgotten. Episodic, semantic, procedural, and working memory perform different computational roles. These functions may require explicit memory architectures that complement large pretrained models rather than simply expanding their parameter counts.

World modeling raises a similar issue. Large multimodal models can learn substantial regularities about objects, actions, space, and temporal change, suggesting that scaling can contribute significantly to internal models of the world. Yet reliable agents must predict consequences across long horizons, represent uncertainty, distinguish hidden causes, and reason about interventions. Explicit predictive state models, simulators, latent dynamics, causal representations, and model-based planning may therefore be required to transform broad statistical knowledge into operational understanding of changing environments.

Causal reasoning is frequently presented as another area where new paradigms may be necessary. Foundation models learn primarily from statistical relationships in observed data, whereas intelligent intervention requires understanding what changes when an agent actively modifies a system. Correlations may support prediction without revealing underlying mechanisms. Structural causal models, counterfactual reasoning, experimentation, and active learning offer frameworks for representing such relationships. Neural models and causal methods may ultimately become complementary rather than competing approaches.

Continual learning also challenges the conventional train-once-and-deploy paradigm. Scaling typically concentrates learning into a large offline optimization process, but AGI must continue adapting after deployment. New knowledge, environments, tools, goals, and skills will appear continuously. Repeatedly retraining a massive foundation model is inefficient and potentially disruptive. Persistent memory, parameter-efficient adaptation, modular learning, skill acquisition, reinforcement learning, and controlled consolidation may therefore become essential architectural extensions to scaled foundation models.

Embodiment creates requirements that cannot be reduced easily to static dataset scaling. Physical agents operate in closed loops in which actions alter subsequent observations. They must learn affordances, dynamics, spatial relationships, uncertainty, and consequences through interaction. Large-scale video and robot datasets can provide powerful priors, but real-world intelligence requires adaptation to environments that cannot be completely represented in advance. Interaction, exploration, simulation, reinforcement learning, and world models therefore introduce learning paradigms in which the agent participates actively in generating its own data.

New paradigms may also involve modular cognitive architectures. Instead of expecting one network to internally implement every capability, future systems may coordinate specialized components for perception, memory, reasoning, planning, retrieval, world modeling, verification, and control. A foundation model could remain the central representational and reasoning engine while external modules provide capabilities that benefit from explicit structure. Such architectures resemble integrated cognitive systems more than isolated predictive models and may provide greater reliability, interpretability, and engineering flexibility.

Neuro-symbolic systems represent one possible architectural direction. Neural networks provide flexible learning from high-dimensional data, while symbolic representations provide explicit variables, rules, relationships, and compositional structures. Pure symbolic AI historically struggled with perception and knowledge acquisition, whereas purely neural systems can struggle with systematic reasoning and explicit constraint satisfaction. Combining these approaches could allow learned representations to remain flexible while structured reasoning provides stronger consistency for tasks requiring precise manipulation of relationships.

Meta-learning introduces another paradigm shift by treating adaptation itself as something that can be learned. Conventional scaling attempts to encode increasingly broad competence into model parameters, whereas meta-learning seeks systems that rapidly construct new competence from limited experience. An AGI may need both: extensive pretrained knowledge and mechanisms that determine how to adapt when existing knowledge is insufficient. Learning strategies, exploration policies, memory usage, and even computational allocation could become adaptive components rather than permanently fixed design choices.

Self-improvement pushes this idea further. A sufficiently capable agent might analyze its own failures, identify missing knowledge, generate training tasks, reorganize memory, discover better strategies, or improve parts of its software and models. This differs fundamentally from conventional scaling because the improvement process becomes partly internal to the intelligent system. However, uncontrolled self-modification creates severe reliability and alignment problems. Any self-improving architecture would require evaluation, verification, rollback, permission boundaries, and stable objectives before modifications could be accepted.

Efficiency may itself become a new intelligence paradigm. Biological brains do not apply maximum computation uniformly to every sensory input. Attention, prediction, novelty detection, and hierarchical processing allocate resources selectively. Future AI could similarly maintain low-cost processing during predictable conditions and activate expensive reasoning only when uncertainty, novelty, conflict, or risk increases. Conditional computation, sparse activation, event-driven processing, hierarchical agents, and adaptive inference could produce systems whose effective intelligence grows faster than their average computational consumption.

The debate also depends on what researchers mean by emergence. Scaling can produce abilities that appear suddenly when measured by particular tasks, even though underlying improvements may have accumulated gradually. Such observations suggest that future scale could reveal capabilities not yet visible in smaller systems. However, emergence cannot be assumed to solve every missing mechanism. If an architecture lacks persistent interaction, memory updating, environmental feedback, or action channels, increasing scale may improve internal prediction without automatically creating these system-level functions.

Evaluation is therefore essential for distinguishing genuine progress from apparent progress. If AI is measured only through static language, coding, mathematics, or knowledge benchmarks, scaling may appear sufficient because those evaluations closely match the strengths of foundation models. AGI evaluation must additionally examine novel-domain adaptation, long-horizon autonomy, persistent memory, causal intervention, embodied interaction, continual learning, uncertainty management, and recovery from failure. Different evaluations may reveal which capabilities scale naturally and which require architectural innovation.

Safety provides another reason not to treat capability scaling as an isolated objective. Larger and more autonomous systems can create greater consequences when reasoning, goals, or actions are incorrect. New architectures must therefore integrate alignment, verification, monitoring, permissions, interpretability, secure execution, and human oversight alongside capability improvements. A system that becomes more powerful without becoming proportionally more reliable and controllable does not necessarily represent progress toward usable AGI. Capability and control must scale together.

The most productive interpretation of scaling versus new paradigms is consequently not either-or but scale-plus-architecture. Scaling can continue improving foundation models, representations, knowledge, multimodal understanding, and reasoning primitives. New paradigms can provide memory, world models, causal structures, continual learning, embodiment, meta-cognition, adaptive computation, and reliable control. Each approach addresses different parts of the general intelligence problem, and advances in one can amplify the effectiveness of the other.

AGI may therefore emerge from increasingly powerful foundation models embedded within increasingly sophisticated cognitive architectures. Scaling supplies broad learned competence, while architectural innovations organize that competence into persistent perception-reasoning-planning-action-learning loops. The central research question is no longer simply how large models should become, but which capabilities should be learned implicitly through scale and which should be represented explicitly through new mechanisms. Discovering the correct division---and the interfaces connecting both sides---may define the next major stage in the path toward general intelligence.

09.04. Timeline and Predictions

Predicting the timeline to artificial general intelligence is inherently uncertain because AGI has no universally accepted operational definition and progress does not follow a simple linear trajectory. Improvements in foundation models, reasoning, multimodal perception, agents, robotics, memory, and world models are occurring simultaneously, but advances in one capability do not guarantee equivalent progress in others. Any AGI timeline should therefore be treated as a set of conditional scenarios rather than a precise forecast of a specific arrival date.

Historical AI development demonstrates why straightforward extrapolation is difficult. Periods of gradual improvement have repeatedly been interrupted by major architectural innovations, new training methods, hardware advances, larger datasets, and unexpected emergent capabilities. Conversely, apparently rapid progress can encounter difficult bottlenecks that require years of research. Future AGI development may similarly alternate between scaling-driven improvement, algorithmic breakthroughs, integration challenges, and periods in which reliability advances more slowly than raw capability.

The near-term trajectory is likely to be dominated by increasingly capable foundation models rather than the sudden appearance of a completely different form of intelligence. Models will continue becoming more multimodal, reasoning-oriented, tool-capable, and integrated with external memory and retrieval. Inference-time computation will allow systems to spend more resources on difficult problems, while smaller specialized models will handle predictable tasks efficiently. The distinction between a language model and an AI agent will consequently become increasingly blurred.

Agentic systems are likely to represent one of the most visible stages in this transition. Instead of responding to individual prompts, systems will increasingly maintain goals, construct plans, use software tools, execute workflows, inspect results, and revise actions. This progression can produce economically important autonomy before full AGI exists. However, increasing task duration exposes accumulated errors, making long-horizon reliability, verification, memory, permissions, and recovery mechanisms central engineering challenges for the next generation of agents.

Multimodal intelligence is also likely to deepen substantially before AGI. Text, images, audio, video, spatial representations, and sensor data will increasingly be processed within shared architectures. This will allow models to reason over richer descriptions of environments and connect abstract language with observable events. Video understanding may become particularly important because temporal data provide information about objects, motion, interaction, causality, and physical change that static text and images represent only indirectly.

World models may emerge as another major intermediate stage. Systems that can predict how environments evolve and estimate the consequences of alternative actions would support planning, simulation, robotics, and autonomous decision-making. Early world models may remain domain-specific or operate reliably only over short horizons, but increasing integration with multimodal foundation models could progressively expand their scope. The transition from recognizing states to predicting state transitions may become one of the most important milestones toward more general agents.

Memory architectures are likely to develop in parallel. Larger context windows will continue to provide useful temporary memory, but persistent agents require mechanisms for storing experiences across sessions and long operational periods. Retrieval, episodic records, semantic knowledge, procedural skills, and memory consolidation may gradually become standard components of agent architectures. This development would allow AI systems to accumulate operational experience without requiring every new piece of information to be incorporated through expensive foundation-model retraining.

Continual learning represents a more difficult milestone and may develop more slowly. Current systems can update external memory and adapt through limited fine-tuning, but safely modifying core capabilities during deployment remains challenging. Future agents may initially separate relatively stable pretrained models from rapidly changing external memory and skill modules. Over time, controlled consolidation mechanisms could transfer selected experience into adaptable model components, creating systems that learn continuously while limiting catastrophic forgetting and undesirable behavioral drift.

Reasoning performance will probably continue improving through a combination of stronger pretrained models and increased inference-time computation. Systems may dynamically decide whether a task requires immediate generation, extended reasoning, external search, simulation, code execution, or independent verification. This adaptive allocation of computation could make future AI substantially more capable without requiring maximum reasoning effort for every problem. Reliability, however, will depend increasingly on verification architectures rather than on reasoning strength alone.

Embodied intelligence is likely to progress on a different timeline from digital agents. Software environments provide structured interfaces and allow errors to be reversed relatively cheaply, whereas robots operate under physical uncertainty, real-time constraints, hardware limitations, and safety requirements. Foundation models, vision-language-action models, imitation learning, reinforcement learning, simulation, and world models will improve robotic adaptability, but general-purpose physical competence may lag behind broad digital cognitive capability.

This difference suggests that AGI may not arrive as a single synchronized threshold across all domains. Systems could achieve extremely broad competence in language, science, programming, digital tool use, and knowledge work while remaining significantly weaker than humans in flexible physical interaction. Alternatively, AGI definitions emphasizing cognitive generality might classify such systems as general intelligence before robust embodied intelligence exists. Timeline predictions therefore depend strongly on which capabilities are considered necessary for the term AGI.

Scientific and engineering automation could accelerate the timeline indirectly. AI systems that assist with algorithm discovery, software engineering, chip design, synthetic data generation, experiment planning, and model evaluation can contribute to improving future AI systems. As these tools become more capable, portions of AI research itself may become increasingly automated. This creates the possibility of feedback loops in which stronger AI accelerates research that produces still stronger AI, although the magnitude and stability of such acceleration remain highly uncertain.

Hardware development provides another source of uncertainty. Improvements in accelerator performance, memory bandwidth, networking, packaging, energy efficiency, and specialized inference hardware can expand the feasible scale of both training and deployment. At the same time, power availability, fabrication capacity, infrastructure cost, and communication bottlenecks may constrain growth. Algorithmic efficiency can partially compensate for these limits, meaning that future capability may depend as much on better use of computation as on the absolute quantity of available hardware.

Data availability will similarly influence future timelines. High-quality human-generated text alone cannot support unlimited scaling, encouraging increased use of multimodal data, synthetic data, simulations, generated curricula, and interactive environments. Future systems may participate directly in producing their own training experiences by identifying uncertainties, generating hypotheses, performing simulations, or selecting informative interactions. The transition from passive dataset consumption toward active experience generation could substantially alter the relationship between data scale and intelligence.

Several bottlenecks could delay AGI even if benchmark capabilities continue improving rapidly. Long-horizon consistency, causal understanding, continual learning, persistent identity and memory, robust world modeling, reliable self-evaluation, and safe autonomous action remain unresolved. Progress may therefore produce systems that appear highly general in short interactions while still failing during extended autonomous operation. The difference between impressive demonstrations and dependable intelligence at scale will become increasingly important when evaluating claims that AGI has been achieved.

Safety and governance may also affect practical deployment timelines independently of technical capability. Systems capable of autonomous planning, code execution, financial actions, infrastructure control, scientific experimentation, or physical operation require stronger monitoring and permission structures than conversational models. Consequently, a technically capable system may exist before society permits unrestricted autonomous deployment. AGI development and AGI deployment should therefore be understood as related but distinct timelines shaped by engineering, regulation, economics, and institutional trust.

Predictions should also distinguish between prototype AGI and mature AGI. A prototype might demonstrate broad cross-domain learning and reasoning while remaining expensive, slow, unreliable, or heavily supervised. Mature systems would require substantially greater robustness, efficiency, continual learning, safety, and operational independence. Similar distinctions occur throughout technology development: demonstrating that a capability is possible is different from transforming it into dependable infrastructure that can operate repeatedly under diverse real-world conditions.

For this reason, milestone-based forecasting is often more informative than assigning a particular year to AGI. Relevant milestones include reliable multimodal reasoning, persistent agent memory, autonomous tool use, long-horizon task completion, robust world models, continual learning, general skill transfer, embodied adaptation, self-evaluation, and safe autonomous operation. Observing how quickly these capabilities improve and integrate provides stronger evidence about progress toward AGI than improvements on isolated benchmarks.

Multiple scenarios therefore remain plausible. Continued scaling combined with inference-time reasoning could produce unexpectedly rapid progress if missing capabilities emerge naturally from increasingly powerful models. A hybrid trajectory may instead dominate, with scaled foundation models combined with explicit memory, world models, planning, causal reasoning, and learning architectures. A slower scenario is also possible if reliability, embodiment, continual learning, energy, data, or safety constraints prove substantially harder than current capability trends suggest.

The most important prediction is consequently structural rather than chronological. Future AI systems are likely to evolve from isolated foundation models toward persistent cognitive architectures combining perception, memory, reasoning, planning, tools, world models, action, learning, meta-cognition, and safety mechanisms. Different components may progress at different rates, and the point called AGI may emerge gradually as their integration crosses practical thresholds of generality, autonomy, adaptability, and reliability rather than appearing as one clearly identifiable event.

The AGI timeline should therefore be understood as a moving frontier rather than a countdown. Scaling, architectural innovation, hardware, data, autonomous research, embodiment, safety, economics, and governance will interact in ways that cannot be forecast precisely. What can be evaluated more reliably is the sequence of capabilities being assembled. As AI progresses from prediction toward reasoning, from reasoning toward agency, from agency toward persistent learning, and from isolated competence toward integrated adaptive intelligence, the boundary between advanced AI and AGI will increasingly become a question of demonstrated system-level capability rather than a predicted date.

09.05. Research Roadmap

A research roadmap toward artificial general intelligence should not be organized around a single model, benchmark, or predicted arrival date. AGI is a systems-level objective requiring coordinated progress in perception, memory, reasoning, planning, learning, action, and meta-cognition. Research must therefore advance individual capabilities while simultaneously studying how they interact inside persistent agents. The roadmap is best understood as progressive integration from powerful foundation models toward unified, adaptive, reliable, and controllable intelligent systems.

Foundation models remain an essential starting point because they provide broad representations of language, knowledge, code, images, audio, video, and increasingly other modalities. Research should continue improving data quality, model architectures, optimization, multimodal alignment, context processing, and computational efficiency. Scaling remains valuable, but its objective should increasingly shift from maximizing parameter count toward maximizing useful capability, generalization, reasoning quality, and knowledge transfer per unit of data and computation.

Representation learning must evolve toward representations that remain useful across different cognitive processes. AGI requires information about objects, concepts, relationships, space, time, actions, uncertainty, and causality to move between perception, reasoning, memory, planning, and control. Research should investigate shared latent spaces, structured representations, multimodal embeddings, object-centric representations, knowledge structures, and neural-symbolic interfaces. The objective is coherent information exchange rather than forcing every cognitive function into one representation format.

Multimodal perception forms the next major research layer. General agents must integrate language with images, video, speech, spatial information, and physical sensor streams rather than treating each modality independently. Research should emphasize temporal perception, cross-modal grounding, spatial reasoning, object permanence, event understanding, and uncertainty estimation. For embodied systems, this must eventually extend to depth, LiDAR, radar, tactile sensing, proprioception, and other observations required for robust interaction with physical environments.

Memory research should move beyond longer context windows toward persistent cognitive memory architectures. Working memory must support immediate reasoning, episodic memory must preserve relevant experiences, semantic memory must organize accumulated knowledge, and procedural memory must retain reusable skills. Research is needed on consolidation, retrieval, compression, forgetting, conflict resolution, provenance, and memory updating. A general agent should progressively transform experience into useful knowledge without allowing its memory system to become an uncontrolled archive of previous interactions.

Reasoning research must focus increasingly on reliability rather than impressive isolated solutions. Systems should learn when to answer immediately, when to decompose a problem, when to search, when to execute code, when to simulate alternatives, and when independent verification is necessary. Inference-time computation, search, reflection, formal reasoning, probabilistic inference, and external verification should be investigated as complementary mechanisms. The central objective is to transform reasoning into a controllable process whose errors can be detected and corrected.

Causal reasoning should become a distinct research direction within this architecture. Statistical prediction is powerful, but autonomous agents need to understand how interventions change environments. Research should connect learned representations with causal discovery, structural causal models, experimentation, counterfactual reasoning, and active learning. Such capabilities would allow agents to distinguish observed correlations from actionable mechanisms and could substantially improve scientific reasoning, diagnosis, planning, and decision-making under unfamiliar conditions.

World-model research should connect perception, causality, prediction, and planning. Agents need internal models capable of representing current states, latent variables, dynamics, uncertainty, other agents, and possible future trajectories. Research should explore predictive latent-state models, multimodal dynamics, hierarchical temporal models, generative simulation, and action-conditioned prediction. World models should eventually support mental simulation in which alternative strategies can be evaluated internally before expensive, irreversible, or dangerous actions are performed externally.

Planning research must extend from generating plausible plans toward maintaining executable plans over long horizons. Hierarchical planning can connect abstract objectives with intermediate goals, skills, and low-level actions, while feedback mechanisms continuously compare expected and observed outcomes. Research should address replanning, constraint satisfaction, resource allocation, uncertainty, partial observability, failure recovery, and coordination among multiple objectives. Planning must operate as a closed-loop process connected continuously to perception and world-state updates.

Tool use provides a practical pathway toward broader capability without requiring every function to exist inside a foundation model. Agents should learn to select and operate search systems, databases, software APIs, code execution environments, simulators, scientific instruments, and specialized models. Research must address tool discovery, interface understanding, result validation, permission management, and failure handling. Tool competence should ultimately become compositional, allowing agents to construct new workflows from previously learned resources when confronting unfamiliar tasks.

Continual learning is necessary to transform agents from static pretrained systems into adaptive intelligence. Research should investigate how operational experience can update memory, skills, representations, and selected model components without catastrophic forgetting or uncontrolled behavioral drift. Self-supervised learning, reinforcement learning, parameter-efficient adaptation, modular updates, replay, and periodic consolidation may provide complementary mechanisms. The system should learn throughout deployment while preserving previously validated capabilities and safety constraints.

Meta-learning should complement continual learning by enabling systems to improve the process of adaptation itself. Instead of applying one fixed learning strategy everywhere, an agent should determine which representations, experiences, tools, or exploration strategies are appropriate for a new problem. Research should examine few-shot adaptation, learned optimization, curriculum generation, active information gathering, and reusable learning strategies. This could reduce dependence on enormous task-specific datasets and enable faster adaptation to genuinely unfamiliar environments.

Meta-cognition should provide supervisory control across the entire cognitive architecture. A general agent needs mechanisms for estimating confidence, detecting knowledge gaps, monitoring reasoning quality, identifying plan failure, allocating computational resources, and deciding when external assistance is necessary. Research should separate generation from evaluation where possible and develop independent verification mechanisms. Meta-cognition should ultimately determine not only what the system thinks, but how much computation and evidence a particular decision deserves.

Embodied intelligence provides the critical test of whether representations and reasoning are grounded in interaction. Robotics research should integrate multimodal perception, world models, planning, manipulation, navigation, control, reinforcement learning, and adaptation within real-time perception-action loops. Simulation and large robot datasets can provide initial experience, but systems must eventually learn safely from physical interaction. Embodied research also exposes weaknesses in uncertainty handling, causality, temporal prediction, and recovery that may remain hidden in purely digital environments.

Autonomous skill acquisition should allow increasingly capable agents to discover reusable behaviors rather than depend entirely on manually defined tasks. Exploration, imitation learning, reinforcement learning, self-supervised prediction, simulation, and automated curriculum generation can support the discovery of skills. These behaviors can be compressed into procedural memory or skill libraries and later composed by planning systems. Research should determine how useful skills are identified, represented, validated, transferred, combined, and refined across different environments.

Multi-agent research extends general intelligence beyond an isolated agent. Many real-world tasks require cooperation, negotiation, specialization, communication, and distributed decision-making. Research should examine how agents share knowledge, divide tasks, coordinate plans, resolve conflicts, and maintain collective memory. Multi-agent systems can also provide environments for studying social reasoning and emergent organization. However, increasing interaction introduces new challenges involving communication reliability, incentives, security, and the propagation of errors among agents.

Self-improvement should be approached gradually and under strict experimental control. Early systems can analyze failures, reorganize memories, generate practice tasks, improve workflows, and propose changes to strategies without directly modifying their core architecture. More advanced research may investigate controlled model or software improvement. Every modification should be evaluated against independent tests, safety constraints, and previous capabilities, with rollback mechanisms available. Self-improvement without reliable evaluation would amplify errors rather than intelligence.

Computational efficiency must develop as a parallel research track. AGI cannot depend on applying maximum computation continuously to every input. Sparse models, mixture-of-experts systems, conditional computation, model compression, specialized hardware, distributed inference, and adaptive reasoning can reduce cost. Agents should eventually allocate computation according to novelty, uncertainty, difficulty, and risk. Predictable situations may require minimal processing, while unfamiliar or consequential situations activate deeper perception, reasoning, simulation, and verification.

Evaluation research must evolve together with the architecture. Static benchmarks should be supplemented with tests of novel-domain generalization, multimodal understanding, memory persistence, causal reasoning, tool use, long-horizon planning, continual learning, embodied adaptation, failure recovery, and autonomous task completion. Evaluation should measure not only whether a system succeeds, but how reliably it succeeds across repeated trials and changing environments. Open-ended interactive environments will become increasingly important for identifying hidden weaknesses in apparently capable systems.

Safety, alignment, and control must operate as continuous research layers rather than final-stage additions. Permission systems, monitoring, interpretability, verification, robustness, secure execution, human oversight, and controllable objectives should be integrated as autonomy increases. Continual learning and self-improvement create additional requirements because safety properties must remain stable while the system changes. Research should therefore treat capability growth and controllability as coupled objectives whose progress must be evaluated together.

The roadmap ultimately converges on unified agent architectures integrating perception, memory, reasoning, world models, planning, tools, action, learning, and meta-cognition within continuous feedback loops. Different components may remain specialized, but they must exchange coherent representations and operate toward persistent goals. Research should progressively increase task duration, environmental diversity, autonomy, and learning requirements while measuring reliability. Integration itself becomes an experimental discipline rather than merely the final engineering step.

Progress toward AGI should therefore be measured through capability milestones rather than a predetermined date. Stronger foundation models provide the cognitive substrate; memory and world models provide continuity and prediction; reasoning and planning provide deliberate intelligence; tools and embodiment provide action; continual and meta-learning provide adaptation; and meta-cognition, evaluation, safety, and alignment provide control. The research objective is to make these capabilities reinforce one another until they form a persistent system able to perceive, reason, learn, plan, and act reliably across unfamiliar domains.