Cracking The Code Of Consciousness

Forbes

24-07-2025

Rohan Pinto is CTO/Founder of 1Kosmos BlockID and a strong technologist with a strategic vision to lead technology-based growth initiatives.
While contemporary AI achieves impressive feats, crafting human-like prose, recognizing complex patterns and mastering games, it obscures a fundamental limitation. Predominantly built on deep learning, today's systems excel at statistical pattern matching across vast datasets. Yet they struggle with genuine causal reasoning, novel situations and creative adaptation. They manipulate symbols without true understanding. The central challenge, then, is not better data processing but building machines capable of genuine thought. This demands a paradigm shift from pattern recognition to cognitive architectures that embody core principles of understanding and reasoning.
The Thinking Mind Vs. The Processing Powerhouse:
Correlation is the lifeblood of modern AI. A large language model (LLM) can learn the statistical likelihoods of word sequences by being fed millions of sentences. Although it does a fantastic job of predicting the next word, it does not have a solid model of the world those words depict. It doesn't need to understand the fundamentals of quantum physics to create a compelling article about it. It doesn't understand; it processes.
True thinking involves:
Internal Representation & Simulation: Constructing and working with intricate, abstract mental models of the world, its people, things and interactions.
Causal Reasoning: Knowing not just that A and B occur together, but also how A causes B enables intervention planning and prediction in new situations.
Abstraction & Transfer: Recognizing fundamental ideas from certain experiences and adaptably applying them to completely different fields.
Meta-Cognition: Knowing what you know and don't know is the capacity to evaluate one's own information, beliefs and thought processes.
Goal-Directed Problem Solving: Pursuing complicated goals by flexible planning, alternative evaluation and strategy adaptation based on logic rather than merely acquired patterns.
Architectures For Thought: Going Beyond Neural Networks
Achieving this requires architectural innovation, combining concepts from cognitive science, neuroscience and computer science:
• The Symbolic Layer: Structured knowledge representations (logic, ontologies and rules) are used to enable explicit reasoning, relational understanding and abstraction management. Consider this the "rules of the game."
• The Neural Layer: Deep learning offers powerful pattern recognition, perception and learning capabilities. This deals with the "game board's" chaotic sensory data.
• The Integration: The magic is in the bidirectional flow. Neural networks detect symbols in sensory data, like identifying a pixel blob as a "cup." Symbols guide learning and enable reasoning, such as applying rules like "if liquid is hot, handle carefully."
This idea sees the brain as a prediction engine that is continually creating models of the world. It reduces "prediction error" (surprise) by either:
• Updating its model (Learning): "I predicted the cup would be cold, but it's hot. Update my model of this cup/material."
• Acting on the world (Inference): "I predicted the cup is hot, so I'll grasp the handle carefully to confirm and avoid burning."
• Implementation: Building such AI involves hierarchical generative models that predict sensory input across abstractions. Instead of merely reacting, the AI acts to reduce uncertainty and refine its world model driven by intrinsic motivation and curiosity.
Thinking is not disembodied. Human cognition is profoundly influenced by interactions with the physical and social worlds.
• Embodiment: AI agents require sensory-motor loops (even if virtual). Learning physics through interaction (for example, a robot arm handling items) gives concepts like "mass," "friction," and "force" direct experience, resulting in a more solid and intuitive understanding than solely textual learning.
• Situatedness: Reasoning must occur in a given situation. To reason effectively, an AI must grasp the scenario, including relevant entities, relationships, goals and restrictions. This necessitates dynamic context management across its architecture.
Building Blocks For A Thinking Machine
World Models: The core. AI requires internal, simulatable representations of its world, which include objects, actors, attributes, spatial/temporal relationships and causal mechanisms. These models must be compositional (made from pieces) and allow for counterfactual reasoning ("what if?").
Causal Reasoning Engines: Mechanisms for modeling intervention ("If I do X, what happens to Y?") and counterfactuals ("Would Y have happened if I hadn't done X?"). Techniques like causal Bayesian networks or structural causal models, when combined with learning, are critical.
Attention & Resource Management: Thinking necessitates directing computational resources. AI requires systems for dynamic attention, which involves devoting "thinking power" to important components of the world model and current goals, similar to human focus.
Learning to Learn (Meta-Learning): The capacity to improve learning algorithms through experience allows for faster adaption to new tasks and efficient knowledge acquisition.
Uncertainty Quantification & Epistemic Humility: A thinking AI must understand its own knowledge limitations, convey confidence (or lack thereof) in its ideas and forecasts and seek information when uncertain. Bayesian techniques are essential here.
Challenges On The Road To Thought
Scalability & Complexity: Integrating symbolic reasoning, neural learning, world modeling and causal engines efficiently at scale is computationally difficult.
Grounding Symbols: One of the main challenges is ensuring that abstract symbols used in thought processes stay grounded in sensory reality.
Defining & Measuring "Thought": How can we really tell if an AI is actually thinking and not just acting like it through clever processing? We need to go beyond checking if it gets tasks right and start testing how deeply it reasons, how flexible it is and how well it can explain itself.
Architectural Unification: A unified architecture that smoothly brings together all essential components has yet to be established.
In conclusion, building AI that thinks isn't about replicating human consciousness, but about engineering systems with human-like understanding and flexible reasoning. This means moving beyond monolithic pattern matchers to structured, integrative architectures. Neuro-symbolic methods offer explicit, data-grounded reasoning, while predictive processing enables proactive modeling and goal-driven behavior. Embodiment helps root abstract concepts in experience. Despite ongoing challenges, the convergence of these ideas points to a future where AI doesn't just process the world, but begins to understand and reason about it in fundamentally new ways. The goal isn't artificial humans, but artificial thinkers that solve problems through true comprehension.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?