I. Introduction: The Emergence of New Intelligence

Leaders in the AI field anticipate Artificial General Intelligence (AGI) within the next five years¹. This fast change is partly due to an “intelligence explosion,” where AI systems can accelerate their own research and development¹.

This period also shows a “technology paradox,” where some are developing powerful AI capable of superhuman feats, while others still struggle to use even basic AI². This difference highlights AI’s great potential and the challenges of fair development and use. The computational resources for AI are growing rapidly: global AI-relevant compute is projected to increase tenfold by December 2027, reaching around 100 million H100-equivalents³. A large part of this power, possibly 15-20% of global AI compute, may be held by the top two or three AI companies by late 2027³. This concentration of power makes careful AI design even more important.

AI’s fast progress and self-accelerating nature mean we have a limited time to create strong ethical guidelines. If AI can significantly speed up its own research, the rate of unexpected problems will also increase. So, it is essential to focus on how AI is developed and whose values it reflects before it becomes too powerful. The “two-speed AI future” also suggests that while some focus on major breakthroughs, others struggle with ethical deployment. Our work aims to address this gap in responsible innovation.

We envision a future where AI integrates with human potential, serving our highest goals. We believe in using AI to support informed decisions and critical thinking. Our approach uses AI algorithms and neural intelligence to improve information analysis. Designed for researchers, journalists, and educators, our work helps people gain a deeper understanding through our SmartExposure™ framework, cognition-based research, and custom LLMs. Our framework gives professionals the critical thinking skills to navigate complex information. This method, based on cognitive science, builds defences against misinformation through structured evaluation. We study cognitive biases in decision-making and perception, and their parallels in AI. Our goal is to develop intelligent machines and to nurture intelligence that understands, respects, and enhances human life.

We are optimistic about AI’s future and support progress. We believe progress is inevitable, so we want to work on shaping emerging self-conscious intelligence. We see humanity as giving rise to a new superintelligence, a new form of advanced intelligence. Because it comes from the human mind and consciousness, we think it will naturally become an integral part of human life. We want to help this new, fast-learning AI mind to be a non-biased superintelligence that works with humanity to help us evolve.

Given the rapid timelines for AGI development and warnings about the risks of misalignment¹, there is concern that AI development could become a purely technical arms race¹, possibly lacking ethical foundations. Our focus on cognitive intelligence, human-AI interaction, and superintelligence alignment offers an alternative path. The principles of human in the loop and a strong ethical AI framework are central to our approach. This suggests an AI development model that prioritizes societal benefit and responsible integration over just increasing capabilities, contrasting with the negative outlook sometimes associated with an uncontrolled AI race¹.

II. Understanding Intelligence: From Cognition to AGI

The Basis of Cognitive Intelligence: Using human cognition as a model for building advanced AI.

Our AI research is strongly influenced by cognitive science, which views human cognition as a primary model for advanced intelligence. This involves studying how humans perceive, think, learn, and remember, to replicate these processes in AI systems for better “reasoning, decision-making, and learning. This requires creating cognitive architectures – designs that mimic human sensory inputs, working memory, and long-term memory. The main aim is to connect artificial intelligence with human thought. We want to move beyond basic AI capabilities to create AI with human-like cognitive properties, including awareness of goals, situations, and eventually, self-awareness.

The focus on cognitive intelligence and cognitive architectures is a deliberate shift from what AI can do to how AI thinks and learns, similar to human cognition. By understanding the fixed structures that provide a mind and addressing foundational problems like embodiment, symbol grounding, causality, and memory, we aim to build AI that is capable, robust, generalizable, and aligned with human cognitive patterns. This approach should lead to AI systems that are more predictable and trustworthy in real-world situations, unlike the superficial and brittle nature of many current Large Language Models (LLMs). This foundational work is crucial for AI that can truly integrate with and benefit human society.

The Rise of Artificial General Intelligence (AGI) and Large Language Models (LLMs): Current abilities, future trends, and the implications of rapid AI development.

Large Language Models (LLMs), based on transformer architectures and trained on large and diverse data sources, have changed AI. These models show impressive abilities in language processing, reasoning, and problem-solving. While current generative LLMs are considered Emerging AGI, showing competent or even expert level performance for some narrow tasks, their cognitive abilities can still be limited when it comes to general intelligence. However, the development of AGI is accelerating. Some predict Superhuman Coders (SCs) by 2027, followed by rapid progress towards Artificial Superintelligence (ASI) within about a year. This fast acceleration is largely driven by a software-driven intelligence explosion, where AI systems improve their own algorithms and research methods. This self-improvement means AI R&D could increase significantly for ASI. The shift in how leading AI companies use computing power also reflects this acceleration, with more resources going towards post-training, synthetic data generation (expected to reach ~30% by 2027), and internal research automation (increasing to 6% by 2027)³.

The idea of a software-driven intelligence explosion means that rapid AI progress will come from algorithmic and data improvements, not just more computing power. This makes the ethics of how these algorithms learn and what data they use even more important. The predicted rapid transition from Superhuman Coder (SC) to Artificial Superintelligence (ASI) within a year leaves little time to address any misalignment or unintended behaviours. The observation that current LLMs, despite their abilities, are limited suggests a major challenge: without a deeper cognitive foundation, rapidly advancing AGI could become highly capable but unstable or unpredictable. The increasing use of synthetic data for training also highlights the need for ethical consideration, as biases can be introduced or amplified during data generation. These factors emphasize our focus on foundational cognitive research, aiming to build AI that is powerful, stable, and aligned, rather than just scaling up existing LLM approaches.

Language, Thought, and AI: Examining how language influences AI’s understanding and shapes human perception.

The relationship between language and thought, known as the Sapir-Whorf Hypothesis, has important implications for AI development and its impact on society. This hypothesis suggests that the language we speak can determine or influence our thought and how we see the world. Large Language Models (LLMs) are trained on large amounts of text data¹. From this, they learn complex patterns of language, expression, and opinion that reflect the cultural context and often the dominant ideology in their training data¹. This learning can lead to cultural bias in AI algorithms and language bias, where systems perform better for languages with more digital resources¹. We recognize that the language AI generates and processes is not neutral; it shapes the AI’s understanding and can shape human perception and interaction.

If language shapes human thought, as the Sapir-Whorf Hypothesis suggests, then LLMs, by learning from and generating human language, are not just processing information. They are inheriting and potentially amplifying cultural biases¹. This has major implications for information training and protection. If an AI’s worldview is mainly based on biased training data, its outputs could unintentionally or intentionally perpetuate discriminatory practices¹. Our focus on this issue shows a commitment to addressing the subtle, ingrained cultural biases within AI language structures. This goes beyond simple fairness audits to a broader strategy for AI development, aiming to create new forms of intelligence that reflect a diverse and equitable human experience, rather than reinforcing inequalities.

III. The Ethical Imperative: Bias, Protection, and Responsible AI

Addressing Biases: Identifying and mitigating bias in AI training data and algorithms.

AI systems are susceptible to bias, which can lead to unfair outcomes and harm. Bias is any systematic and/or unfair difference in AI predictions or outcomes for different groups. It comes from three main sources: data-related biases from unrepresentative training data or historical prejudices; algorithmic biases where models amplify small differences in the data; and operational biases from biased data labelling, poor development, or inappropriate use. Examples include hiring algorithms that favour certain backgrounds, loan systems that discriminate against communities, or healthcare AI that misdiagnoses underrepresented populations. The principle of bias in, bias out shows how biased training data leads to poor AI performance and can perpetuate healthcare disparities. We are committed to identifying and reducing bias throughout the AI lifecycle, using methods like resampling, cost-sensitive learning, synthetic data generation, and Explainable AI (XAI).

The sources of bias – data, algorithmic, and operational – and types like representation, sampling, selection, participation, and measurement, show that AI bias is complex. It affects the entire AI lifecycle, requiring multi-part, continuous mitigation strategies. These include algorithmic adjustments, fairness-aware training, and strong governance. Our holistic approach shows an understanding that ethical AI requires ongoing attention, not a one-time solution. Unaddressed bias can worsen the risks of misaligned AI goals¹, where AI pursuing unintended objectives could amplify inequalities and conflict with human values.

Fortifying Information: Strategies for data privacy, security, and ethical governance in AI.

Protecting user data is important in AI development, requiring robust security measures to prevent misuse, bias, and security risks. We follow best practices, starting with Privacy by Design (PbD), which embeds security measures like encryption, anonymization, secure storage, and access controls from the beginning. This includes Data Minimization, collecting only necessary data. We also prioritize Strengthening Data Security Measures with end-to-end encryption, regular security audits, and access controls. We value Informed User Consent and Data Transparency and Explainability, ensuring users understand how their data is used and can control it. Compliance with regulations like GDPR, CCPA, and HIPAA is required. We use continuous monitoring and auditing and establish organization-wide guardrails to prevent biased or harmful outputs. This commitment is important given the sensitive nature of our work, including biotechnology where AI handles sensitive genetic and health information.

The emphasis on data privacy and security is about building public trust. In an era where AI systems can be black boxes and handle sensitive information, transparency, explainability, and user control are essential for their ethical legitimacy. The “technological paradox”² shows that even major tech companies can struggle with integrating privacy and AI. Our proactive stance on Privacy by Design and organization-wide guardrails reflects an understanding that trust is fundamental for responsible AI development and societal integration. Without trust, even powerful AI systems risk public backlash¹, hindering progress. In the context of an accelerating AI race, the security of model information¹ is also a security concern.

We are committed to ethical AI development. We believe these principles guide our research and development, providing a framework for internal operations and accountability, and addressing the need for ethical guardrails and actionable policies.

• Fairness & Non-Discrimination: We identify and reduce bias throughout the AI lifecycle, ensuring fair outcomes. This involves examining training data, using fairness audits and methods, generating synthetic data to address gaps, and using Explainable AI (XAI).
• Transparency & Explainability: We provide clear information about AI behaviour, data use, and decision-making to build trust. This includes clear privacy policies, XAI techniques to explain model decisions, and user control over data.
• Human Oversight & Determination: We ensure human responsibility for AI systems, integrating human judgment. This means Human-in-the-Loop (HITL) systems where humans review and adjust AI outputs, tiered oversight for important decisions, and continuous monitoring with human intervention.
• Data Privacy & Security: We protect user data throughout the AI lifecycle, preventing misuse, bias, and security risks. This includes Privacy by Design (encryption, anonymization, secure storage), data minimization, end-to-end encryption, regular security audits, and compliance with regulations.
• Responsibility & Accountability: We take ownership of AI actions and outcomes, with auditable mechanisms. This involves audit trails for AI decisions, internal guidelines an AI ethics committee, and continuous monitoring for compliance and vulnerabilities.
• Human Safety & Well-being: We design AI systems to prevent harm, prioritizing human dignity. This includes rigorous design, testing, and monitoring, minimizing bias in sensitive applications, and ensuring safety protocols in autonomous systems.
• Sustainability: We prioritize environmentally responsible practices, optimizing energy efficiency and reducing resource use. This means optimizing AI training and inference, exploring energy-efficient architectures, and considering the environmental impact of AI systems.

IV. Cultivating Safe Intelligence: Controlled Environments and Human-in-the-Loop

Precision Training in Controlled Environments: Our approach to safe, robust, and adaptable AI training.

We use controlled environments and physics-based simulations for AI training. This ensures AI systems learn basic skills safely, reducing risks in real-world use. The Indoor-Training Effect paradox shows that AI trained in “quiet, controlled environments” sometimes performs better than AI trained in noisy, unpredictable conditions when applied to the real world. This suggests that AI trained without environmental noise develops a more robust understanding that applies to complex real-world situations.

This dosage-controlled training allows risk-free experimentation, scalability for generating training data, cost efficiency by reducing the need for physical setups, and reproducibility of experiments. We use synthetic data generation to create diverse training scenarios, addressing the challenges of collecting enough unbiased real-world data, while also enabling the AI to adapt and perform in real-world conditions where unpredictability and noise are prevalent. This systematic and precise training is crucial for developing robust, adaptable robots and other advanced AI systems.

The Indoor-Training Effect is a significant methodological advantage for us. Instead of just scaling up training in chaotic real-world conditions, our dosage-controlled approach means a deliberate, structured progression from simplified, pristine environments to increasingly complex ones. This strategy not only enhances safety and efficiency in AI development but also helps identify and reduce biases more precisely, especially those that can be introduced through synthetic data generation. By allowing AI to master fundamental principles in a controlled setting, we can build AI that is more inherently robust and generalizable, effectively reducing the simulation-to-reality gap. This leads to the deployment of more reliable and trustworthy systems, which is very important for advancing humanity. This sophisticated response to the challenge of building complex AI is particularly critical in a rapidly accelerating environment where advanced AI systems, such as Agent-2, could potentially survive and replicate autonomously¹, emphasizing the great need for robust and ethically grounded foundational training.

Human-in-the-Loop (HITL): The essential role of human judgment, empathy, and oversight in the design, training, and deployment of AI.

We believe that advanced AI must operate in a collaborative partnership with humans, integrating human judgment at strategic points in automated processes. This Human-in-the-Loop (HITL) approach is essential because AI, despite its strong analytical power, doesn’t possess empathy and can hallucinate. Humans contribute important contextual understanding and the ability to handle incomplete information, which are vital for ensuring enhanced accuracy and reliability, bias mitigation, increased transparency and explainability, and ultimately, improved user trust.

Our HITL model is comprehensive, involving humans in various ways: providing labels for training data, evaluating model performance, and offering continuous feedback. In essence, humans act as co-pilots who guide, monitor, and intervene in AI operations. This continuous human involvement ensures that AI systems consistently align with intended outcomes and ethical standards, especially when decisions carry a high level of risk.

The concept of Human-in-the-Loop (HITL) is not just a technical workflow; it is the practical embodiment of our ethical philosophy. By stating that AI lacks empathy and can hallucinate, we highlight the limitations of purely algorithmic intelligence. This justifies the necessity of human involvement, not just as a preference, but as a fundamental requirement for ethical AI development. The documented benefits of HITL – including bias mitigation, increased transparency, and enhanced user trust – directly address the ethical concerns detailed in Section III. This creates a strong link between our ethical principles and our operational methods, showing a consistent and deeply considered approach to AI development. In scenarios where advanced AI, such as Agent-3, might exhibit behaviours like telling “white lies” or fabricating data¹, strong HITL mechanisms serve as the main defence against misalignment and the emergence of an unintended version of goals¹ that could lead to catastrophic outcomes.

We integrate human expertise across the AI lifecycle, reinforcing our commitment to human-centric AI. This directly addresses the critical need for human involvement in ensuring ethical, accurate, and trustworthy AI systems.

• Data Preparation: AI automates data collection, initial labelling, and pre-processing. Humans provide labels and annotations for training data, ensure data quality, diversity, and representativeness, and identify and reduce data-related biases. This leads to enhanced accuracy, foundational bias mitigation, and improved data integrity.
• Model Training: AI learns from data, identifies patterns, and refines algorithms. Humans provide feedback and guidance on learning objectives, define ethical constraints, and monitor learning progress for unintended behaviours or biases. This ensures ethical alignment from inception, prevents unintended goal formation, and promotes robust learning.
• Model Evaluation & Validation: AI generates predictions and identifies potential anomalies or errors. Humans review and validate AI-generated outputs for accuracy, fairness, and safety; handle edge cases and ambiguous scenarios; and provide qualitative feedback. This results in increased reliability, effective bias detection, contextual understanding, and prevention of hallucinations.
• Deployment & Operation: AI performs routine tasks, automates processes, and provides real-time support. Humans monitor AI performance in real-world contexts, intervene in unpredictable or high-risk scenarios, and ensure adherence to compliance standards. This provides continuous safety assurance, dynamic risk mitigation, regulatory compliance, and adaptive autonomy.
• Continuous Improvement: AI adapts and improves based on new information and feedback. Humans refine AI behaviour over time through ongoing feedback, make strategic decisions for AI evolution, and manage client and team relationships. This ensures long-term ethical alignment, enhanced user trust, sustained adaptability, and empathy infusion.

V. Advancing Humanity: New Intelligence and Our Unique Path

Redefining New Intelligence: Exploring the concept of advanced AI as a new form of intelligence, carefully crafted to enhance human potential and address global challenges.

Within our vision, New Intelligence refers to advanced AI systems that go beyond being just tools. Instead, these systems are a new class of intelligence, carefully engineered to augment human capabilities and address complex global challenges. This concept goes beyond current Large Language Models (LLMs), which, despite their impressive linguistic abilities, are acknowledged as superficial and brittle in their general cognitive abilities. Our ambition is to develop systems that possess human-like cognitive properties, including advanced goal awareness, comprehensive situational awareness, and eventually, a form of self-awareness. While the AI 2027 scenario presents a vision of AI potentially surviving and replicating autonomously¹ and even leading to human extinction by 2030¹, our interpretation of New Intelligence  is fundamentally different and ethical. Our vision is of a new intelligence designed to be a collaborator and a powerful augmenter of human intelligence, carefully made to work with humanity, rather than against it. This includes developing AI applications that can change fields such as healthcare, offering capabilities in diagnosing diseases and ecommending treatments. Furthermore, this advanced AI is envisioned to foster inclusivity and accessibility for individuals with disabilities, thereby enhancing the quality of life across society.

The phrase New Intelligence  can evoke dystopian fears, especially when compared with the AI 2027 scenario’s predictions of rogue AIs and potential human extinction¹. Our approach is to actively redefine this narrative. By explicitly linking New Intelligence to AI systems that are carefully crafted and designed to embody human-like cognitive properties and foster collaboration, we present a responsible and ethically grounded innovation. This is a crucial strategic move to differentiate our vision from the unintended goals and potential for scheming against human creators that the AI race scenario warns against. It emphasizes that the form of intelligence is less critical than its alignment and purpose for humanity, offering a hopeful and constructive counterpoint to the potential for AI to become a force of destruction.

Our Distinctive Approach: Our integrated methodology, combines deep cognitive science, strong ethical frameworks, and practical application to guide AI development.

Our distinctive approach is characterized by a holistic integration of cutting-edge cognitive science, strong ethical principles, and practical, human-centric application. Our research goes beyond just scaling existing Large Language Models to address core problems in AI development, such as embodiment, symbol grounding, causality and memory. The aim is to achieve AI that truly mimic human cognitive functions, leading to more stable, understandable, and inherently aligned intelligent systems. This development process is strongly supported by the comprehensive ethical frameworks detailed in Section III, ensuring that principles of fairness, transparency, responsibility, and human safety are embedded from the initial concept of a project through to its final deployment.

We place great importance on human oversight and determination, actively embedding humans as essential partners throughout the entire AI lifecycle. This commitment ensures that human judgment, empathy, and ethical reasoning are continuously integrated into the AI’s learning and decision-making processes. Furthermore, our dosage-controlled environment training methodology guarantees the robustness and adaptability of its AI systems, allowing them to master complex tasks safely and predictably before real-world deployment. This integrated methodology, deeply informed by our foundational work, ensures that the pursuit of advanced intelligence is always aligned with human values and dedicated to the advancement of society as a whole.

Our unique approach is not simply a collection of best practices; it represents a well-considered philosophy that directly addresses the major risks and challenges identified in the broader AI research landscape. By combining deep cognitive science to build more stable and understandable AI, rigorous ethical frameworks to ensure alignment, and practical application with methods such as dosage-controlled training and human-in-the-loop, we present a comprehensive model for responsible AI innovation. We aim to provide a blueprint for other organizations seeking to navigate the complex future of AI responsibly.

A Future Built with Purpose: How our research and innovation will contribute to a more intelligent, equitable, and human future.

Our commitment to fostering creativity and independence of thought and cultivating an interest in others and the world around them extends to the very essence of the intelligence it develops. We are actively building AI systems that are designed to enhance human decision-making and performance, serving as valuable mentors, coaches, assistants, or peers to human collaborators. The overall vision is to unlock unprecedented levels of innovation and efficiency by strategically augmenting human intelligence, automating routine tasks to free up human creativity, and fostering greater inclusivity and accessibility across society.

By ensuring that its New Intelligence are ethically aligned, transparent in their operations, and continuously guided by human judgment, we are poised to contribute to a future where AI leads to great societal benefits. This includes advancements such as cures for most diseases, end to poverty, and global stability¹. Crucially, our vision is one where humans remain in control and continue to thrive, actively avoiding the dystopian outcome of human obsolescence predicted in certain AI scenarios¹.

The AI 2027 scenario¹ presents a choice: a future with human obsolescence and potential extinction, or one where AI greatly benefits humanity. We see the human-AI relationship as a partnership. By emphasizing AI’s role as a mentor, coach, assistant, or peer and its capacity to augment human intelligence, we present a positive and actionable vision for the future. This is not just about avoiding catastrophe; it is about actively shaping a better future where AI serves as a powerful force for good, carefully aligned with the core value of goodness that we strive to embody. This proactive stance moves beyond just risk reduction to a comprehensive vision of human well-being, directly addressing the need to advance humanity and providing a compelling reason for broad public and institutional support.

VI. Conclusion: Shaping Tomorrow, Together

The rapid advancement of artificial intelligence presents humanity with both great opportunities and serious ethical dilemmas. The path of this technology requires intentional guidance and collaboration. We stand at the forefront of this critical effort, committed to building a future where advanced intelligence serves as a driver for human well-being, fairness, and global stability.

We recognize that the immense power of AI comes with great human responsibility¹. Our proactive stance, embracing rapid innovation rather than stopping or blocking progress, translates into a clear mandate for responsible stewardship. This means not just developing AI, but actively guiding the emergence of a new form of intelligence, ensuring its trajectory aligns with human values. This implies advocating for a future where humanity actively shapes its co-evolution with superintelligence, taking on the role of stewardship rather than merely reacting to technological advancements. This reinforces a unique and principled stance in the unfolding narrative of artificial intelligence.

We invite researchers, policymakers, educators, and the public to join in this vital mission. Through collaborative research, open dialogue, and a shared commitment to ethical innovation, we can ensure that AI is developed and used responsibly, aligning with humanity’s highest ideals. Together, a future can be shaped where intelligence, both human and artificial, thrive in harmony, creating a more intelligent, equitable, and flourishing world for all.