Yulian Bed, Roman Chiypesh, Vadym Shatan

Research Institute for Artificial Intelligence Studies
Carpathian University named after Augustin Voloshyn

Uzhhorod, 25 September 2025.

Abstract

The article presents a review and analytical assessment of the latest advances in artificial intelligence (AI) between 2023 and 2025. It examines key lines of progress—multimodal and agentic models, neuro-symbolic approaches, embodied AI and robotics, autonomous algorithm discovery, and applications in healthcare, education, and industry. The study highlights architectural and infrastructural innovations and identifies major challenges and risks associated with AI development. It emphasizes the prospects for integrating neural and symbolic paradigms, the growing role of agentic systems, and the importance of regulation.

Keywords: artificial intelligence; multimodal models; agentic AI; neuro-symbolic AI; embodied AI; autonomous algorithm discovery; healthcare; education; innovation.

Introduction

In 2023–2025, artificial intelligence (AI) firmly established itself as a principal technological driver of our time, shaping not only the IT sector but also healthcare, education, industry, and the broader economy. According to Stanford University’s AI Index Report 2025, the compute required to train the most powerful models doubles roughly every five months—far outpacing Moore’s-law-style trends [1]. While this fosters resource concentration among a few market leaders, the gap is gradually narrowing thanks to open research and innovations in architectures and optimization.

Accordingly, AI progress in 2023–2025 reflects a shift from a pure “scaling race”—ever-larger parameter counts and datasets—toward competition in architectural and applied innovation. This includes new approaches to multimodal systems and agentic models, integration of neural and symbolic methods, and hardware optimization—now a decisive factor for future advances.

Multimodal and Agentic Models

A defining trend of 2023–2025 has been the rapid development of multimodal models capable of simultaneously processing text, images, audio, and video. This progress reflects not only technical advances but also a conceptual rethinking of artificial intelligence—from specialized systems toward universal “foundation models.”

• GPT-4o (OpenAI, 2024) marked a milestone by introducing a fully integrated multimodal architecture. Unlike earlier systems that stitched together separate models, GPT-4o is able to process and generate speech, images, and video natively, demonstrating real-time interaction and lowering inference costs.
• Gemini (Google DeepMind, 2023–2025) continues to evolve as a series of multimodal agents, emphasizing reasoning, planning, and integration with external tools.
• Claude (Anthropic) and Mistral models introduced new levels of interpretability and efficiency, while open-source ecosystems (LLaMA, Mixtral, Qwen, etc.) significantly expanded access to cutting-edge multimodality.

At the same time, a second major breakthrough of this period was the advancement of agentic AI—systems capable of autonomous action, reasoning, and decision-making in complex environments.

• The rise of AutoGPT, BabyAGI, and subsequent frameworks demonstrated the potential for chaining reasoning steps and delegating tasks to AI agents.
• By 2025, most major laboratories and corporations had shifted from producing static models toward agent-based ecosystems, where multiple AI systems cooperate, specialize, and interact with human users.

Agentic AI is reshaping how researchers and businesses imagine the next phase of progress: not merely smarter models, but AI societies of cooperating agents that can plan, execute, and evaluate tasks across digital and physical domains.

Neuro-Symbolic AI

Another major direction of AI development in 2023–2025 has been the integration of neural networks with symbolic reasoning systems. This trend aims to overcome the limitations of purely statistical approaches, which often struggle with logic, abstraction, and explicit knowledge representation.

• Research at MIT, Stanford, and DeepMind showed that combining large language models (LLMs) with symbolic logic modules significantly improves consistency in reasoning, mathematical proofs, and scientific discovery.
• Neuro-symbolic architectures allow AI to not only generate coherent text but also verify hypotheses, execute formal operations, and provide explanations of its reasoning—a critical step toward transparent and trustworthy AI.
• Such models are increasingly applied in law, medicine, and education, where high standards of accuracy and accountability are required.

The neuro-symbolic approach is also viewed as a pathway to reducing the risks of hallucinations in generative AI. By grounding neural outputs in symbolic frameworks, researchers are moving toward systems that combine creative flexibility with logical rigor.

Thus, 2023–2025 became a period when neuro-symbolic AI transitioned from theoretical research to practical applications, gradually forming a new paradigm that may define the trajectory of artificial intelligence in the coming decade.

Embodied AI and Robotics

In 2023–2025, the field of embodied AI—artificial intelligence integrated with physical agents and robotic systems—achieved a series of important breakthroughs. Unlike disembodied models that operate purely in digital space, embodied AI directly interacts with the physical environment, expanding both its challenges and opportunities.

• Boston Dynamics, Tesla, Figure AI, and Agility Robotics demonstrated humanoid robots capable of performing increasingly complex manipulations, navigation, and coordination tasks. Their progress reflects not only mechanical engineering achievements but also the integration of advanced AI models for real-time decision-making.
• The development of sim-to-real transfer technologies allowed agents trained in virtual environments to adapt effectively to real-world conditions, significantly accelerating the pace of robotics research.
• Embodied AI systems are now being tested in logistics, manufacturing, agriculture, and healthcare, where they assist with physically demanding, repetitive, or high-risk tasks.

A parallel trend is the emergence of interactive environments (e.g., MineDojo, Habitat, MuJoCo), which serve as large-scale training grounds for embodied agents. These platforms enable AI to acquire generalized skills—navigation, object manipulation, multimodal perception—that can be reused across different domains.

By 2025, embodied AI is no longer viewed as a futuristic vision but as a strategically important area of technological competition, directly tied to the creation of autonomous service robots, industry 4.0 automation, and even defense applications.

Autonomous Algorithm Discovery

One of the most striking breakthroughs of 2023–2025 has been the progress in AI systems capable of discovering new algorithms independently. This development signals a shift from using AI merely as a tool for optimization to positioning it as a co-creator of scientific and technological knowledge.

• DeepMind’s AlphaDev (2023) demonstrated that reinforcement learning could generate novel algorithms for sorting and data processing that outperformed human-designed solutions in efficiency.
• Similar initiatives by OpenAI, Meta AI, and independent research groups showed that large models are able to propose innovative architectures for neural networks, optimization techniques, and coding strategies.
• These results suggest that AI is becoming not only an object of design but also a subject of innovation, capable of contributing to the evolution of computer science itself.

The implications of autonomous algorithm discovery are profound:

1. Acceleration of scientific progress — AI can rapidly test and refine approaches, reducing the time needed for breakthroughs.
2. Optimization of resources — new algorithms frequently surpass traditional methods in speed and energy efficiency, which is particularly critical given the rising cost of training large models.
3. Philosophical and ethical challenges — questions arise about authorship, intellectual property, and responsibility when algorithms are created by machines rather than humans.

By 2025, autonomous algorithm discovery is emerging as a central trend in AI research, blurring the boundary between human creativity and machine innovation.

Applications of AI in Healthcare, Education, and Industry

The years 2023–2025 confirmed the practical significance of artificial intelligence, as its applications became deeply integrated into key sectors of society.

Healthcare

• AI has increasingly become an indispensable tool for diagnostics, treatment planning, and drug development.
• Medical imaging enhanced by AI enables faster and more accurate detection of cancers, cardiovascular diseases, and neurological disorders.
• Generative models are being used to design new molecules and predict protein structures, building on the achievements of AlphaFold and related systems.

AI-driven telemedicine platforms expand access to healthcare in remote and underserved regions, making consultations more effective and affordable.

Education

The education sector has witnessed a rapid transformation thanks to adaptive learning technologies:

• AI-powered platforms personalize curricula, track student progress, and provide tailored feedback.
• Large language models are widely used as tutors and assistants, helping students write, analyze, and solve problems in real time.
• Universities and schools increasingly integrate AI into administrative processes, content creation, and academic research.

Industry

AI is becoming the backbone of the Fourth Industrial Revolution:

• Predictive maintenance systems reduce downtime and increase efficiency in manufacturing.
• Smart logistics and supply chain optimization help mitigate disruptions caused by geopolitical crises and pandemics.
• In finance, AI ensures fraud detection, algorithmic trading, and personalized customer service.

The overall impact of these applications is reflected in rising productivity, reduced costs, and improved quality of services. However, they also raise critical issues—job displacement, privacy concerns, and the need for new regulatory standards.

Architectural and Infrastructural Innovations

The years 2023–2025 have also been marked by a shift in focus from sheer scale expansion (larger datasets and billions of parameters) to architectural optimization and infrastructure improvement. These changes were necessary to overcome the limitations of energy consumption, costs, and latency in AI systems.

• Mixture-of-Experts (MoE) models (e.g., Mixtral, DeepSeek) showed that distributing computations across specialized subnetworks can dramatically reduce costs while maintaining high performance.
• Sparse attention mechanisms and low-rank adaptation methods (LoRA) enabled more efficient fine-tuning of large models without retraining them from scratch.
• The development of quantization and pruning techniques allowed AI models to be deployed on mobile devices and edge systems, democratizing access to advanced AI.

On the infrastructure side:

• New AI supercomputers (Frontier, Aurora, Fugaku, and specialized clusters built by Google, Microsoft, and NVIDIA) provided unprecedented training capacity.
• The race for energy-efficient hardware intensified, with innovations in GPUs, TPUs, and neuromorphic chips.
• Cloud providers began offering AI-as-a-Service platforms that allow small businesses and universities to experiment with state-of-the-art models without massive infrastructure investments.

Thus, architectural and infrastructural advances during this period not only expanded the boundaries of AI capabilities but also ensured their greater accessibility and sustainability.

Challenges and Risks

Despite impressive achievements, the rapid development of AI in 2023–2025 has raised a number of serious challenges and risks that require careful regulation and ethical reflection.

Hallucinations and Reliability

Large language models and multimodal systems still suffer from generating hallucinations—false or misleading information presented as fact. This problem limits the trustworthiness of AI in critical fields such as law, healthcare, and scientific research.

Concentration of Resources

The enormous cost of training frontier models has led to the concentration of AI development in the hands of a few global corporations. This centralization raises concerns about monopolization, unequal access, and geopolitical dependence.

Ethical and Legal Issues

AI challenges established legal and ethical frameworks:

• Intellectual property questions arise over AI-generated texts, images, and algorithms.
• Accountability and responsibility remain unclear when AI systems make autonomous decisions.
• Privacy and surveillance risks grow with the widespread use of AI in biometrics and data analysis.

Labor Market Transformation

Automation and AI-driven productivity gains threaten to displace millions of jobs in manufacturing, logistics, and services. At the same time, new opportunities emerge in AI development, data analysis, and digital infrastructure management.

Security and Geopolitics

AI technologies are increasingly viewed as a matter of national security. States are investing in AI for defense, cyber operations, and intelligence. This creates risks of militarization and escalation of technological competition.

Prospects for Development

The trajectory of AI in 2023–2025 suggests several key directions that will shape the next stage of technological evolution:

1. Integration of Neural and Symbolic Approaches

The convergence of deep learning with symbolic reasoning is expected to reduce hallucinations, improve explainability, and enhance reliability in critical domains.

• Agentic AI Ecosystems

Autonomous multi-agent systems will increasingly cooperate with humans and with each other, forming digital “societies of agents” capable of solving complex, multi-step tasks.

• Embodied Intelligence

Advances in robotics and embodied AI point to a future where autonomous agents will operate not only in virtual but also in physical environments—factories, hospitals, homes, and even outer space.

• Sustainable AI

Research will focus on reducing energy consumption and the environmental footprint of large models. Energy-efficient architectures and neuromorphic hardware are central to this agenda.

• Global Regulation and Standards

The growing influence of AI will require international agreements, legal frameworks, and ethical guidelines. The European Union, the United States, and other regions are already drafting AI Acts and strategies that will define the global rules of the game.

In general, AI is entering a stage where the key issue is no longer whether it can achieve superhuman performance in certain areas, but how to ensure that such progress remains safe, equitable, and beneficial for humanity as a whole.

Conclusions

The period of 2023–2025 has become a decisive stage in the development of artificial intelligence. AI has advanced from incremental improvements to qualitative breakthroughs:

• the emergence of multimodal and agentic systems;
• the practical application of neuro-symbolic models;
• the rise of embodied AI and robotics;
• and the first successes in autonomous algorithm discovery.

At the same time, AI has deeply transformed healthcare, education, industry, and science, proving its status as a foundational technology of the 21st century. Architectural and infrastructural innovations have made progress more efficient and sustainable, though they have not eliminated risks related to centralization, ethical dilemmas, and security concerns.

The prospects for the near future lie in balancing innovation and responsibility: building AI systems that are powerful yet explainable, autonomous yet accountable, and globally accessible while respecting human dignity and rights.

Artificial intelligence is no longer just a technological tool; it has become a structural factor in global development, shaping the economy, politics, culture, and human thought itself. The task of the coming decade is to ensure that this force is directed toward the good of humanity.

References

1. Stanford University. AI Index Report 2025. Stanford Institute for Human-Centered AI. Available at: https://aiindex.stanford.edu (accessed: 25 September 2025).
2. OpenAI. GPT-4o Technical Report. OpenAI, 2024. Available at: https://openai.com/research/gpt-4o (accessed: 25 September 2025).
3. Google DeepMind. Introducing Gemini: A new multimodal AI model. DeepMind Blog, 2023–2025. Available at: https://deepmind.google (accessed: 25 September 2025).
4. Anthropic. Claude AI System Card. Anthropic, 2024. Available at: https://www.anthropic.com
   (accessed: 25 September 2025).
5. Meta AI. LLaMA Models and Research Releases. Meta AI, 2023–2025. Available at: https://ai.meta.com (accessed: 25 September 2025).
6. MIT Computer Science & Artificial Intelligence Laboratory (CSAIL). Neuro-Symbolic AI Research. MIT, 2023–2025. Available at: https://www.csail.mit.edu (accessed: 25 September 2025).
7. DeepMind. AlphaDev Discovers New Algorithms. DeepMind Blog, 2023. Available at: https://deepmind.google/discover/blog/alphadev (accessed: 25 September 2025).
8. Boston Dynamics. Next-Generation Humanoid Robots. Boston Dynamics, 2024. Available at: https://www.bostondynamics.com (accessed: 25 September 2025).
9. Tesla. Optimus Robot Development Update. Tesla, 2023–2025. Available at: https://www.tesla.com/AI (accessed: 25 September 2025).
10. Agility Robotics. Digit: Human-Centric Robotics. Agility Robotics, 2023–2025. Available at: https://agilityrobotics.com (accessed: 25 September 2025).
11. MineDojo. Open-Ended Embodied AI Research Environment. Carnegie Mellon University, 2023. Available at: https://minedojo.org (accessed: 25 September 2025).
12. NVIDIA. AI Infrastructure and Supercomputing Platforms. NVIDIA, 2023–2025. Available at: https://www.nvidia.com (accessed: 25 September 2025).
13. European Union. AI Act: Proposal for a Regulation on Artificial Intelligence. European Commission, 2024. Available at: https://artificial-intelligence-act.eu
    (accessed: 25 September 2025).
14. U.S. Government. National AI Research and Development Strategic Plan (2023 Update). Washington, D.C., 2023. Available at: https://www.whitehouse.gov/ostp/ai
    (accessed: 25 September 2025).