3 days Ago
Dr. Son Nguyen is the cofounder & CEO of Neurond AI , a company providing world-class artificial intelligence and data science services. Artificial intelligence technologies are changing faster than ever.
What started as tools that could answer questions or generate images is now developing into systems that plan, adapt and collaborate like never before. Established giants like OpenAI and Google now share the stage with agile newcomers such as China’s DeepSeek—a rising star making waves with fresh approaches to how AI solves problems, proving that groundbreaking ideas can come from anywhere. In 2025, we’re witnessing more than just smarter algorithms; we’re seeing the arrival of AI systems that learn while they work, autonomous agents that strategize like seasoned professionals and models small enough to fit in your pocket yet powerful enough to compete against supercomputers.
These advancements aren’t isolated—they’re converging to create AI that remembers, modifies and collaborates with human-like continuity. Based on my research experience and running an AI company, let’s explore the most important trends poised to reshape AI technologies in 2025. An AI agent is a system or program that performs tasks autonomously on behalf of a user or another system using artificial intelligence techniques.
Unlike conventional AI, which waits for instructions and simply responds, agentic AI actively figures out what needs to be done and takes action to achieve specific goals. It uses advanced tools like ML (to learn from data), NLP (to understand and use language) and reasoning (to make decisions). This proactive approach allows it to adapt to new situations, learn from what it does and handle complex tasks.
In terms of business operation, an AI agent can autonomously interpret customer requests through NLP, retrieve relevant information and provide personalized responses—all without human intervention. It might even escalate complex issues to human representatives only when necessary, improving efficiency and reducing response times. The increase in interest in AI agents is also reflected in broader technological and market trends in early 2025.
Gartner forecasts that agentic AI will be integrated into 33% of enterprise software applications in 2028, compared to less than 1% in 2024. Inference time compute refers to the computational resources and time required to run a machine learning model to make predictions or inferences on new data. This means allowing the model to spend extra milliseconds (or minutes) “thinking” during real-world use to improve its predictions without requiring retraining.
Grok 3 's use of chain-of-thought prompting exemplifies how inference time computing can be harnessed effectively. This technique encourages the model to generate intermediate reasoning steps, much like a human would when solving a problem. For instance, when tasked with a complex math problem, it doesn’t jump straight to the answer—it breaks the problem into logical chunks, evaluates each step transparently and arrives at a solution that’s not only accurate but also explainable.
Crucially, inference can be tuned and improved without retraining the underlying model. By prioritizing high-quality training data and enhanced inference-time “thought training,” we can create significantly smarter AI agents. AI developers don’t always publicly disclose precise figures about the parameters of their large language models.
However, it’s believed that the current generation of LLMs contains one to two trillion parameters (e.g., 1.
8 trillion for GPT 4 ). The next generation is expected to reach even more parameters . This significant jump promises to unlock even more advanced capabilities, including enhanced reasoning, improved contextual understanding and more fluent and nuanced language generation.
Still, bigger isn’t always better—unless you’re training AI to navigate the complexity of human language. While today’s trillion-parameter models demonstrate impressive language processing abilities, such as writing emails or summarizing text, their 2025 successors aim to better understand language and context. A very large language model could, for instance, parse a legal document while referencing regional laws, historical court cases and even cultural biases in legal language.
Not every AI needs to be a supercomputer. While LLMs have gained much attention, the rise of small language models is equally transformative this year. Smaller models, some with just 3 billion parameters, can punch above their weight.
They’re able to achieve comparable performance to their larger counterparts while demanding fewer computational resources. This portability enables SLMs to run on personal devices like laptops and smartphones, democratizing access to powerful AI capabilities, reducing inference times and lowering operational costs. Microsoft’s Phi-3 represents the most powerful and efficient SLM.
This 3.8B smartphone-friendly model handles coding and math problems efficiently. The secret? Better training data.
Leveraging high-quality textbooks, code repositories and synthetic exercises, researchers are distilling expertise into compact systems. Forgetful AI is becoming a challenge. Most of the current generative AI and LLMs have the “memory problem.
” They may struggle to recall anything beyond the last few messages and can only effectively parse the most recent prompt, limiting their ability to maintain context in lengthy conversations. The development of near-infinite memory is set to revolutionize this issue. 2025’s systems can maintain ongoing conversations and recall all previous interactions over months or years.
Google Gemini , leveraging this advanced memory capability, can provide highly personalized and context-aware responses by drawing on a user’s entire interaction history. This feature allows it to seamlessly pick up where past conversations left off, adapt to evolving preferences and deliver tailored insights without requiring users to repeat themselves. AI in 2025 isn’t just a tool—it’s a teammate, a strategist rolled into one.
Agentic AI will handle tasks that once required entire teams. Models with tens of trillions of parameters will decode problems we thought were too messy for machines. Yet the most transformative advances might come from the smallest models: SLMs are already putting AI in places it’s never been.
But the real story of 2025 isn’t just about size, speed or even intelligence; it’s about how AI fits into our workflows, devices and the mess of our lives. These trends will blur the lines between human and machine intelligence, unlocking productivity and creativity. Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives.
Do I qualify?.
