Agentic AI refers to intelligent systems capable of reasoning, planning, and taking action to achieve specific goals. They can break down complex problems into multi-step plans and interact with tools and databases. While current models struggle with consistent logical reasoning in complex scenarios, the development of more robust models in 2025 is expected to lead to well-performing AI agents that can effectively tackle intricate tasks. This trend is driven by a clear demand for AI that can go beyond simple execution to handle more nuanced problems.

Inference time compute is a development where new AI models extend the processing time during inference, essentially allowing them to "think" before generating an answer. The duration of this "thinking" is variable, depending on the complexity of the reasoning required. What makes this significant is that this reasoning can be tuned and improved without needing to retrain the underlying model. This means that AI's reasoning capabilities can be enhanced not only during training (with better data) but also during inference (with improved chain-of-thought processing), ultimately contributing to smarter AI agents.

In 2025, AI models are expected to span a wide range of sizes. Frontier models are anticipated to grow significantly larger, potentially reaching upwards of 50 trillion parameters, far exceeding the 1-2 trillion parameters seen in 2024. Conversely, 2025 is also projected to be the year of "very small models," with only a few billion parameters. These smaller models are notable because they can run on personal devices like laptops and phones without requiring extensive computing resources or large data centers, making AI more accessible and capable of performing specific tasks efficiently on local devices.

In 2024, common enterprise AI use cases focused on improving customer experience, IT operations and automation, virtual assistants, and cybersecurity. Looking ahead to 2025, AI is expected to enable more advanced use cases. This includes customer service bots capable of solving complex problems (beyond just routing tickets), AI systems that can proactively optimize entire IT networks, and security tools that can adapt to new threats in real-time. The shift indicates a move from basic support functions to more sophisticated, problem-solving, and adaptive AI applications within enterprises.

"Near-infinite memory" refers to the growing ability of AI models, particularly chatbots, to retain a vast amount of past information and conversations with users. While earlier models had limited context windows (e.g., 2,000 tokens), current models can handle hundreds of thousands or even millions of tokens. This trend suggests that in the near future, customer service chatbots, for example, will be able to recall every conversation they've ever had with a user. This could lead to a more personalized and seamless user experience, as AI systems will have a comprehensive understanding of past interactions and preferences.

Human-in-the-loop augmentation involves combining human expertise with AI systems to achieve better outcomes than either could achieve alone. An example cited is a study where a chatbot outperformed physicians in clinical reasoning, but doctors using the chatbot performed worse than the chatbot alone. This highlights a current challenge: effectively integrating AI tools into professional workflows without requiring the human professional to be an expert in AI usage or prompt engineering. The goal for 2025 is to develop better systems that allow professionals to seamlessly leverage AI's capabilities to enhance their own performance.

The reasoning capabilities of Large Language Models (LLMs) can be improved in two primary ways. Firstly, at the training stage, by using better quality training data to refine the model's parameters. Secondly, and increasingly significant for 2025, is through "inference time compute." This involves extending the time an LLM spends "thinking" before providing an answer, and crucially, improving this inference-time reasoning through techniques like "chain of thought training." This allows for enhancements to reasoning without the need for full model retraining.

The source actively solicits the audience's input for identifying important AI trends. In a previous "2024 Trends" video, the speaker turned the final trend over to the viewers, and received hundreds of insightful thoughts. This approach is being replicated for 2025, suggesting that the collective wisdom and observations of the AI community and users are considered a valuable source for identifying emerging and significant trends in the field.