Artificial Intelligence: What to expect in 2025

As we step into 2025, the world of Artificial Intelligence (AI) is poised to undergo seismic shifts. The breakthroughs we’ll witness are not merely technical feats but transformative shifts that will redefine industries, economies, and societies. But what makes a technology trend in AI worth watching? It’s the potential for scale, the promise of solving long-standing challenges, and the ethical questions it raises. Below, I delve into top key AI trends expected to take centre stage in 2025, sorted by their disruptive potential and substantiated by ongoing advancements.

AGI: Yes, it’s coming in 2025

While there has been a lot of debates around what exactly is AGI, the broadest definition is that Artificial General Intelligence (AGI) refers to an advanced form of AI capable of understanding, learning, and applying knowledge across a wide range of tasks at a level comparable to human intelligence. Unlike narrow AI, which excels in specific tasks like language translation or image recognition, AGI can adapt to new problems, reason abstractly, and make autonomous decisions without being explicitly programmed for each scenario. AGI represents the ultimate vision of AI development, where machines achieve human-like cognitive capabilities and versatility.

The current AI systems are beyond narrow AIs and heading towards AGI. ChatGPT O3 demonstrates remarkable capabilities, such as understanding context, generating coherent and creative responses, performing complex reasoning tasks, and adapting to a wide range of topics. These abilities can make it feel like interacting with an intelligent entity capable of abstract thinking. Its performance on various benchmarks suggests a degree of generalisation and adaptability far beyond earlier models. One of the standout features of O3 is its ability to decompose intricate problems into manageable steps, allowing for more accurate and reliable solutions. This methodical approach enables the model to handle tasks that require multi-step reasoning, such as advanced mathematics, coding challenges, and scientific problem-solving. By simulating a “chain of thought,” ChatGPT O3 can plan ahead and reason through tasks more effectively than its predecessors.

Within 2025, we most likely will hit the milestone of AGI. The arrival of AGI holds the potential to revolutionise society and industries by addressing challenges that require deep understanding, creativity, and adaptability. AGI could accelerate breakthroughs in medicine by designing cures for complex diseases, address climate change through innovative solutions, and redefine education with highly personalised learning systems. However, AGI also comes with significant ethical, social, and regulatory challenges. Preparing for AGI means creating robust frameworks for alignment, collaboration, and oversight to harness its potential responsibly.

In short, AGI isn’t just an advancement in technology—it’s a paradigm shift that could redefine what it means to innovate, learn, and thrive in a rapidly changing world.

The Era of Agentic AI

Agentic AI refers to advanced systems capable of reasoning, planning, and making decisions independently, without requiring constant human input or pre-set instructions. Unlike traditional AI, which reacts to specific commands, agentic AI can evaluate situations, identify goals, and take actions to achieve them proactively.

For example:

Autonomous Vehicles: These systems can navigate unexpected road conditions, such as detours or hazardous weather, by making real-time decisions.

Virtual Agents: AI systems might negotiate contracts or make financial decisions on behalf of users, acting as independent advisors or representatives. But before we get to major decisions, we would see minor transactional decisions being made sooner with some risk appetite.

In essence, agentic AI goes beyond simply following rules; it acts as a problem-solver, adapting to dynamic environments.

By automating complex or labor-intensive tasks, agentic AI can revolutionise industries like logistics, where it can optimise supply chains, or healthcare, where it might assist in personalised patient care or remote surgeries. In high-risk environments, such as mining or disaster response, agentic AI can take over dangerous tasks, reducing human exposure to harm.

However, this level of autonomy raises critical questions about ethics and accountability. For instance, if an agentic AI makes an error, who is responsible—its creator, its user, or the AI itself?Governments and organisations will need robust laws to regulate how agentic AI systems operate and ensure their decisions align with societal values.

Agentic AI is not just about smarter systems; it’s about creating machines that can think and act independently, unlocking unprecedented potential while posing new challenges for humanity to navigate responsibly.

Near Infinite Memory AI Systems: Remembering Context

Near infinite memory AI systems can store and recall vast amounts of contextual information from past interactions, unlike traditional AI, which forgets quickly. This enables AI to deliver more personalised and adaptive responses by remembering user preferences, interactions, and history. For example: A virtual assistant might recall your favourite music or past goals. An AI doctor could track your medical history and provide tailored advice.

These can lead to massive benefits

• Personalisation: AI can offer responses and services uniquely tailored to each user.
• Adaptation: Over time, AI improves its behaviour by learning from past interactions.

Applications include personalised healthcare, adaptive education, and improved customer service. However, storing such detailed data raises privacy concerns, making robust data governance essential to ensure responsible and secure use.

The Emergence of Very Large Language Models (VLLMs)

Very Large Language Models (VLLMs) are AI systems with an immense number of parameters—often in the billions or even trillions. These parameters act like connections in a virtual brain, allowing the models to understand language with incredible depth and precision.

Llama comes with 7, to 70 billion parameters. OpenAI and Google do not disclose the number of parameters anymore but sources estimate that GPT 4 and Gemini 2 contain hundreds of billions to over a trillion parameters. In June 2023, just a few months after GPT-4 was released, George Hotz publicly explained that GPT-4 was comprised of roughly 1.76 trillion parameters. And Gemini, as some experts claim had 1.56 billion parameters as of 2023. The count may have gone way up now.

Trained on vast and diverse datasets, VLLMs can perform complex tasks such as:

Language Understanding: Grasping context, tone, and intent in conversations.

Creative Writing: Generating stories, articles, or poetry that feel human-like.

Multilingual Translations: Translating text across many languages with cultural sensitivity and accuracy.

These models excel at interpreting and generating text that aligns with human expectations, bridging the gap between people and machines.

Despite these challenges, VLLMs represent a cornerstone of modern AI, shaping how we interact with technology in both personal and professional settings. Their ability to process and generate nuanced language makes them indispensable across industries.

We might already be running out of fresh training data but in 2025, chances are strong that the parameter sizes will continue to go up.

The Rise of Very Small Models: Efficiency at the Edge

Very Small Models are designed for resource-constrained devices, such as smartphones, IoT sensors, and wearables, focusing on efficiency while delivering robust, task-specific AI capabilities. These models are vital for edge computing, enabling AI to function locally without needing powerful servers, making AI more accessible and responsive.

Where Small Models Excel

Task-Specific Applications: Small models excel at specific tasks, such as voice assistants on smartphones and real-time translations.

Edge Devices and IoT: They power healthcare wearables and IoT sensors, offering low-latency decision-making in critical applications, even in remote areas.

Accessibility in Remote Areas: Small models help deploy AI where infrastructure is limited, including education tools and disaster response drones.

Focus for 2025

Efficiency and Optimisation: The focus will be on energy efficiency and enabling transfer learning, so models can learn with minimal data.

Privacy and Security: Federated learning and on-device encryption will enhance data privacy and security, making local AI more reliable.

Hybrid Systems: Combining small models with larger ones for local tasks and deeper insights from the cloud will maximize efficiency.

In 2025, very small models will help bring AI to the edge, making it more efficient, secure, and accessible, transforming industries and daily life.

Sustainable AI: Reducing AI’s Carbon Footprint

As the size of AI models increases, so do their energy demands, leading to significant environmental concerns. For example, training large language models (LLMs) requires immense computational power, which translates to high electricity consumption.

GPT-3: Training GPT-3 consumed approximately 1,287 MWh of electricity, emitting about 550 tons of CO?, equivalent to the emissions of a gasoline-powered car driving 1.3 million miles.

GPT-4: Training GPT-4 is estimated to have used between 51,773 MWh and 62,319 MWh, over 40 times higher than GPT-3’s consumption.

Just as one example of growing energy demand, Microsoft has undertaken significant steps. In a 20-year agreement with Constellation Energy, Microsoft plans to purchase about 835 MW of power from the Three Mile Island Unit 1 nuclear plant, aiming to supply energy to its AI data centers. It has committed to pairing its constant electricity usage with clean energy sources, supporting its decarbonisation goals. And so are other organisations doing.

As AI models grow in size, their energy demands surge, leading to significant environmental impacts. Sustainable AI focuses on developing energy-efficient algorithms, leveraging renewable energy for training, and prioritising low-power deployments. Techniques like model distillation and edge computing are central to this effort.

Unchecked growth in AI energy consumption could negate its societal benefits by exacerbating climate change. By 2025, sustainable AI practices will be a competitive differentiator, with organizations and nations prioritising environmentally friendly technologies to align with global climate goals.

The focus in 2025 will be on energy efficiency and enabling transfer learning, so models can learn with minimal data. Combining small models with larger ones for local tasks and deeper insights from the cloud will maximise efficiency.

The Promise of Open-Source AI Models: Democratising Innovation

Open-source AI models, like Meta’s Llama, represent a significant shift in how advanced AI technology is shared and utilised. These models are made freely accessible to researchers, developers, and organisations, allowing anyone to use, modify, and build upon them. Open-source AI fosters collaboration, transparency, and creativity by providing the underlying architecture and training data to the community, enabling widespread experimentation and innovation.

Unlike proprietary models, which are controlled by a few large organisations, open-source models decentralise access to cutting-edge AI capabilities. They empower smaller businesses, academic institutions, and even individuals to develop advanced AI solutions without the massive resource investment typically required.

Open-source AI is a game-changer in democratising access to technology. Models like Llama enable underrepresented regions, startups, and independent researchers to contribute to and benefit from AI advancements. This openness drives innovation in areas like healthcare, education, and sustainability, where localised solutions often require cost-effective AI tools.

Moreover, open-source models enhance transparency and trust. By allowing experts to inspect and improve the code, they help identify and mitigate biases, improve performance, and foster ethical AI development. They also accelerate AI adoption across industries, enabling tailored solutions for specific challenges without the high costs associated with proprietary systems.

In a world where the benefits of AI are often concentrated in the hands of a few, open-source models like Llama represent a powerful movement toward equitable access, collaborative problem-solving, and global technological progress.

Augmented Working: Productivity Reimagined

Augmented working involves the seamless integration of AI into human workflows. From AI co-pilots in software development to intelligent assistants for medical diagnosis, this trend enhances human decision-making rather than replacing it.

It is about transforming how we approach productivity by integrating AI into human workflows in a way that enhances, rather than replaces, human capabilities.

In sectors like software development, AI co-pilots assist developers by suggesting code, identifying bugs, and even automating repetitive tasks. Similarly, in healthcare, AI-driven assistants can analyse patient data and provide diagnostic insights, helping doctors make informed decisions faster.

This shift reduces the cognitive load on workers, allowing them to focus on creative and high-value tasks. By automating routine processes, augmented working not only boosts productivity but also helps mitigate employee burnout. It empowers workers to make better decisions with real-time, data-driven insights.

As AI continues to evolve, augmented working is poised to reshape industries by allowing humans and machines to collaborate seamlessly, driving both efficiency and innovation.

Ultimately, it represents a vision of work where AI augments human potential, fostering a more dynamic and fulfilling work environment.

In 2025, productivity will be significantly boosted by the widespread adoption of AI-driven automation and augmented working. AI systems will seamlessly integrate into human workflows, handling repetitive tasks and providing data-driven insights, allowing workers to focus on higher-value, creative activities. AI co-pilots will assist in industries like software development, healthcare, and customer service, enhancing decision-making and reducing cognitive load.

Generative AI Beyond Text: The Video Frontier

Generative video AI lets users create videos from text descriptions or simple inputs, combining visuals, sound, and stories to make realistic or creative videos. This technology uses advanced AI models and large amounts of data to produce videos that make sense and fit together, reducing the need for traditional video production teams.

With generative video AI, creating content becomes faster and cheaper. It helps businesses and individuals communicate more effectively, whether for marketing, learning materials, or other creative projects. In the entertainment industry, it allows faster and more affordable content creation, which is great for independent creators. When combined with virtual reality, it also offers users interactive, immersive experiences.

Some of the top tools for creating AI-generated videos include Synthesia, Sora, Runway, Canva, DeepBrain AI etc.

Companies like Nike and L’Oréal use Synthesia for training, marketing, and internal communications. By automating video production, they reduce the need for expensive filming and post-production processes. It is estimated that these organisations save up to 80% on video production costs while significantly speeding up content creation.

In 2025, generative video AI is expected to revolutionise content creation by producing highly realistic and interactive videos from text prompts. AI-generated videos will become more cost-effective, enabling individuals and small businesses to create professional-grade content. Integration with augmented and virtual reality will enhance user experiences, making them more immersive and engaging. Additionally, video AI tools will be more accessible, offering advanced features to users without technical expertise, democratising video production for various industries.

Real-Time Automated Decision-Making: Lightning-Fast Insights

Real-time AI enables systems to analyse data, derive insights, and make decisions within milliseconds. This capability is vital in dynamic environments like stock trading, autonomous driving, and emergency healthcare responses. As AI models become more advanced, they can process vast amounts of live data quickly and accurately, allowing for instantaneous decisions that would traditionally take minutes or hours.

The speed and accuracy of automated decision-making are critical in industries where delays could result in financial loss, operational inefficiency, or life-threatening consequences. By 2025, advancements in this area will redefine what “immediate” means in decision-making contexts. For instance, in autonomous vehicles, real-time decision-making will improve safety by allowing cars to adapt to sudden changes in road conditions. In finance, AI will enable high-frequency trading algorithms to identify and exploit market opportunities with unprecedented speed.

Beyond this, with real time tools from the top AI firms may affect the voice based job profiles like customer care too.

NextGen Voice Assistants: Conversational AI Evolved

Voice assistants are becoming more conversational and emotionally intelligent, capable of understanding nuances in tone, intent, and context. Multilingual capabilities and real-time learning are also on the horizon. This progress will allow voice assistants to adapt to diverse conversational environments, offering more fluid, natural exchanges between humans and machines.

These advancements will redefine accessibility and human-computer interaction. From assisting the elderly to serving as emotional companions, next-gen voice assistants will find applications across diverse demographics. Enhanced empathy and personalised interactions will make voice assistants invaluable in caregiving settings, helping people with disabilities or elderly individuals navigate daily tasks. Additionally, businesses will leverage these more intelligent voice assistants to provide better customer support, improving user experience and engagement across various industries.

Final Thoughts

The rise of AGI-level reasoning and deeper token-based thinking is likely to amplify trends like Agentic AI and Real-Time Decision-Making. All these trends collectively signal an era of unprecedented transformation.

Which of these trends do you believe will have the most profound impact in 2025? How should we prepare for the future they promise?