AI trends to watch in 2025

2024 was a wild year for AI. OpenAI’s ChatGPT got even broader adoption in everyone's daily lives. That common phrase “let’s Google it” is slowly becoming “Let me check in ChatGPT”. On the other hand, a wide selection of tools also uses their main competitor's solution - Claude.ai by Anthropics (as it is superior in coding tasks).

Additionally, we saw further advancements in Vision, with smoothly generated videos by OpenAI Sora, Runway, and video-based Segment Anything-2, which greatly improved tracking and labelling in Computer Vision applications.

This marks our third year predicting the future of AI! Join us as we dive into where AI advancements are headed in 2025, and don't forget to revisit our predictions for 2023 and 2024.

AI avatars present in the broader world

In AI, we went all in with GenAI and fully delegated specific tasks (i.e. first point of contact in customer support). However, the interface with which we interact is still text or audio. I predict that we will see a significant rise in avatars, improving our user experience.

A good example is Synthesia, which offers predefined avatars with a smooth experience, more than 80 languages, and greatly improves visual quality regarding facial expression, mouth-to-sound synchronisation, and hand gestures.

LLM agents creation for a broader audience

Flowrise, no-code solution for agentic workflow creation.

In 2024, we saw a large selection of PoCs and MVPs (often released in beta versions) that used agentic workflows, i.e., editing files in chatGPT (where one of the visible steps is to generate code loading and changing) or AI assistants like Devin. Unfortunately, tool creation often requires a programmer, while a large portion of “Prompt Engineers” have never seen multithreaded Python code.

We can see glimpses of that trend with Microsoft introducing the Copilot agent in SharePoint or a no-code solution from Flowrise. Both tools are still in the early stages, however, thanks to that we might see the rise of popularity in agents in a scale comparable to SaaS solutions in the 2000s and 2010s.

Closing the gap between open-source and proprietary LLMs

In the coming year, we may see a narrowing performance gap between open-source models and services provided by major players in the LLM space, such as OpenAI and Antrophic. A herald of such a trend may be the latest Qwen/Qwen2.5-Coder-32B-Instruct model, which has overtaken GPT4-o in programming tasks.

More mature Agentic systems

The year 2024, with its surprising enhancements in generative AI models, promised us all autonomous AI agentic developers in the form of the Devin project. This inspiration results in emerging open-source projects that try to implement similar functionality of the presented assistant, such as devika or OpenHands (formerly OpenDevin). When I tested these tools (the first half of 2024), they exhibited many early-stage problems: a large number of bugs, lack of out-of-the-box functioning, and lack of promised functionality for autonomous solution of complex programming tasks.

As time passes, solutions of this type will begin to gain stability and provide more and more useful functionality. It is a matter of time before they become the indispensable companion of every engineer in the IT world. It may not be the perspective of 2025, but it could be the year to bring these projects to a state where they start to be useful and, thanks to the snowball effect, begin to improve rapidly.

Shift to more data-efficient LLM models

In recent years, we have gotten used to Copilots, Chatbots and all the applications using LLMs as their backbone. By this time, all those were following the life-cycle from “They will take our jobs” to “They are useless”. I feel like we are a little stuck with developing those models, and feeding them with more and more data doesn’t guarantee better results. Even if we try, the data we can collect is a finite set. We need to figure out a better architecture to acquire better results.

We know it’s possible because our brains can do it with less data. We just don’t know how to do it properly, but we are trying. Our understanding of brain structure is growing, and engineers can use it to improve and generate new ideas. I believe that 2025 and the following years will bring us more efficient models that will learn faster and with less data and will be the beginning of NEXT-GEN AI.

Personalized real-time doctors for healthy lifestyles

Garmin Connect, health statistics.

Most of us like to collect data about ourselves. We have smartwatches, rings, bands, and many more devices that gather our lifestyle and medical data. This data can generate insights about our health status and be analysed in real-time. Having a personal doctor who keeps track of our health and informs us about any problems immediately is a market-needed solution in times of healthy lifestyles and AI-powered devices.

On the other hand, Machine Learning models used in professional medical devices continue to achieve better results than humans and improve. Therefore, AI will continue to be used to analyze medical images, and models that analyze and describe the studies will join it.

Further adoption of AI based tools: not only LLMs

The upcoming year will be dominated by the wide adoption of AI models in all business domains. Companies will focus mainly on solutions utilizing Large Language models, but we can also see solutions using both LLMs and Computer Vision, such as software for document understanding. We can expect a broad spectrum of AI-based content analysis applications, like, for example, customer feedback analysis, reasoning about large customer data, suggestions for product improvements, automatic email handling, problem identification, issues prioritisation, the first line of facing the customer, marketing campaigns planning, content creation and more.

On the other hand, we may also see progress in Bot-developers. The AI models may slowly replace the lowest-grade and the cheapest IT developers. The AI capabilities are already beyond simple IT tasks and can successfully write quality code.

LLM disillusionment

2025 will be a year of rethinking and redefining approaches to LLM-based applications. After initial enthusiasm, we will conclude PoCs in terms of reflections on running costs, data privacy, and the degree of generality that can be achieved with only prompt-engineering. This will mark the start of the development of smaller models with quantization or knowledge distillation that would allow more personalized fine-tuning of privacy-preserving models without sending any data to external API endpoints. This would also be connected with the realization that LLMs, especially external ones like ChatGPT or Claude, are not a perfect-fit solution for literally every problem.

Cybersecurity in AI, AI in Cybersecurity

The second hottest area in IT, after AI, is right now everything that falls under the umbrella term of cybersecurity. The most natural consequence is increasing interest in applications of AI in cybersecurity - and there are quite a lot of possible applications of AI, starting from unsupervised detection of anomalous behaviours up to very specialized models for detecting specific attack vectors. Having said that, I think that there will also be great interest in securing AI models from malicious usage. This would be a good year for AI Red Teams to find model vulnerabilities, perform adversarial attacks, and extract information in training datasets. I hope that ML engineers will provide a counterbalance to them in hardening AI solutions and making models more and more robust to adversarial inputs.

Conclusions

Let's see what next year brings us and how accurate we have been! And remember, AI is just one part of the tech puzzle.

Reviewed by Michał Matłoka, Maria Kucharczyk