AI trends worth following in 2025
This is our third year of forecasting the future of AI! Come along as we explore where AI progress is going in 2025.
AI had a turbulent 2024. OpenAI’s ChatGPT became an even bigger part of people’s everyday lives, and the familiar “let’s Google it” is gradually turning into “Let me check in ChatGPT”. At the same time, many tools rely on the offering of its main rival – Claude.ai by Anthropics (since it performs better on coding tasks).
We also saw Vision move forward, with smooth video generation from OpenAI Sora and Runway, as well as the video-based Segment Anything-2, which considerably improved tracking and labelling in Computer Vision applications.
This is the third year in a row that we are forecasting the future of AI! Come along as we explore where AI progress is going in 2025.
AI avatars spreading into the wider world
We have gone all in on GenAI in AI and handed over certain tasks entirely (e.g. the first point of contact in customer support). Yet we still interact with it through text or audio. My prediction is that avatars will grow substantially, making the user experience better.
Synthesia is a good example: it provides ready-made avatars with a smooth experience and over 80 languages, and it markedly raises visual quality in terms of facial expressions, lip-sync with the audio, and hand gestures.
Building LLM agents for a wider audience
Flowrise, a no-code tool for creating agentic workflows.
2024 brought a large number of PoCs and MVPs (frequently released as betas) built on agentic workflows, for example editing files in chatGPT (where one of the visible steps is generating code that loads and modifies them) or AI assistants such as Devin. Sadly, building such tools usually takes a programmer, whereas many “Prompt Engineers” have never laid eyes on multithreaded Python code.
First signs of this trend are already visible, with Microsoft launching the Copilot agent in SharePoint and the no-code offering from Flowrise. Both are still at an early stage, but they may help agents become as popular as SaaS solutions were in the 2000s and 2010s.
Open-source LLMs catching up with proprietary ones
Next year, the performance difference between open-source models and the services offered by the big LLM players, such as OpenAI and Antrophic, may shrink. An early sign of this could be the recent Qwen/Qwen2.5-Coder-32B-Instruct model, which has beaten GPT4-o on programming tasks.
Agentic systems growing up
2024, with its unexpected leaps in generative AI models, promised everyone autonomous agentic AI developers in the shape of the Devin project. That inspired new open-source projects aiming to replicate the showcased assistant’s functionality, such as devika and OpenHands (previously OpenDevin). When I tried these tools (in the first half of 2024), they showed plenty of early-stage issues: lots of bugs, not working out of the box, and missing the promised ability to solve complex programming tasks autonomously.
Over time, such solutions will become more stable and offer increasingly useful features. It is only a matter of time before they turn into an essential companion for every engineer in IT. That may not happen in 2025, but this could be the year these projects reach a point where they become useful and, thanks to the snowball effect, start getting better quickly.
Moving toward more data-efficient LLM models
Over the last few years, we have grown accustomed to Copilots, Chatbots and every other application built on top of LLMs. So far, all of them have gone through the cycle from “They will take our jobs” to “They are useless”. My impression is that progress on these models has somewhat stalled, and feeding them ever more data does not ensure better outcomes. Even if we tried, the data available to us is a finite set. To get better results, we need to come up with a better architecture.
We know this can be done, since our brains manage it with far less data. We simply do not yet know the right way to do it, but we are working on it. Our knowledge of how the brain is structured keeps expanding, and engineers can draw on it to improve things and come up with new ideas. I expect 2025 and the years after it to deliver more efficient models that learn faster from less data, marking the start of NEXT-GEN AI.
Personal, real-time doctors for a healthy lifestyle
Garmin Connect, health statistics.
Most of us enjoy gathering data about ourselves. We wear smartwatches, rings, bands, and plenty of other devices that record our lifestyle and medical data. That data can be analysed in real time to produce insights into our health. In an era of healthy lifestyles and AI-powered devices, a personal doctor that monitors our health and alerts us to any problem right away is something the market needs.
Meanwhile, Machine Learning models in professional medical devices keep improving and keep outperforming humans. AI will therefore continue to be used for analysing medical images, and models that analyse and describe examinations will be added to that.
Wider adoption of AI-based tools: not just LLMs
Next year will be shaped by AI models being widely adopted across every business domain. Companies will mostly concentrate on solutions built on Large Language models, but we will also see solutions combining LLMs with Computer Vision, such as document understanding software. Expect a wide range of AI-powered content analysis applications – for instance customer feedback analysis, reasoning over large volumes of customer data, product improvement suggestions, automated email handling, problem detection, issue prioritisation, first-line customer contact, marketing campaign planning, content creation and more.
We may also see progress in Bot-developers. AI models could gradually replace the most junior and cheapest IT developers. AI already goes beyond simple IT tasks and is capable of writing good-quality code.
Disillusionment with LLMs
In 2025, approaches to LLM-based applications will be reconsidered and redefined. Once the initial excitement fades, we will wrap up PoCs with reflections on operating costs, data privacy, and how much generality prompt-engineering alone can deliver. This will kick off the development of smaller models using quantization or knowledge distillation, enabling more personalized fine-tuning of privacy-preserving models without sending any data to external API endpoints. It will also come with the recognition that LLMs – particularly external ones such as ChatGPT or Claude – are not the ideal answer to literally every problem.
Cybersecurity in AI, AI in Cybersecurity
After AI, the second hottest field in IT at the moment is everything covered by the broad term cybersecurity. Naturally, this leads to growing interest in using AI for cybersecurity – and AI has many potential uses there, from unsupervised detection of anomalous behaviour to highly specialized models that detect particular attack vectors. That said, I expect strong interest in protecting AI models from malicious use as well. It will be a good year for AI Red Teams to uncover model vulnerabilities, carry out adversarial attacks, and extract information from training datasets. I hope ML engineers will counter them by hardening AI solutions and making models increasingly resistant to adversarial inputs.
Conclusions
Let’s see what the coming year has in store and how close our predictions turn out to be! And keep in mind that AI is only one piece of the tech puzzle.