2024 was a real turning point for generative AI, marked by spectacular technological advances. La The rise of open source, the improvement of model performances and the drastic fall in costs have redefined the landscape. At the same time, companies are adopting these tools with better controlled and more robust methods, while regulatory, security or cybersecurity issues take a central place.
What lessons can we learn from this key year, and what challenges await us in 2025?
The performances of the models converge and the border of reasoning is blurring.
- The undivided dominance of OpenAI is being questioned and open source still seriously competes with proprietary solutions. This is in particular thanks to a increasingly dynamic open source research that brings together industry and academia . However, transparency is often limited, with the exception of some cases like DeepSeek.
- Models go beyond simple language and venture into maths, the medicine, the code and other specialized areas. However, AI is still lagging behind on complex tasks (very advanced mathematics, visual reasoning, planning). For these tasks, new, tougher benchmarks are emerging because the old ones are already outdated.
- Image, video and audio: the multimodal generation is becoming convincing.
- At the same time, China massively dominates AI patent filings over the last decade even if the US dominates the cutting-edge models. The total number of AI patents is exploding. We could therefore continue to see AI advance at a very significant speed over the coming year.
Methods are evolving and businesses are becoming more professional
- With the extension of LLM context windows combined with advances in embeddings and retrieval methods, Augmented Retrieval Generation (RAG) is establishing itself as the reference for exploiting generative AI in business.
- The implementation of monitoring the performance of solutions has become essential in the industry to ensure the sustainability of developments.
- The critical importance of conducting rigorous feasibility studies based as much on the needs of end users as on purely technical and technological aspects.
- The rise of the roles of Chief Knowledge Officer, making it possible to structure in depth the access and production of knowledge in a company.
Model costs have fallen dramatically
- -100x at OpenAI, -60x at Anthropic in one year. This decrease can be explained in part by the increased availability of GPU resources.
- Quantization now makes it possible to compress models effectively open source without sacrificing performance.
- The fine-tuning of the models reaches a new level with methods like DPO and PPO, making adaptation more effective and better aligned with real uses.
- The generation of synthetic data is an essential tool for training more advanced models at a lower cost. In particular, it allowed the effective retraining of CoT (Chain of Thought) models without an explosion of expenses (cf. DeepSeek).
Challenge #1 Security, Regulation & Policy
AI regulation
- The regulation of AI is progressing, mainly in Europe, and physical constraints (computing power, emissions) are becoming a strategic factor.
Securing
- A change of environment from safety to acceleration (i.e. AI Doomers vs AI Accelerationists) is taking place as companies that previously warned us about the impending extinction of humanity need to increase their sales to businesses and the use of their consumer applications.
- The risks of sophisticated attacks (Prompt injection (as a figurehead) in the long term are of increasing concern to researchers.
2025 challenges
Cost and sustainability of AI models
- Move from models to viable products, with questions about costs and sustainability.
Reliability and robustness
- Ensure that progress on benchmarks results in reliable and robust value in the real world, beyond specific tasks.
Performance and energy consumption
- Reconciling performance and energy consumption of LLM models to meet corporate CSR issues.
Setting up AI factory
- Set up AI Factories rather than operating use case by use case in a disorganized way: i. Improving a company's knowledge management and making it one of the main drivers of growth ii. Streamlining scattered initiatives iii. Systematically estimate the impacts of the AI products developed.
2025 promises to be a decisive year: it is no longer just a question of improving models, but of transforming them into viable and sustainable solutions, while controlling their environmental and economic impact.
If you want to deepen these topics and structure your strategy around generative AI, Eulidia, a pure Data & AI player, is at your disposal to support you.
