The past year has been transformative for artificial intelligence, marked by breakthrough innovations, emerging regulations, and a shift toward practical AI tools that enhance productivity. As we look ahead to 2025, let's review the major developments of 2024 and explore what the future might hold.
Part I: 2024 Year in Review
OpenAI in the front
OpenAI maintained its position at the forefront of AI innovation in 2024. The release of GPT-4o in May marked a significant milestone as the first frontier multimodal model, capable of understanding and generating content across different forms of media. This was followed by o1 in September, along with its lighter counterpart o1-mini, and the announcement of o3 in December.
Perhaps most notably, OpenAI's Sora project revolutionized video generation. Announced in February and enhanced with Sora Turbo in December, this technology demonstrated unprecedented capabilities in creating realistic video content from text descriptions.
Agentic AI and AI Assistants
Rather than replacing jobs, AI in 2024 focused on enhancing human productivity through innovative tools and assistants. Google introduced several groundbreaking projects: Astra, an AI assistant for phones and smart glasses, and Mariner, a Chrome extension that enables Gemini to interact with your browser. These tools represent a shift from simple chat interfaces to interactive agents that can understand and manipulate our digital environment.
Anthropic joined this trend by enabling Claude to use computers the way humans do—viewing screens, moving cursors, and interacting with interfaces. This development opened new possibilities for AI assistance in everyday computer tasks.
The developer community saw particularly exciting advances. Github's Copilot and Cursor AI have been very well received by the developer community. On the other hand, autonomous coding agents like Devin and Github Workspace are still rough around the edges. These tools aren't replacing developers; instead, they're supercharging their productivity by handling routine tasks and suggesting improvements.
AI Regulation
2024 saw significant progress in AI regulation, particularly with the EU AI Act. This groundbreaking legislation set the first comprehensive framework for AI regulation, establishing rules for AI system development and deployment while balancing innovation with safety and ethical concerns.
Part II: Looking Ahead to 2025
Small, specialized models
While decoder-only models like GPT and Claude have dominated headlines, in 2025 encoder-only models are still an important piece in AI applications. Just one week before ending the year, ModernBERT was released, representing a significant advance in this space that was long overdue. These models are particularly crucial for Retrieval-Augmented Generation (RAG) setups, where they excel at information retrieval and classification at a significantly lower cost to run.
In a similar fashion, the current trend of agentic systems will favor small, specialized models. These models will be fine-tuned to be as good as frontier models in really specific tasks, but used in conjunction in an agentic workflow, will allow for really powerful and resource efficient AI systems.
Multimodal models and Test-time compute
We can expect significant advances in multimodal capabilities, with AI systems becoming more natural at processing and responding to various forms of input—text, voice, images, and video—simultaneously. These improvements will make AI interactions feel more natural and contextually aware.
The concept of "test-time compute"—giving models more time to think—emerged as a game-changer in 2024 and will likely become more prominent in 2025. Frontier models like OpenAI’s o3 are crushing reasoning, math and coding benchmarks. They are consistently beating humans at really complex, deep reasoning tasks, making us wonder how far Artificial General Intelligence really is.
Test-time compute is not only showing remarkable results for frontier models. Used in small edge models, they are surpassing the performance of much larger models when given more processing time. For instance, Llama 3.2 3B outperformed Llama 3.1 70B in the Math-500 benchmark using a test-time compute strategy of 256 iterations.
The evolution of AI regulation
Following the EU's lead, we're likely to see more regions implement AI regulations in 2025. Interestingly, major tech companies are actively calling for regulation, potentially as a strategy to shape the regulatory environment and raise barriers to entry for competitors.
Conclusion
2024 has been a year of remarkable progress in AI. We have seen many new ways to use AI in our day-to-day lives, as well as new methods to improve models that are making giant strides closer to AGI. As we move into 2025, the focus appears to be shifting toward more specialized and efficient AI systems, improved reasoning capabilities through techniques like test-time compute, and a focus on AI tools to improve productivity.
Top comments (0)