The most important AI trends in 2024

The year 2022 represented a turning point when artificial intelligence (AI) became a topic of public debate, and 2023 was the year when it began to be used in business. The year 2024 brings new challenges and opportunities as researchers and businesses look for ways to best integrate AI into everyday life.

In this article, we'll take a look at some of the most important trends in AI that we should prepare for and be able to take advantage of in the coming year.

Multimodal AI and video

The next wave of advancements will not only focus on improving performance within a specific domain, but also on multimodal models that can use multiple types of data as input. While models that operate across different data modalities are not exactly a new phenomenon, text-to-image models like CLIP and speech-to-text models like Wave2Vec have been around for years. They usually work in only one direction and have been trained to accomplish a specific task.

The upcoming generation of interdisciplinary models includes proprietary models such as OpenAI's GPT-4V or Google's Gemini, as well as open source models such as LLaVa, Adept or Qwen-VL. This generation can move freely between natural language processing (NLP) tasks, computer vision, and even bring video into play. In late January, Google announced Lumiere, a model for creating video from text that can also perform tasks from an image or use images as a style reference.

These advancements allow for the creation of more intuitive AI apps and virtual assistants, bringing new possibilities for interacting with technology.

Smaller language models and open source enhancements

In domain-specific models - particularly LLM, we have arguably reached a point of diminishing returns from more parameters.

Only the largest companies have the funds and server space to train and maintain energy-intensive models with hundreds of billions of parameters. It is estimated that training a single model the size of GPT-3 requires the annual electricity consumption of more than 1,000 homes. A standard day of ChatGPT demands is equivalent to the daily energy consumption of 33,000 U.S. households.

Smaller models, meanwhile, are much less resource-intensive. A March 2022 study by Deepmind showed that training smaller models on more data yields better performance than training larger models on less data. Thus, much of the ongoing innovation in LLM has focused on getting more performance from fewer parameters.

The power of open models will therefore continue to grow. In December 2023, Mistral released "Mixtral", which integrates 8 neural networks, each with 7 billion parameters. The company claims that it not only outperforms the 70B variant of the Llama 2 parameter in most benchmarks at 6 times the inference speeds, but even matches or outperforms the much larger GPT-3.5 OpenAI in most standard benchmarks.

These advances in smaller models have three important advantages:

• They make the AI more understandable: the larger the model, the harder it is to determine how and where it makes important decisions. Understandable AI is essential to understanding, improving, and trusting the outputs of AI systems.
• They help democratize AI: smaller models that can be run at lower cost on more affordable hardware allow more amateurs and institutions to study, train, and improve existing models.

They can be run locally on smaller devices: enabling more sophisticated AI in scenarios such as edge computing and the Internet of Things (IoT). In addition, running models locally on a user's smartphone, for example, helps circumvent many of the privacy and cybersecurity concerns that arise when interacting with sensitive data.

GPU shortage and cloud costs

The trend towards smaller models will be driven as much by necessity as by business pressure, as the cost of cloud computing rises as hardware availability decreases.

While large companies are investing in in-house computing resources, an increasing number of users are relying on cloud services. It is therefore important for companies to strike a balance between efficient smaller models and more powerful but more expensive models.

These trends present new challenges and opportunities in AI and are important for the further development of the technology in 2024.