Introduction

Artificial intelligence (AI) is one of the most exciting and rapidly evolving fields of technology in the 21st century. In 2023, we witnessed remarkable advances in AI, such as generative models that can create realistic images, texts, and sounds from simple prompts, and multimodal models that can understand and generate different types of data, such as text, images, audio, and video.

But what will 2024 bring for AI? What are the trends, challenges, and opportunities that will shape the future of this technology? To answer these questions, we turned to one of the leading experts in AI, Martin Signoux, the head of policy at Meta France, formerly known as Facebook France. Signoux is a visionary thinker who has been involved in AI research and development for over a decade. He is also a member of the High-Level Expert Group on Artificial Intelligence (AI HLEG), a body that advises the European Commission on AI ethics and governance.

Signoux shared his predictions for AI in 2024 in a recent blog post, where he outlined eight scenarios that he believes will happen or become more prominent in the next year. Here is a summary of his predictions:

Prediction 1: AI smart glasses become a thing

Signoux predicts that as multimodality rises, leading AI companies will double down on AI-first wearable devices. And what’s better than the glasses form factor to host an AI-assistant?

AI smart glasses are devices that can project information and images onto the lenses, as well as capture and process data from the environment. They can also interact with the user through voice, gestures, or eye movements. Signoux believes that AI smart glasses will become more popular and accessible in 2024, as they offer a convenient and immersive way to access AI services and applications.

Some examples of AI smart glasses that are already in development or on the market are Google Glass, Meta’s Project Aria, and Snap’s Spectacles. These devices aim to provide users with augmented reality (AR) experiences, such as overlaying digital information on the physical world, or enhancing social interactions with filters and effects. Signoux expects that AI smart glasses will also enable new use cases, such as education, entertainment, health, and productivity.

• Prediction 2: ChatGPT won't be to AI assistant what Google is to search

Signoux predicts that 2023 started with ChatGPT taking all the light and ends with Bard, Claude, Llama, Mistral and thousands of derivatives. As commoditization continues, ChatGPT will fade as THE reference, leading to a valuation correction.

ChatGPT is a generative model that can produce natural language responses to any text input. It is based on GPT-3, a large-scale language model developed by OpenAI, a research organization backed by Elon Musk and other tech billionaires. ChatGPT was launched in 2023 by Meta, and quickly became the most popular and widely used AI assistant in the world, surpassing Siri, Alexa, and Cortana.

However, Signoux argues that ChatGPT is not as dominant and innovative as it seems. He points out that ChatGPT is essentially a rebranded and fine-tuned version of GPT-3, which is available to anyone who can access OpenAI’s API. He also notes that ChatGPT faces fierce competition from other generative models, such as Bard, Claude, Llama, and Mistral, which are developed by different companies and organizations, and offer more diversity and customization. Signoux predicts that ChatGPT will lose its market share and hype in 2024, as users and investors realize that it is not the ultimate AI assistant.

• Prediction 3: So long LLMs, hello LMMs

Signoux predicts that Large Multimodal Models (LMMs) will keep emerging and oust LLMs in the debate; multimodal evaluation, multimodal safety, multimodal this, multimodal that. Plus, LMMs are a stepping stone towards truly general AI-assistant.

LLMs, or Large Language Models, are AI models that can generate natural language texts based on a large corpus of data. Examples of LLMs are GPT-3, BERT, and T5. LLMs have been the focus of AI research and development in the past few years, as they have shown impressive results in various natural language processing (NLP) tasks, such as question answering, summarization, and translation.

However, Signoux predicts that LLMs will be overshadowed by LMMs, or Large Multimodal Models, in 2024. LMMs are AI models that can understand and generate different types of data, such as text, images, audio, and video. Examples of LMMs are DALL-E, CLIP, and COCO-GPT. LMMs have the potential to enable more rich and diverse AI applications, such as content creation, image captioning, and video synthesis.

Signoux also believes that LMMs are a stepping stone towards truly general AI-assistant, or AGI, which is the ultimate goal of AI research. AGI is the hypothetical AI that can perform any intellectual task that a human can. Signoux argues that LMMs are closer to achieving AGI than LLMs, as they can handle more complex and varied data and tasks.

• Prediction 4: No significant breakthrough, but improvements on all fronts

Signoux predicts that new models won't bring real breakthroughs (GPT-5) and LLMs will remain intrinsically limited and prone to hallucinations. We won’t see any leap making them reliable enough to "solve basic AGI" in 2024. Yet...iterative improvements will make them “good enough” for various tasks. Improvements in RAG, data curation, better fine-tuning, quantization, etc, will make LLMs robust/useful enough for many use-cases, driving adoption in various services across industries.

Signoux predicts that 2024 will not see any major breakthroughs in AI model development, such as a new model that surpasses GPT-3 or GPT-4 in performance and capabilities. He argues that LLMs will remain intrinsically limited and prone to hallucinations, which are errors or inconsistencies in the generated texts. He also claims that we won’t see any leap making them reliable enough to "solve basic AGI" in 2024, meaning that they won’t be able to achieve human-like intelligence and reasoning.

However, Signoux also predicts that 2024 will see iterative improvements in various aspects of AI model development, such as RAG, data curation, better fine-tuning, quantization, etc. RAG, or Retrieval-Augmented Generation, is a technique that allows LLMs to retrieve relevant information from external sources, such as Wikipedia, to improve the quality and accuracy of the generated texts. Data curation is the process of selecting, cleaning, and organizing the data that is used to train the LLMs, to ensure that they are representative and unbiased. Fine-tuning is the process of adjusting the parameters of the LLMs to optimize them for specific tasks or domains, such as medical or legal texts. Quantization is the process of reducing the size and complexity of the LLMs, to make them faster and more efficient.

Signoux predicts that these improvements will make LLMs robust and useful enough for many use-cases, driving adoption in various services across industries, such as education, health, entertainment, and business.

Prediction 5: Small is beautiful

Signoux predicts that Small Language Models (SLMs) are already a thing, but cost-efficiency and sustainability considerations will accelerate this trend. Quantization will also greatly improve, driving a major wave of on-device integration for consumer services.

Signoux predicts that 2024 will see a rise in the popularity and development of SLMs, or Small Language Models, which are AI models that can generate natural language texts with a smaller size and complexity than LLMs. SLMs have several advantages over LLMs, such as lower cost, lower energy consumption, and lower latency. SLMs are also more suitable for on-device integration, which means that they can run on the user’s device, such as a smartphone or a smartwatch, without relying on the cloud or the internet.

Signoux expects that cost-efficiency and sustainability considerations will accelerate this trend, as users and developers become more aware and concerned about the environmental and economic impact of LLMs. LLMs require a lot of computational resources and electricity to train and run, which can be expensive and harmful to the planet. SLMs, on the other hand, are more affordable and eco-friendly, as they consume less power and resources.

Signoux also expects that quantization will greatly improve, driving a major wave of on-device integration for consumer services. Quantization is the process of reducing the size and complexity of the LLMs, to make them faster and more efficient. Quantization can also be applied to SLMs, to make them even smaller and more compact, without compromising their quality and functionality. Signoux predicts that quantization will enable more SLMs to run on the user’s device, offering more privacy, security, and convenience.

• Prediction 6: An open model beats GPT-4, yet the open vs closed debate progressively fades

Signoux predicts that looking back at the dynamism and progress made by the open source community over the past 12 months, it’s obvious that open models will soon close the performance gap. We’re ending 2023 with only 13% left between Mixtral and GPT-4 on MMLU. But most importantly, Signoux predicts that an open model will beat GPT-4, yet the open vs closed debate will progressively fade. He explains that the open source community has been making remarkable progress in developing and improving AI models, and that it is only a matter of time before an open model surpasses GPT-4, the latest and most powerful LLM developed by OpenAI. He cites Mixtral, a hybrid model that combines several open models, such as BART, T5, and PEGASUS, as an example of a strong contender that is closing the performance gap with GPT-4. He also mentions MMLU, or Multimodal Language Understanding, as a benchmark that measures the ability of AI models to understand and generate different types of data, such as text, images, audio, and video.

However, Signoux also predicts that the open vs closed debate will progressively fade, as the AI community and the public become more aware and respectful of the different motivations and trade-offs behind each approach. He argues that both open and closed models have their pros and cons, and that there is no one-size-fits-all solution for AI development. He suggests that instead of focusing on the dichotomy of open vs closed, the AI community and the public should focus on the values and principles that guide AI development, such as transparency, accountability, fairness, and safety.

• Prediction 7: AI regulation will be a hot topic

Signoux predicts that AI regulation will be a hot topic in 2024, as more countries and regions will adopt or propose legal frameworks and guidelines to govern the development and use of AI. He expects that the European Union will lead the way, as it has already published a draft proposal for the Artificial Intelligence Act, which aims to create a harmonized and risk-based approach to AI regulation in the EU. He also anticipates that other countries and regions, such as the US, China, India, and Africa, will follow suit, or at least engage in dialogues and collaborations with the EU and other stakeholders.

Signoux believes that AI regulation is necessary and beneficial, as it can help ensure that AI is used in a responsible and ethical manner, and that it respects the fundamental rights and values of humans. He also thinks that AI regulation can foster trust and confidence among the users and the developers of AI, as well as create a level playing field and a competitive edge for the AI industry.

• Prediction 8: AI ethics will be more than a buzzword

Signoux predicts that AI ethics will be more than a buzzword in 2024, as more AI practitioners and researchers will adopt and implement ethical principles and practices in their work. He expects that AI ethics will become more mainstream and accessible, as more resources and tools will be available to help AI practitioners and researchers to assess and address the ethical implications and impacts of their AI projects. He also hopes that AI ethics will become more inclusive and diverse, as more voices and perspectives will be heard and represented in the AI ethics discourse and decision-making.

Signoux asserts that AI ethics is more than a buzzword, as it is a crucial and integral part of AI development and innovation. He argues that AI ethics is not only a moral duty, but also a strategic advantage, as it can help enhance the quality and reliability of AI, as well as increase the social acceptance and adoption of AI. He also claims that AI ethics is not only a challenge, but also an opportunity, as it can inspire and motivate AI practitioners and researchers to create more human-centric and value-driven AI solutions.

Conclusion

These are the predictions of Martin Signoux, the head of policy at Meta France, for AI in 2024. Signoux is a prominent and influential figure in the AI field, and his predictions reflect his vision and insights on the current and future state of AI. Whether his predictions will come true or not remains to be seen, but one thing is certain: AI will continue to be one of the most fascinating and impactful technologies of our time.