The world’s attention shifted when OpenAI made generative AI accessible to the public with ChatGPT. Then, the game changed again when GPT-4 was released – able to browse the internet with Bing and provide even more accurate information, including sources. Undoubtedly, the uprising of GenAI is already significantly impacting our way of working. But how will recent developments change our lives in the (near) future? Here’s everything you should know about GenAI trends in 2024.
What is GenAI?
First, a quick recap on generative AI. GenAI are complex deep learning models that learn from training data to generate new content, such as text, images, videos, or music. These models are trained on diverse datasets to provide human-like, original content. The best-known example is OpenAI’s ChatGPT, a generative AI chatbot based on natural language processing (NLP), which made the capabilities of GenAI widely available to everyone, even without coding knowledge.
Trend 1: Using GenAI models will become the standard
According to a Gartner poll, 55% of organizations have already hopped on the GenAI train to discover or use GenAI models in their businesses. It’s not just talk: real investments are being made into changing our way of working. For example, we will most likely see GenAI being increasingly used for simulations, facilitating research, predicting energy demand, drug discovery, self-driving cars, and even space discovery. The expectation is that GenAI applications will be integrated into software more and more. Some examples include Windows Copilot or Neural Filters in Adobe Photoshop.
“GenAI will have a great impact on our way of working. GenAI has the potential to become as big as the introduction of the personal computer or the smartphone. It will take time, but in the years to come it will be an integral part of our daily routine.” Yannick Maltha, CEO of UbiOps
Read more: GenAI: why were 1 million users using OpenAI ChatGPT in its first 5 days
Trend 2: Open source LLMs are catching up
The boom of open source started when Meta released Llama in 2023. A couple of months later, OpenAI responded by releasing the public version of ChatGPT. In December 2023, Mistral released Mixtral, the top-performing open-source LLM according to many benchmarks. The battle for the title of “best model” between open source and closed source is starting to get close. It’s just a matter of time until open-source models catch up with closed-source models that are only accessible through an API, such as ChatGPT (GPT-4) by OpenAI or Claude by Anthropic.
While some people joke that English is the newest coding language, GenAI provides many opportunities for developers, particularly through the use of open-source, off-the-shelf AI models. There’s no need to start from scratch, thanks to the great efforts of the open source community. Pre-trained, open-source models can be downloaded from popular repositories like Hugging Face and can be customized via:
1. fine-tuning – retraining the model on specific data;
2. prompt engineering – guiding the model on how it’s supposed to perform a given task;
3. retrieval augmented generation (RAG) – empowering the model with domain-specific, up-to-date knowledge, and increasing the accuracy and auditability of its responses.
Companies can have different reasons for opting for open source instead of closed source. For some, opting for open-source models provides unparalleled customization and control over their AI applications, allowing fine-tuning for specific enterprise needs. For others, the flexibility of open-source models enables them to avoid the limitations of closed-source models, particularly in accessing and integrating their own data securely. In the long term, we believe that open-source solutions will offer cost-effectiveness and the ability to switch between different models, ensuring adaptability in the rapidly evolving landscape of generative AI.
Developers have already created thousands of customized models based on open-source models like Llama. Our expectation is that in 2024 we will see a great deal more. Right now, creating a new application based on Mistral is easy.
“2024 will be the year where open source is catching up with closed-source LLMs.” Yannick Maltha, CEO of UbiOps
Read more: Deploy your Generative AI and Foundation models with UbiOps
Trend 3: Specialized GenAI models for specific purposes
“Enterprise adoption is driven by the development of data-infused, context-specific LLMs that will be much more compute-efficient to run. Organizations will use several, possibly many domain-specific models next to each other.” Yannick Maltha, CEO of UbiOps.
Manuvir Das, Vice President of Enterprise Computing at NVIDIA, summarizes it by stating that one size doesn’t fit all. Due to the different natures of companies and the different types of data and information that they work with, companies will need to adapt GenAI to their goals and purposes to be most efficient. For Large Language Models (LLMs), this means that the vast majority will be infused with domain-specific data, using information from curated external services with the help of techniques such as RAG.
Although fine-tuning models requires more compute resources (GPUs), inferencing (i.e. running the model in production) is becoming much more efficient in terms of compute. Most open-source models such as Llama 2 and Mixtral can run on a single GPU, which will drive deployment and adoption across all levels of enterprise.
Trend 4: Step aside LLMs, it’s time for LMMs
Large Multimodal Models (LMMs) are deep learning models that work with data of more than one type: text, images, videos, etc. Stable diffusion, for example, is a multimodal model because it takes in a text input and outputs an image (i.e., text-to-image). Humans naturally work with and learn from multimodal data, therefore working with AI that can do the same has clear benefits.
There are already a lot of LMMs out there. You may be familiar with some of the most famous examples such as Midjourney, Adobe Firefly, or even ChatGPT (with a paid account). In fact, last year, a new LMM was released every week.
We believe that while 2023 was the year of the LLM, 2024 will be the year of the LMM. The bar has been raised for what consumers expect the next generation of AI assistants to be capable of. The main difficulty with LMMs – and most likely the first hurdle that organizations will face when trying to deploy LMMs – lies in the data: images, videos, and other unstructured data can be tricky to process and manage. The ability to build data pipelines will be crucial for success.
Trend 5: Collaboration with GenAI
If you have read this far, you may be wondering: was this article written with GenAI? Good question! The answer is no. We indeed asked for the help of our new best friend (sorry, Google, we still love you!) to aid in seeking information, creating structure, and ensuring we didn’t forget any trends – but what you are reading was indeed written by a human. Why? Because humans are still able to detect pretty quickly whether a text was written by an AI, partly because it tends to lack originality.
We instead like to think of using GenAI as a collaboration, much like many people collaborate with their laptops, phones, the internet, and many more technologies. Collaborating with AI helps to automate repetitive tasks, freeing up more time for creativity and innovation within job roles. Think of helpdesk chatbots that can take care of simple questions, or automated systems for more extensive jobs like creating financial statements and prediction models. According to Forrester, the key here is that AI will collaborate with humans, not necessarily replace them as previously thought.
“LLMs are much more than just the chatbots we associate them with. They will become smart agents that will automate many repetitive tasks for us. Don’t see them as a threat, but see them as an assistant or copilot that works alongside you. Another field where GenAI can be particularly helpful is creating new synthetic data to fill in for missing data, which will be used to train new models.” Yannick Maltha, CEO of UbiOps.
Trend 6: Sustainability in AI
Another trend that we are seeing the beginnings of is a concern for sustainability in AI. While chatting with a GenAI chatbot may not feel like pollution, the reality is somewhat different. AI models use up a lot of computing power, which comes from servers. A considerable amount of electricity is required, often from old-fashioned power plants and non-renewable sources. In 2024, we expect to see efforts from businesses and independent initiatives to lower the carbon footprint of their AI practices.
Read more: Bytesnet & UbiOps partnership
The Next Steps to Take with GenAI in 2024
Remember: only 17 years ago, in 2007, the first iPhone was released. Nowadays, it’s the most normal thing in the world to be reachable 24/7, find our way anywhere thanks to Google maps, and so much more. With the fast adoption of GenAI, one can only predict what the future will look like – or perhaps we could just ask a generative AI for the answer. The ongoing shift will continue in 2024 with many innovations and surprises to come. No doubt GenAI will soon become as mainstream as owning a laptop.