After ChatGPT’s debut in November 2022, 2023 emerged as a pivotal year for artificial intelligence. The breakthroughs of the last year, featuring a dynamic open-source environment and advanced multimodal models, have established a foundation for noteworthy progress in AI.
However, even as generative AI continues to enchant the technology sector, perspectives are evolving to become more balanced and mature. Organizations are shifting their emphasis from mere experimentation to practical applications. The Top AI trends observed this year indicate a growing level of sophistication and caution in how AI is developed and implemented, with considerations for ethics, safety, and the changing regulatory framework.
Here are the top 10 AI Trends and machine learning to watch for in 2024.
1. Multimodal AI
Multimodal AI surpasses conventional single-mode data processing by integrating multiple types of inputs, including text, images, and audio—moving closer to replicating the human capability to interpret a variety of sensory information.
“The interfaces of the world are multimodal,” stated Mark Chen, head of frontiers research at OpenAI, during a November 2023 talk at the EmTech MIT conference. “We desire our models to perceive what we see and hear what we hear, and we want them to generate content that engages more than one of our senses.”
OpenAI’s GPT-4 model’s multimodal features let the program to react to both audio and visual input. During his presentation, Chen used the example of snapping pictures of the interior of a refrigerator and asked ChatGPT to recommend a dish using the components shown. If the request is made aloud using ChatGPT’s voice mode, the exchange may even include an audio component.
“The real power of these capabilities is going to be when you can marry up text and conversation with images and video, cross-pollinate all three of those, and apply those to a variety of businesses,” stated Matt Barrington, Americas emerging technologies leader at EY, despite the fact that the majority of generative AI trends initiatives today are text-based.
The practical uses of multimodal AI can be numerous and growing. For instance, multimodal models can evaluate medical images in the context of genetic data and patient history to increase the precision of diagnosis in the healthcare industry. By providing fundamental design and coding skills to people without a formal education in those areas, multimodal models can broaden the range of work that different employees can perform at the job function level.
Barrington remarked, “I can’t draw to save my life,” “All right, I can now. I can use a tool like [picture creation] because I can speak fairly well, and I can use AI to realize some of my concepts that I could never physically sketch.”
In addition, by providing models with fresh data to work with, multimodal capabilities could improve them. “As our models get better and better at modeling language and start to hit the limits of what they can learn from language, we want to provide the models with raw inputs from the world so that they can perceive the world on their own and draw their own inferences from things like video or audio data,” Chen commented.
2. Agentic AI
Reactive AI is giving way to proactive AI with agentic AI. AI agents are sophisticated systems with the capacity for independent action, proactivity, and autonomy. AI agents, in contrast to traditional AI systems, are made to comprehend their surroundings, set goals, and take action to accomplish those goals without direct human interaction. Traditional AI systems primarily react to user inputs and follow preset programming.
For instance, an AI agent might be trained to gather information, examine AI trends, and take preventative measures in response to dangers like early warning indications of a forest fire in environmental monitoring. Similarly, an AI financial agent may actively manage a portfolio of investments by employing adaptive algorithms that respond instantly to shifting market conditions.
“2023 was the year of being able to chat with an AI,” wrote Peter Norvig, a fellow at Stanford’s Human-Centered AI Institute and a computer scientist, yesterday. “The capacity for agents to complete tasks for you will be available in 2024. Plan a trip, make reservations, and connect to additional services.
Furthermore, integrating multimodal and agentic AI may create new opportunities. Chen provided an example of an application intended to determine the contents of an uploaded image in the presentation cited above. It used to be necessary for someone wishing to create such an application to train their own picture recognition model and then figure out how to implement it. However, all of this might be achieved by natural language prompting in multimodal, agentic models.
“I firmly believe that multimodal in conjunction with GPTs will enable the development of computer vision applications without the need for code, much like prompting enabled the development of numerous text-based applications without the need for code,” Chen stated.
3. Open source AI
The process of creating big language models and other powerful generative AI systems is costly and necessitates vast amounts of data and computation. However, by allowing developers to build upon the work of others, an open source model lowers costs and increases access to AI. Organizations and researchers can contribute to and expand upon existing code thanks to open source AI, which is generally available to the public for free.
Developer involvement with AI, especially generative AI, has significantly increased over the past year, according to GitHub data. With initiatives like Stable Diffusion and AutoGPT generating hundreds of new contributors, generative AI projects made an appearance in the top 10 most popular projects on the code hosting platform in 2023.
There weren’t many open source generative models available at the beginning of the year, and they frequently performed worse than proprietary alternatives like ChatGPT. However, around 2023, the field grew considerably to include strong open source competitors including Mistral AI’s Mixtral models and Meta’s Llama 2. By giving smaller, less resourced organizations access to advanced AI models and technologies that were previously unattainable, this could change the dynamics of the AI trends in 2024.
“It gives everyone easy, fairly democratized access, and it’s great for experimentation and exploration,” said Barrington.
Because more people are looking at the code, there is a higher chance of finding biases, errors, and security flaws, hence open source approaches can help promote transparency and ethical development. However, analysts are also worried about open source AI being abused to produce negative content, including misinformation. Furthermore, even for conventional software, it can be challenging to develop and maintain open source, let alone sophisticated and computationally demanding AI models.
4. Retrieval-augmented generation
Hallucinations, or believable-sounding but inaccurate answers to users’ questions, are still a concern with generative AI technologies, despite their widespread adoption in 2023. Due to the potential for disastrous hallucinations in situations involving customers or business criticality, this shortcoming has been a barrier to enterprise adoption. A method for lessening hallucinations called retrieval-augmented generation (RAG) has gained popularity and could have a significant impact on the adoption of AI trends in businesses.
RAG improves the precision and applicability of AI-generated content by combining text creation and information retrieval. It gives LLMs access to outside data, which aids in the production of more precise and contextually sensitive answers. Additionally, avoiding the necessity to directly storing all knowledge in the LLM results in a smaller model, which boosts efficiency and cuts expenses.
“You can use RAG to go gather a ton of unstructured information, documents, etc., [and] feed it into a model without having to fine-tune or custom-train a model,” added Barrington.
These advantages are especially alluring for enterprise applications where current factual knowledge is essential. For instance, companies can develop more effective and educational chatbots and virtual assistants by combining RAG with foundation models.
5. Customized enterprise generative AI models
The most popular tools among consumers investigating generative AI are large, all-purpose ones like ChatGPT and Midjourney. The increasing need for AI systems that can satisfy specialized needs, however, may make smaller, more focused models the most resilient for corporate use cases.
Although starting from scratch with a new model is an option, it is a resource-intensive process that many firms will not be able to afford. Most firms instead alter pre-existing AI models to create customized generative AI, such as by adjusting their architecture or fine-tuning on a domain-specific data set. This may be less expensive than using API calls to a public LLM or creating a new model from scratch.
As an example, according to Shane Luke, vice president of AI and machine learning at Workday, “calls to GPT-4 as an API are very expensive, both in terms of cost and in terms of latency — how long it can actually take to return a result.” “We’re working hard to optimize so that we can do the same thing, but it’s really focused and precise. As a result, it may be a much more manageable and smaller model.
The ability of customizable generative AI models to meet user needs and niche markets is its main advantage. From document review to supply chain management to customer service, customized generative AI solutions can be developed for nearly any situation. This is particularly important for industries with high specialized terminology and practices, such as healthcare, finance and legal.
The largest LLMs are necessary in many corporate use cases. According to Luke, “it’s not the state of the art for smaller enterprise applications,” even though ChatGPT may be the best available chatbot for consumer-facing applications that can answer any question.
Barrington anticipates that as the capabilities of AI developers start to converge, businesses will investigate a wider variety of models in the upcoming year. “We’re expecting, over the next year or two, for there to be a much higher degree of parity across the models — and that’s a good thing,” he stated.
At Workday, which offers a suite of AI services for teams to test internally, Luke has witnessed a similar situation on a lesser scale. According to Luke, he has observed a gradual trend toward a mix of models from several suppliers, including Google and AWS, even though staff initially mostly used OpenAI services.
Because it provides companies more control over their data, creating a bespoke model rather than using a publicly available off-the-shelf technology frequently increases privacy and security. Luke used creating a model for Workday tasks that handle private information like medical records and disability status as an example. “Those aren’t things that we’re going to want to send out to a third party,” he explained. “In general, our clients wouldn’t be comfortable with that.”
Having these privacy and security advantages, more stringent AI regulations in the upcoming years may force businesses to concentrate on proprietary models, according to Gillian Crossan, a risk advisory principal and worldwide leader in the technology sector at Deloitte.
“It’s going to encourage enterprises to focus more on private models that are proprietary, that are domain-specific, rather than focus on these large language models that are trained with data from all over the internet and everything that that brings with it,” she added.
6. Need for AI and machine learning talent
It’s not easy to design, train, and test a machine learning model, let alone deploy it to production and keep it running in a complicated corporate IT environment. Therefore, it should come as no surprise that there will be a continued need for talent in AI and machine learning until 2024 and beyond.
“The market is still really hot around talent,” Luke stated. “It’s very easy to get a job in this space.”
Professionals that can bridge the gap between theory and reality are becoming increasingly needed, especially when AI and machine learning are incorporated into commercial operations. This calls for the ability to implement, track, and manage AI systems in practical environments; this field is commonly known as machine learning operations, or MLOps.
According to a recent O’Reilly survey, the top three competencies that their companies need for generative AI projects were AI programming, data analysis and statistics, and operations for AI and machine learning. But there aren’t many people with these kinds of skills. “That’s going to be one of the challenges around AI — to be able to have the talent readily available,” Crossan stated.
Look for enterprises, not just large tech firms, to hire people with these kinds of abilities in 2024. Building internal AI and machine learning capabilities is set to be the next step in digital transformation, as AI efforts gain popularity and IT and data become almost ubiquitous as business operations.
Additionally, Crossan stressed the value of diversity in AI projects at all levels, from the board to the technical teams creating models. “One of the big issues with AI and the public models is the amount of bias that exists in the training data,” she stated. “And unless you have that diverse team within your organization that is challenging the results and challenging what you see, you are going to potentially end up in a worse place than you were before AI.”
7. Shadow AI
As workers in a variety of job roles grow more interested in generative AI, companies are dealing with the problem of shadow AI, which is the application of AI inside a company without the IT department’s express consent or supervision. This approach is growing in popularity as AI becomes more widely available, allowing even nontechnical professionals to use it on their own.
When workers desire to experiment with new technology more quickly than authorized channels permit or require quick fixes for an issue, shadow AI usually emerges. Employees can easily test out AI chatbots in their web browsers without undergoing IT review and approval procedures, which is particularly prevalent for simple AI chatbots.
On the plus side, investigating applications for these new technology shows initiative and creativity. However, there is also a risk because end users frequently don’t have the necessary knowledge about security, data privacy, and compliance. For instance, a user may enter trade secrets into an LLM that is visible to the public without being aware that doing so would reveal that private information to outside parties.
“Once something gets out into these public models, you cannot pull it back,” Barrington stated. “So there’s a bit of a fear factor and risk angle that’s appropriate for most enterprises, regardless of sector, to think through.”
By 2024, companies will have to manage shadow AI with governance frameworks that strike a compromise between fostering innovation and safeguarding security and privacy. In order to understand how different departments wish to use AI, this might involve establishing clear guidelines for appropriate AI use, offering platforms that have been approved, and promoting cooperation between IT and business executives.
“The reality is, everybody’s using it,” Barrington remarked, referring to a recent study by EY that found 90% of participants employed AI in their jobs. “Whether you like it or not, your people are using it today, so you should figure out how to align them to ethical and responsible use of it.”
8. A generative AI reality check
In 2024—a stage known as the “trough of disillusionment” in the Gartner Hype Cycle—organizations are likely to experience a reality check as they move past the early enthusiasm surrounding generative AI to actual adoption and integration.
“We’re definitely seeing a rapid shift from what we’ve been calling this experimentation phase into [asking], ‘How do I run this at scale across my enterprise?'”, Barrington said.
Organizations are facing generative AI’s drawbacks as early excitement starts to fade, including issues with output quality, security and ethics issues, and integration challenges with current workflows and systems. Tasks like guaranteeing data quality, training models, and maintaining AI systems in production might be more difficult than first thought. The complexity of deploying and growing AI in a corporate setting is frequently underestimated.
“It’s actually not very easy to build a generative AI application and put it into production in a real product setting,” Luke said.
The bright side is that, although painful in the near term, these growing pains may eventually lead to a more balanced and healthy perspective. Setting reasonable goals for AI and gaining a more sophisticated knowledge of its capabilities will be necessary to get past this stage. AI initiatives should have a clear framework in place for assessing results and be closely linked to business objectives and real-world use cases.
“If you have very loose use cases that are not clearly defined, that’s probably what’s going to hold you up the most,” Crossan said.
9. Increased attention to AI ethics and security risks
The growing number of deepfakes and advanced AI-generated material is causing concerns about the possibility of identity theft, other forms of fraud, and misinformation and manipulation in politics and the media. AI can also make ransomware and phishing attempts more effective by making them more realistic, flexible, and difficult to identify.
Technologies for identifying AI-generated content are being developed, although this is still a difficult task. Existing AI watermarking methods are rather simple to get around, and AI detection tools can produce false positives.
The growing prevalence of AI systems also emphasizes how crucial it is to make sure they are open and equitable, for instance, by thoroughly screening algorithms and training data for bias. Crossan underlined that when creating an AI strategy, these compliance and ethics issues should be integrated throughout.
“You have to be thinking about, as an enterprise … implementing AI, what are the controls that you’re going to need?” “I said,” she said. And that begins to assist you in making some plans for the regulation so that you can work together on it. ‘Oh, now we need to worry about the controls.’ isn’t what you’re doing after all this AI experimentation. You do it simultaneously.
Luke noted that smaller, more precisely customized models can also be considered for ethical and safety reasons. “These smaller, tuned, domain-specific models are just far less capable than the really big ones — and we want that,” he stated. “They’re less likely to be able to output something that you don’t want because they’re just not capable of as many things.”
10. Evolving AI regulation
Unsurprisingly, with laws, rules, and business frameworks fast changing both domestically and internationally, 2024 is looking to be a crucial year for AI regulation given these ethics and security concerns. Global operations and AI development strategies may be significantly impacted by changing compliance regulations in the upcoming year, so organizations will need to remain knowledgeable and flexible.
Members of the EU’s Parliament and Council have secured a tentative agreement on the EU’s AI Act, which is the first comprehensive AI law in the world. Adopting it would oblige corporations utilizing generative AI to be transparent, prohibit specific applications of AI, and put duties on developers of high-risk AI systems. Noncompliance might result in fines of millions of dollars. Furthermore, new legislation alone might not have an effects in 2024.
“Interestingly enough, the regulatory issue that I see could have the biggest impact is GDPR — good old-fashioned GDPR — because of the need for rectification and erasure, the right to be forgotten, with public large language models,” said Crossan. “How do you control that when they’re learning from massive amounts of data, and how can you assure that you’ve been forgotten?”
The AI Act, when combined with the GDPR, might establish the EU as a global regulator of AI, potentially impacting AI development and use standards across the globe. “They’re certainly ahead of where we are in the U.S. from an AI regulatory perspective,” said Crossan.
Although the United States does not currently have complete federal legislation like to the AI Act of the European Union, experts advise firms to consider compliance before official rules are established. “For instance, at EY, “we’re interacting with our clients to get ahead of it,” Barrington wrote. Businesses might have to catch up when restrictions do take effect if they don’t.
Conclusion
The fact that 2024 is an election year in the United States and that the current presidential candidates represent a diverse spectrum of views on tech policy issues further complicates matters. Theoretically, a subsequent administration might reverse or amend Biden’s executive order and nonbinding agency recommendations, changing the executive branch’s approach to AI oversight.