While AI trends may appear to be following a similar trajectory of hype and adoption as previous enterprise technology trends such as cloud and machine learning, they are very different in the following ways:
- AI requires large amounts of computing in the process of digesting and recreating unstructured data.
- AI is changing the way some organizations look at organizational structures and careers.
- AI content that can be mistaken for photographs or original artwork is shaking up the art world, and some worry it could be used to influence elections.
Here are our predictions for five AI trends, often referring to generative models, to watch in 2024.
AI adoption is moving closer and closer to integration with existing applications
Many generative AI use cases brought to market for enterprises and businesses are integrated with existing applications rather than creating entirely new use cases. The most high-profile example of this is the proliferation of copilots, which refer to generative AI assistants. Microsoft has installed his Copilots next to his 365 suite of products, and SoftServe and many other companies offer Copilots for industrial work and maintenance. Google offers a variety of co-pilots for everything from video creation to security.
But all of these co-pilots are designed to cull existing content or create content that's closer to what humans write at work.
See also: Which is better for the job: Google Gemini or ChatGPT? (Tech Republic)
Even IBM called for a reality check on trending technologies, pointing out that tools like Google's 2018 Smart Compose are technically “generative” but aren't considered to change the way we work. did. The key difference between Smart Compose and modern generative AI is that some AI models today are multimodal. This means you can create and interpret photos, videos, and graphs.
“I would argue that in 2024, we're going to see a lot of innovation around it (multimodality),” Arun Chandrasekaran, distinguished vice president and analyst at Gartner, said in an interview with TechRepublic. Told.
At NVIDIA GTC 2024, many startups on the show floor ran chatbots on Mistral AI's large-scale language models, as open models allow you to create custom-trained AIs with access to enterprise data. With proprietary training data, AI can answer questions about a specific product, industrial process, or customer service without feeding a company's proprietary information into a trained model and exposing that data to the public internet. can do. There are many other open models for text and video, including Meta's Llama 2, Stability AI's model suite including Stable LM and Stable Diffusion, and Abu Dhabi's Technology Innovation Institute's Falcon family.
“There is a strong interest in bringing enterprise data into LLM as a way to build on the model and add context,” Chandrasekaran said.
Customization of the open model can be done in several ways, including prompt engineering, search extension generation, and fine-tuning.
AI agent
Another way AI could become more integrated with existing applications in 2024 is through AI agents, which Chandrasekaran called a “watershed moment” in the advancement of AI.
AI agents automate tasks for other AI bots. This means that the user does not need to specifically instruct individual models. Instead, he can provide the agent with instructions in one natural language. This essentially forces the team to work together on the various commands needed to carry out the instructions.
Sachin Katti, Intel senior vice president and general manager of the network and edge group, also talked about AI agents and how AI can work with each other during a briefing ahead of the Intel Vision conference, April 9-11. He suggested that delegating could help accomplish department-wide tasks. .
Search Augmented Generation Dominates Enterprise AI
Search extension generation allows LLM to check responses against external sources before providing them. For example, the AI might match the answer to a technical manual and provide the user with a footnote with a direct link to the manual. RAG aims to increase accuracy and reduce hallucinations.
RAG offers organizations a way to improve the accuracy of their AI models without increasing their bills. RAG produces more accurate results and speeds engineering and fine-tuning compared to other common methods of adding enterprise data to LLM. This will be a hot topic for him in 2024 and could continue to be a hot topic later this year.
Organizations express quiet concerns about sustainability
AI is used to create climate and weather models that predict catastrophic events. At the same time, generative AI is energy and resource intensive compared to traditional computing.
What does this mean for AI trends? Optimistically, being aware of energy-intensive processes will help companies run them and right-size their usage. will create more efficient hardware. On a less optimistic note, generative AI workloads are likely to continue to consume large amounts of power and water. Either way, generative AI could become an issue that contributes to national debates about energy use and grid resilience. Currently, AI regulations are primarily focused on use cases, but in the future, AI energy use may also fall under specific regulations.
Tech giants are tackling sustainability in their own ways, such as Google purchasing solar and wind energy in specific regions. For example, NVIDIA touted its ability to save energy in data centers while running AI by using fewer server racks with more powerful GPUs.
AI data center and chip energy usage
The 100,000 AI servers that NVIDIA plans to send to customers this year could generate between 5.7 and 8.9 TWh of electricity per year, a fraction of the power currently used in data centers. This is according to a paper by PhD candidate Alex de Vries published in October 2023. However, as the paper speculates, if NVIDIA alone adds 1.5 million AI servers to the grid by 2027, the servers will use 85.4 to 134.0 TWh per year. A much more serious impact.
Another study found that creating 1,000 images with Stable Diffusion XL produces about the same amount of carbon dioxide as driving 6.1 miles in an average gas-powered car.
“We found that multi-objective generative architectures are orders of magnitude more expensive than task-specific systems for a variety of tasks, even when controlling the number of model parameters,” says researcher Alexandra Sasha Luccioni writes Yasin Jarnait of Hugging. Faith Strubel and Emma Strubel of Carnegie Mellon University;
In the journal Nature, Microsoft AI researcher Kate Crawford noted that GPT-4 training used about 6% of the local district's water.
The changing role of the AI specialist
Rapid engineering will be one of the hottest skill sets in the tech industry in 2023, with people rushing for six-figure salaries to instruct ChatGPT and similar products to generate useful responses. I was about to take it home. The hype has faded somewhat, and as mentioned above, many companies that use generative AI heavily are customizing their own models. In the future, agile engineering may become part of a software engineer's regular job, but not as a specialty, but simply as one of the ways software engineers do their regular work.
Utilizing AI in software engineering
“The use of AI in software engineering is one of the fastest growing use cases we see today,” Chandrasekaran said. “Nimble engineering is an organizational imperative in the sense that anyone who interacts with AI systems (which will be many of us in the future) needs to know how to guide and steer these models. I believe it's going to be an important skill across the board. But of course, software engineering people need to really understand prompt engineering at scale and some of the advanced techniques of prompt engineering.”
How the role of AI is assigned varies greatly between individual organizations. It remains to be seen whether most people working in Prompt Engineering will have Prompt Engineering as a job title.
Executive positions related to AI
A January 2024 survey of data and technology executives by MIT's Sloan Management Review found that organizations are sometimes cutting back on the number of chief AI officers. While we've seen “confusion about responsibilities” for highly specialized leaders like AI and data chiefs, in 2024 we'll be able to create value from data, regardless of where it comes from, and There is a possibility that things will become normalized around a “general technology leader” who reports to. from.
See: The role of an AI chief and why your organization needs one in the future. (Tech Republic)
Meanwhile, Chandrasekaran said that while chief data and analytics officers and chief AI officers are “less prevalent,” their numbers are growing. It is difficult to predict whether these two will remain separate roles from CIO or CTO, but it depends on what core competencies the organization is looking for and how CIOs hold many other responsibilities at the same time. It may depend on whether you feel it's too much.
“There is no question that these roles (AI officer, data and analytics officer) are coming up more and more in conversations with customers,” Chandrasekaran said.
On March 28, 2024, the U.S. Office of Management and Budget released guidance on the use of AI within federal agencies. This includes a requirement for all federal agencies to appoint a chief AI officer.
Both AI art and glazing on AI art will become more common.
As art software and stock photo platforms embrace the gold rush of easy images, artists and regulators are looking for ways to identify AI content to avoid misinformation and theft.
AI art is becoming more common
Adobe Stock now provides tools to create AI art and mark catalogs of stock images as AI art. On March 18, 2024, Shutterstock and his NVIDIA announced his 3D image generation tool in early access.
OpenAI recently promoted filmmakers using its photorealistic Sora AI. The demo was criticized by artist advocates, including Fairly Training AI CEO Ed Newton Rex, formerly of Stability AI. They called them “artist washing.”: When you train on people's work without permission or payment, while soliciting positive comments from a small number of creators about the AI models you generate. ”
Two possible responses to AI artwork could be further developed into 2024. Watermarks and glazing.
Watermark AI art
The leading standard for watermarking is OpenAI (Diagram A) and Meta collaborated to tag images generated by AI. However, watermarks that appear visually or in metadata can be easily removed. Some say the watermark doesn't go far enough in preventing misinformation, especially before and after the 2024 U.S. election.
Diagram A
SEE: The U.S. federal government and major AI companies agreed last year to a list of voluntary initiatives that include watermarking. (Tech Republic)
Poisoning original art to AI
Artists who want to prevent their AI models from being trained on original art posted online can use Glaze or Nightshade, two data poisoning tools created by the University of Chicago. Data poisoning adjusts the artwork to the point that it becomes unreadable to the AI model. We're likely to see more tools like this in the future, as both AI image generation and protecting artists' original works will continue to be a focus in 2024.
Is AI overhyped?
AI was so popular in 2023 that it was inevitably overhyped heading into 2024, but that doesn't mean it's not in use. In late 2023, Gartner declared that generative AI had reached “peak hype.” This is the height of the hype that emerging technologies are known for before they are commercialized and standardized. The peak is followed by a “trough of disillusionment,'' followed by a rise to a “slope of enlightenment'' and, ultimately, increased productivity. Perhaps the peak or trough of generative AI means it is overhyped. But many other products have gone through hype cycles before, and many eventually reach a “productivity plateau” after an initial boom.
