Image Credit: STEPHANIE ARNETT/MITTR | EYEEM
About this time last year, we acted irresponsibly. We dabbled in future forecasting in a field where nothing stays the same.
Our four main predictions for 2023 were as follows:
✔ Multimodal chatbots would be the wave of the future.
✔ Lawmakers would draft stringent new regulations.
✔ Open-source startups would put pressure on big tech.
✔ AI would fundamentally alter the pharmaceutical industry.
It would be too soon to tell whether this would come to pass.
The open-source boom is still going strong, but the first drugs created using AI are still a few years away from the market.
And here we go again.
The apparent was disregarded. Big language models will be in the lead for a while. Over time, regulators will become more daring. Researchers, regulators, and the general public will continue focusing on AI issues, such as prejudice, copyright, and doomers, even after 2024. This article delves more into our four major concerns regarding generative AI.
We’ve narrowed it down to a handful of trends. In 2024, keep an eye out for these things. (Return next year to assess our progress.)
1. Localized Chatbots
Tech companies that have invested much money into generative AI will have to show that their products can make money by 2024. Google and OpenAI, two of the most prominent names in artificial intelligence, are taking a modest approach to this problem. They are working on systems anyone can use to build little chatbots tailored to their needs without coding knowledge. Both have released online resources that everyone may use to construct generative AI apps.
By 2024, generative AI may have practical applications for the average, non-technical individual, and the number of individuals experimenting with various AI models is expected to rise. GPT-4 and Gemini are examples of state-of-the-art AI models that are multimodal, meaning they can process images, videos, and text. Lots of new apps might become available with these new capabilities. To illustrate the point, a realty agent may easily ask the individualized AI to write a property description by uploading photographs and videos of fresh listings, fine-tuning a robust model to produce comparable content at the push of a button, and uploading text from past listings.
2. Videos will be the second wave of generative AI.
It’s incredible how quickly the extraordinary becomes the ordinary. In 2022, generative models that could generate photorealistic images first appeared and quickly became ubiquitous. A deluge of breathtaking photos of the pope in Balenciaga and award-winning artwork filled the internet thanks to tools like Adobe’s Firefly, Stability AI’s Stable Diffusion, and OpenAI’s DALL-E. There is sexist sexual stereotyping and imitation fantasy art for every pug with pompoms, so it’s not all rainbows and unicorns.
Transforming text into video is the next big thing. In other words, it will likely amplify text-to-image’s flaws and shortcomings.
Just over a year ago, we first tasted the capabilities of generative models when we trained them to composite several still photos into short videos. The output was shaky and distorted. But technology has advanced at a tremendous pace.
Every few months, Runway, a startup that co-created Stable Diffusion and develops generative video models, releases new versions of its tools. Even its most recent model, Gen-2, produces incredibly high-quality video only a few seconds long. The top clips are very similar to what Pixar could release.
3. Controversial election misinformation spread by AI will be pervasive.
A record number of people will go to the polls in 2024, and if the last election is any indication, AI-generated election disinformation and deepfakes will be a significant issue. Politicians are already turning these instruments into weapons. Two Argentine presidential contenders attacked their opponents with AI-generated photos and videos.
During the Slovak elections, a deepfake featuring a liberal pro-European party leader joking about child pornography and threatening to increase beer prices went viral. Moreover, in the United States, Donald Trump has supported an organization that uses artificial intelligence to create memes that contain racist and sexist clichés.
4. Multitasking Bots
Roboticists are beginning to develop more versatile, general-purpose robots by studying the fundamental techniques underlying the present surge in generative AI.
Recently, artificial intelligence has moved away from utilizing numerous separate models, each trained to perform a particular task (such as picture identification, sketching, or captioning), and toward employing a single, unified model that can do all these things and more. Researchers can train OpenAI’s GPT-3 to handle coding difficulties, write movie scripts, pass high school biology exams, and more by showing it a few new examples, a process known as fine-tuning. Visual and verbal problems are solved by multimodal models such as GPT-4 and Gemini from Google DeepMind.
A one-size-fits-all paradigm might allow robots to multitask, eliminating the need to train separate bots for tasks such as opening doors and flipping pancakes. In 2023, several works that debuted in this domain were published.
Updated from last year’s Gato, DeepMind’s Robocat came out in June. It learns to handle many robot arms (rather than just one, as is usual) using its data generated via trial and error.
With the help of 33 academic labs, the business released RT-X, a new general-purpose model for robotics, and a massive new general-purpose training dataset in October. Comparatively, several prestigious educational groups are investigating related technologies, for example, Berkeley’s RAIL (Robotic Artificial Intelligence and Learning) group.
More data is needed. Generative AI uses image and text data collection. On the other hand, there are few reliable data sources for robots to learn how to carry out the kinds of industrial and household chores humans want.
An NYU group headed by Lerrel Pinto is working on a solution to this problem. He and his coworkers work on methods enabling robots to learn autonomously, creating their training data. Even more covertly, Pinto has enlisted the help of concerned citizens to film their neighborhoods using an iPhone camera attached to a garbage picker. In the past few years, major corporations like Meta have begun distributing massive data sets for robot training. One such set is Ego4D.
Already, this method is demonstrating potential in autonomous vehicles. Wayve, Waabi, and Ghost are leading the charge of a new generation of autonomous driving AI that consolidates vehicle control into a single big model instead of several smaller ones. Small businesses have been catching up to industry heavyweights like Waymo and Cruise. London’s tight, congested streets are now the site of Wayve’s autonomous vehicle testing. A comparable boost is about to be bestowed upon robots worldwide.
