SAN FRANCISCO — At an event in San Francisco in November, Sam Altman, CEO of artificial intelligence company OpenAI, was asked what surprises the field would bring in 2024.
Online chatbots such as OpenAI’s ChatGPT will take “a leap forward that no one expected,” Altman immediately responded.
Sitting beside him, James Manyika, a Google executive, nodded and said, “Plus one to that.”
The AI industry this year is set to be defined by one main characteristic: a remarkably rapid improvement of the technology as advancements build upon one another, enabling AI to generate new kinds of media, mimic human reasoning in new ways, and seep into the physical world through a new breed of robot.
In the coming months, AI-powered image generators such as DALL-E and Midjourney will instantly deliver videos as well as still images. And they will gradually merge with chatbots such as ChatGPT.
That means chatbots will expand well beyond digital text by handling photos, videos, diagrams, charts, and other media. They will exhibit behavior that looks more like human reasoning, tackling increasingly complex tasks in fields such as math and science. As technology moves into robots, it will also help to solve problems beyond the digital world.
Many of these developments have already started emerging inside the top research labs and in tech products. But in 2024, the power of these products will grow significantly and be used by far more people.
“The rapid progress of AI will continue,” said David Luan, CEO of Adept, an AI startup. “It is inevitable.”
OpenAI, Google, and other tech companies are advancing AI far more quickly than other technologies because of the way the underlying systems are built.
Most software apps are built by engineers, one line of computer code at a time, which is typically a slow and tedious process. Companies are improving AI more swiftly because the technology relies on neural networks, and mathematical systems that can learn skills by analyzing digital data. By pinpointing patterns in data such as Wikipedia articles, books, and digital text culled from the internet, a neural network can learn to generate text on its own.
This year, tech companies plan to feed AI systems more data — including images, sounds, and more text — than people can wrap their heads around. As these systems learn the relationships between these various kinds of data, they will learn to solve increasingly complex problems, preparing them for life in the physical world.
(The New York Times sued OpenAI and Microsoft last month for copyright infringement of news content related to AI systems.)
None of this means that AI will be able to match the human brain anytime soon. While AI companies and entrepreneurs aim to create what they call “artificial general intelligence” — a machine that can do anything the human brain can do — this remains a daunting task. For all its rapid gains, AI remains in the early stages.
Here’s a guide to how AI is set to change this year, beginning with the nearest-term advancements, which will lead to further progress in its abilities.
Until now, AI-powered applications mostly generated text and still images in response to prompts. DALL-E, for instance, can create photorealistic images within seconds of requests such as “a rhino diving off the Golden Gate Bridge.”
But this year, companies such as OpenAI, Google, Meta, and New York-based Runway are likely to deploy image generators that allow people to generate videos, too. These companies have already built prototypes of tools that can instantly create videos from short text prompts.
Tech companies are likely to fold the powers of image and video generators into chatbots, making the chatbots more powerful.
Chatbots and image generators, originally developed as separate tools, are gradually merging. When OpenAI debuted a new version of ChatGPT last year, the chatbot could generate images as well as text.
AI companies are building “multimodal” systems, meaning the AI can handle multiple types of media. These systems learn skills by analyzing photos, text and potentially other kinds of media, including diagrams, charts, sounds, and video, so they can then produce their text, images, and sounds.
That isn’t all. Because the systems are also learning the relationships between different types of media, they will be able to understand one type of media and respond with another. In other words, someone may feed an image into a chatbot and it will respond with text.
“The technology will get smarter, more useful,” said Ahmad Al-Dahle, who leads Meta’s generative AI group. “It will do more things.”
Multimodal chatbots will get stuff wrong, just as text-only chatbots make mistakes. Tech companies are working to reduce errors as they strive to build chatbots that can reason like a human.
When Altman talks about AI’s leaping forward, he is referring to chatbots that are better at “reasoning” so they can take on more complex tasks, such as solving complicated math problems and generating detailed computer programs.
The aim is to build systems that can carefully and logically solve a problem through a series of discrete steps, each one building on the next. That is how humans reason, at least in some cases.
Leading scientists disagree on whether chatbots can truly reason like that. Some argue that these systems merely seem to reason as they repeat behavior they have seen in internet data. But OpenAI and others are building systems that can more reliably answer complex questions involving subjects such as math, computer programming, physics and other sciences.
“As systems become more reliable, they will become more popular,” said Nick Frosst, a former Google researcher who helps lead Cohere, an AI startup.
If chatbots are better at reasoning, they can then turn into “AI agents.”
As companies teach AI systems how to work through complex problems one step at a time, they can also improve the ability of chatbots to use software apps and websites on their behalf.
Researchers are essentially transforming chatbots into a new kind of autonomous system called an AI agent. That means the chatbots can use software apps, websites, and other online tools, including spreadsheets, online calendars, and travel sites. People could then offload tedious office work to chatbots. But these agents could also take away jobs entirely.
Chatbots already operate as agents in small ways. They can schedule meetings, edit files, analyze data, and build bar charts. However, these tools do not always work as well as they need to. Agents break down entirely when applied to more complex tasks.
This year, AI companies are set to unveil more reliable agents. “You should be able to delegate any tedious, day-to-day computer work to an agent,” Luan said.
This might include keeping track of expenses in an app such as QuickBooks or logging vacation days in an app such as Workday. In the long run, it will extend beyond software and internet services and into the world of robotics.
In the past, robots were programmed to perform the same task over and over again, such as picking up boxes that are always the same size and shape. But using the same kind of technology that underpins chatbots, researchers are giving robots the power to handle more complex tasks — including those they have never seen before.
Just as chatbots can learn to predict the next word in a sentence by analyzing vast amounts of digital text, a robot can learn to predict what will happen in the physical world by analyzing countless videos of objects being prodded, lifted, and moved.
“These technologies can absorb tremendous amounts of data. And as they absorb data, they can learn how the world works, how physics work, how you interact with objects,” said Peter Chen, a former OpenAI researcher who runs Covariant, a robotics startup.
This year, AI will supercharge robots that operate behind the scenes, such as mechanical arms that fold shirts at a laundromat or sort piles of stuff inside a warehouse. Tech titans such as Elon Musk are also working to move humanoid robots into people’s homes.
This article originally appeared in The New York Times.