The Next Frontier of Large Language Models
Artificial intelligence is moving forward at very fast rates, with big steps forward happening every couple of months. Recent releases from OpenAI, Anthropic, Meta, Cohere, X.AI, and Google give us an exciting look at the cutting edge of large language models (LLMs) and give us a taste of the huge potential these technologies hold.
The fact that the newest models are so smart and can reason is one of the most amazing improvements. On difficult academic tests in subjects ranging from math and science to law and medicine, systems like GPT-4, Claude 3 Opus, and Gemini 1.5 Pro are performing almost as well as humans [1][2][3]. Not even a year or two ago, it would have been impossible for them to have complex conversations, answer follow-up questions, or even admit when they were wrong.
The bigger context windows of these models are another important improvement. This lets them handle and reason over a lot more inputs. Now models like Gemini 1.5 Pro can handle situations with up to 1 million tokens, which is the same as a multivolume book. [3:1] This opens up powerful new uses, such as looking at whole papers, handling long audio and video files, and understanding huge codebases.
The newest LLMs are also pushing the limits of how we understand things in different ways. They can handle more than just text. They can also handle pictures, sounds, videos, and even organized data like PDFs and websites. Models like GPT-4 and Claude 3 have shown that they are very good at visual conceptualization [1:1][2:1]. Now we can make systems that are more natural and easy to use, and they can work with any mode a user chooses.
Of course, having a lot of power means having a lot of stuff to do. As these models get better, it's very important that they are built and used in an honest and responsible way. Leading AI labs are putting a lot of money into testing their models for possible abuse and coming up with ways to reduce safety risks. [1:2][2:2][3:2][4][5][6]. But there is still a lot of work to be done to make sure that LLMs are reliable and in line with human ideals before they can be used in the real world.
It's amazing how quickly LLMs are changing things, and they have the power to make a huge difference. These models could be used to power a new breed of AI helpers. They could also change the way we use computers and help solve some of the world's biggest problems. At the same time, they make us think deeply about the future of work, creation, and even what intelligence is. As we move closer to this exciting new territory, it's up to the AI community to make sure we build useful tools that help people reach their full potential.
Introducing Claude 3, Anthropic, March 4, 2024 ↩︎ ↩︎ ↩︎
Our next-generation model: Gemini 1.5, Google, February 15, 2024 ↩︎ ↩︎ ↩︎
Announcing Grok-1.5, X.AI, March 28, 2024 ↩︎
Build the future of AI with Meta Llama 3, Meta, Retrieved June 2023 ↩︎