“Next-generation AI will be able to invent breakthrough technologies and cure diseases”
The next generation of AI systems will be able to perform tasks autonomously without human intervention, according to the company's CEO Sam Altman, and this is made possible by models like OpenAI's new o1 model.
Disclaimer 1: This is a free translation of Ryan Morrison's column, which he wrote for the Tom's Guide portal. The translation was prepared by the editors of Technocracy. To stay up to date with new materials, subscribe to “Voice of Technocracy” — we regularly talk about news about AI, LLM and RAG, and also share useful mustreads and current events.
Speaking at T-Mobile Capital Markets Day, Altman praised the o1 models and their ability to “reason,” which he says will open up entirely new possibilities for AI that were previously unavailable with previous GPT-class models.
Altman believes that these reasoning models, which can work through a problem before providing a solution, will enable the development of level-3 AI, which he describes as agent systems.
Agent systems are when ChatGPT can effectively act on its own to provide the best possible response to the user, including performing tasks in other services. This will lead to the fourth level – systems capable of innovation.
What is the main change in AI?
During our conversation, Altman acknowledged the strengths of the current batch of GPT models, including GPT-4, which underpins ChatGPT and Advanced Voice. These models are multimodal by design, but operate like any previous AI, token by token.
“The GPT series of models were great at System-1 thinking, but we needed systems that could reason. If an AI could reason about problems, that would be of enormous value. o1 is the first system that can reason about complex problems, and if you give it a complex problem, you get amazing results.” — Sam Altman
Here, “System 1” refers to fast, intuitive, and automatic cognitive processes. “System 2,” which is what o1 models aim for, involves more deliberate, logical thinking, albeit slower. These systems were described by Daniel Kahneman, one of the founders of behavioral economics and a Nobel laureate.
Altman emphasized the significance of this progress: “This approach will eventually look as significant as the output of GPT. Imagine that the current models that can reason are at the level of GPT-2, but very soon they will reach the level of GPT-4.”
In other words, models with logical reasoning are currently in their early stages of development: o1-preview roughly corresponds to GPT-2, which preceded the launch of ChatGPT in November 2022.
Despite the early stage, he expressed confidence in rapid progress: “But you will see updates in the coming months as we move from o1-preview to o1. The improvement curve is very steep, and what the models cannot solve today, they will be able to solve in a few months.”
Altman also noted the potential for new and innovative applications: “We’re going to see completely new ways to use these models… We’re so early in the o1 process that there will be new ways to use it, and it’s going to take time for us and users to figure out how to do that.”
Why is o1 so important?
OpenAI o1 is a completely new class of large language models. Previous generations and approaches, including the GPT family, responded to a user's query token by token, often resulting in hallucinations or completely incorrect information.
There are a few workarounds for this, including larger context windows that allow the AI to access previously fixed parts, and memory functions that do the same thing across multiple chats. However, these are only temporary solutions, and a paradigm shift was needed.
With o1, OpenAI changed its approach, moving to a “train of thought” concept, where after you give an AI model a question, it goes away and works through the problem step by step, much like a human might think through a problem before providing an answer. I'm sure many of us got our grades down in school for not showing our solution properly; now, AI should be able to show its train of thought, too.
Speaking to T-Mobile, Altman cited healthcare and education as great areas where models like o1 could have a significant impact. “If you imagine every student getting personalized instruction that’s tailored to them and other learning opportunities,” that’s huge. He’s especially hopeful that AI will help make scientific discoveries. “If AI helps us invent new things, cure diseases, come up with better energy sources, that would be a huge win.”
Altman concluded by reiterating OpenAI’s commitment to deep learning and their belief in the path to artificial general intelligence (AGI), while remaining open to changing their approach based on continuous learning. He envisions a future where AI agents enabled by this technology will have a huge impact on the world.
Altman says we should expect the first full version of o1, not a mini or preview, in the coming months, and o2 and next-generation versions in the coming years. It's unclear whether the “o” family of models will become agents or whether this will be another paradigm shift like the transition from GPT to o.
