(Dec 21): OpenAI is preparing to launch a new artificial intelligence (AI) model that it said is capable of more advanced human-like reasoning than its current offerings, ratcheting up the competition with rivals such as Alphabet Inc’s Google.
The new o3 model, unveiled during a live-streamed event on Friday, spends more time computing an answer before responding to user queries, with the goal of solving more complex multi-step problems. The company will also introduce a smaller version of the model called o3-mini.
On the live stream, OpenAI shared some early details of how o3 bests o1, the reasoning model it introduced in September, when answering complicated questions related to topics such as coding. OpenAI is also asking safety and security researchers to apply to test the models as part of its process before launching new software. The company plans to release the o3-mini model at the end of January and the o3 model shortly after that, OpenAI chief executive officer Sam Altman said during the event.
OpenAI kicked off an AI arms race with the release two years ago of ChatGPT, a chatbot that was initially powered by a large language model called GPT-3.5. OpenAI followed that up in 2023 with GPT-4, which it described as more accurate and creative, and more recently with o1, its first reasoning model. The spokesperson said OpenAI decided not to name the new model o2 “out of respect” for the British telecommunications brand that goes by that name. The Information previously reported the name.
Other top AI developers are pushing ahead with increasingly advanced technology, too. Earlier this month, Google debuted a new version of its flagship model, Gemini, that it said is twice as fast as the previous model and can “think, remember, plan and even take action on your behalf”. Meta chief executive officer Mark Zuckerberg also recently teased plans to introduce Llama 4 next year.
However, several leading companies, including OpenAI and Google, are confronting diminishing returns from their costly efforts to develop newer models, Bloomberg News has previously reported. That’s due in part to the challenge of finding enough new, untapped sources of high-quality, human-made training data. To help get around that, companies are turning to new tactics, including more emphasis on so-called reasoning.
Along with the model previews, OpenAI released research describing a new approach it’s using to ensure that systems like o1 and o3 do what they should and avoid, say, helping users carry out illegal activities. Called “deliberative alignment”, the technique has the models follow a series of safety-related steps as they consider how to respond to a user’s query.
Alignment, as the issue is sometimes called, represents a technical challenge for those building large language models, which are typically trained on huge swaths of internet data. The effort is complicated by the fact that people’s ethics and values vary — as do their ideas of what AI should and should not be permitted to do.
OpenAI’s latest announcement capped off 12 days of live-streamed product events. The startup has used the launch series to introduce a more expensive new ChatGPT Pro subscription option and begin rolling out an AI video generation tool called Sora, among other new products.
Uploaded by Tham Yek Lee