by We may soon see AI step up to the next level, with upgrades to artificial intelligence (AI) systems developed by OpenAI and Meta. OpenAI’s GPT-5 is the new “engine” inside the ChatGPT AI chatbot, and Meta’s upgrade will be called Llama 3. Among other things, the current version of Llama powers chatbots on Meta social media platforms.
Statements to the media from executives at both OpenAI and Meta indicate that some forward planning capability will be incorporated into these upgraded systems. But how exactly will this innovation change the capabilities of AI chatbots?
Imagine that you are driving from home to work and you want to choose the best route – that is, the best order of options in some way, based on cost or timing, for example. An AI system would be able to choose the best route from two existing routes. But it would be a much more difficult task for him to generate the best route from scratch.
A path is ultimately a sequence of different choices. However, making individual decisions in isolation is unlikely to lead to an optimal overall solution.
For example, sometimes you have to make a small sacrifice at the beginning, to benefit later: you might join a slow queue to enter the motorway, to move faster later. This is the essence of a planning problem, a classic topic in artificial intelligence.
There are parallels here with board games such as Go: the outcome of a game depends on the entire sequence of moves, and some moves are aimed at unlocking opportunities that can be exploited later.
The AI company Google DeepMind developed a powerful AI to play this game called AlphaGo, based on an innovative approach to planning. Not only was he able to explore a tree of available options, but also to improve that ability with experience.
Of course, the real point isn’t about finding the best ways to drive or play games. The technology that powers products such as ChatGPT and Llama 3 is called Large Language Models (LLMs). The point here is to provide these AI systems with the ability to evaluate the long-term consequences of their actions. This skill is also essential for solving mathematical problems, so it may unlock other abilities for LLMs.
Large language models are designed to predict the next word in a given sequence of words. But in practice, they are used to predict long strings of words, as answers to questions from human users.
This is currently done by adding one word to the answer, then another word and so on, adding to the initial sequence. This is known in the jargon as an “autoregressive” prediction. However, LLMs can sometimes paint themselves into unfathomable corners.
Expected development
It was an important goal for LLM designers combining planning with deep neural networks, the type of algorithm – or set of rules – that sits behind the models. Deep neural networks were originally inspired by the nervous system. They can improve what they do through a process called training, in which they are exposed to large data sets.
The wait for LLMs who can plan may be over, according to comments from OpenAI and Meta executives. However, this does not come as a surprise to AI researchers, who have been anticipating such a development for some time.
At the end of last year, OpenAI CEO Sam Altman was fired and then rehired by the company. At the time, there was a rumor that the play was related to the company’s development of an advanced algorithm called Q*, although this explanation. since replaced. Although it is not clear what Q* does, at the time, the name struck a chord with AI researchers because it echoed the names of existing planning methods.
Commenting on those rumours, the head of AI Meta said, Yann LeCun, write on X (formerly Twitter that it was challenging to plan in LLMs to replace the auto regression process, but that almost all top labs were working on it. He also thought that Q* was probably OpenAI’s attempt to include planning in its LLMs.
LeCun was on to something in what he said about the best labs, because recently, Google DeepMind published a patent application that hinted at planning capabilities.
Interestingly, the inventors were listed as members of the AlphaGo team. The method described in the application looks a lot like the one that guides AlphaGo towards its goals. It would also be compatible with the current neural network architectures used by large language models.
That brings us comments from executives at Meta and OpenAI about the capabilities of their upgrades. Joelle Pineau, vice-president of AI research at Meta, told the FT newspaper: “We are working hard to find out how to get these models not only to talk, but to reason, to plan . . . to remember.”
If this is successful, we may see a progression in planning and reasoning, moving from simple step-by-step generation of words to planning entire conversations, or even negotiations. Then we might see AI step up to the next level.
This article from The Conversation is republished under a Creative Commons license. Read the original article.
Nello Cristianini does not work for, consult with, share in, or be funded by any company or organization that would benefit from this article, nor has she disclosed any material interests beyond their academic appointment.
