With upgrades to artificial intelligence (AI) systems developed by OpenAI and Meta, we may soon see artificial intelligence rise to the next level. OpenAI’s GPT-5 will be the new “engine” of AI chatbot ChatGPT, while Meta’s upgrade will be called Llama 3. Among other things, the current version of Llama powers chatbots on Meta’s social media platforms.
Statements made to the media by both OpenAI and Meta executives suggest that some of the forward-planning capability will be included in these upgraded systems. So how exactly will this innovation change the capabilities of artificial intelligence chatbots?
Imagine you’re driving from home to work and you want to choose the best route—in some sense, the optimal set of choices based on cost or schedule. An AI system will have the perfect ability to choose the better of the two available routes. However, creating the most suitable route from scratch will be a much more difficult task for him.
A route ultimately consists of a number of different options. However, making individual decisions alone is unlikely to lead to an optimal overall solution.
For example, sometimes you have to make a small sacrifice at the beginning to get some benefit later: maybe you can get into a slow queue to enter the highway in order to move faster later. This is the essence of the planning problem, one of the classic issues of artificial intelligence.
There are parallels here with board games like Go: the outcome of a match depends on the overall order of moves, and some moves aim to unlock opportunities that can be exploited later.
Artificial intelligence company Google DeepMind has developed a powerful artificial intelligence based on an innovative planning approach to play this game called AlphaGo. Not only was he able to explore the tree of options available, but he was also able to hone this ability with experience.
Of course, the main thing is not to find optimal routes for driving or playing. The technology that powers products like ChatGPT and Llama 3 is called Large Language Models (LLM). The point here is to give these AI systems the ability to consider the long-term consequences of their actions. This skill is also essential for solving math problems, thus potentially unlocking other abilities for graduate students.
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, such as answers to questions from human users.
This is now a word to the answer, followed by another word, etc. It is done by adding and thus expanding the first row. This is known in the jargon as “autoregressive” forecasting. However, Masters can sometimes paint themselves into corners that are impossible to get out of.
expected development
An important goal for the LLM designers was to: Combine planning with deep neural networks, the type of algorithms or set of rules behind the models. Deep neural networks were initially inspired by the nervous system. They can improve what they do through a process called training, where they are exposed to large amounts of data.
The wait for a master’s degree in planning (LLM) may be over, according to comments from OpenAI and Meta executives. However, this is not a surprise for artificial intelligence researchers who have been waiting for such a development for a while.
Late last year, OpenAI’s CEO Sam Altman was fired by the company and then rehired. At the time, this drama was said to involve the company developing an advanced algorithm called Q*, but this statement was later revealed. since it was changed. Although it wasn’t clear what Q* did, at the time the name attracted the attention of AI researchers because it echoed the names of existing methods for planning.
Commenting on these rumors, Meta’s head of artificial intelligence said, Yann LeCun wrote on X (formerly Twitter) Replacing the automated regression process with planning in the MSc was challenging, but nearly every top lab was working on it. He also thought Q* was probably OpenAI’s attempt to incorporate planning into graduate studies.
LeCun was on to something in what he said about top labs because recently Google DeepMind published a patent application that points to planning capabilities.
Interestingly, the inventors listed were members of the AlphaGo team. The method described in the application is very similar to the method that guides AlphaGo towards its goals. It will also be compatible with existing neural network architectures used by large language models.
This brings us to comments from Meta and OpenAI executives about the capabilities of their upgrades. Joelle Pineau, Meta’s vice president of AI research, told the FT newspaper: “We’re working hard to figure out how to make these models not just talk but actually reason and plan… have memory.”
If this works, we could see progress in planning and reasoning, moving from simple, step-by-step word generation to planning entire conversations and even negotiations. Then we may see AI truly rise to the next level.
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Nello Cristianini does not work for, consult, own shares in, or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond his academic duties.