Upgrades to OpenAI and Meta’s artificial intelligence (AI) platforms are expected soon, potentially bringing AI to new heights. The AI chatbot ChatGPT will use OpenAI’s GPT-5 as its new “engine,” while Meta’s upgrade will be called Llama 3. The current version of Llama is used, among other things, to power chatbots on Meta’s social media accounts.
According to statements made to the media by representatives of OpenAI and Meta, these improved systems will include some capacity for forward planning. However, what specific changes would this invention make to AI chatbots’ capabilities?
Envision yourself choosing the ideal route to drive from home to work; that is, the set of options that is best in some way, depending on factors like timing or cost, for example. It would be absolutely possible for an AI system to decide between the two available options. However, creating the best route from start would be a significantly more challenging process.
A path is essentially a series of distinct decisions. Individual decision-making, however, is unlikely to result in the best overall outcome.
For example, sometimes you have to give up something in order to gain something later on. You might have to join a long line to get onto the freeway, for example, in order to travel more quickly afterwards. This is the fundamental idea behind a planning issue, which is a well-known subject in AI.
Similarities exist between this and board games like Go, where a player’s overall strategy determines the result of a match and where certain plays are intended to create openings that can be later taken advantage of.
AlphaGo is a game that the AI company Google DeepMind created a strong AI to play using a novel method to planning. It was not only capable of exploring a tree of possible choices, but it could also get better at it with practice.
Naturally, this is not really about playing games or discovering the best routes to drive. Large Language Models are the technology behind programs like ChatGPT and Llama 3. (LLMs). Giving these AI systems the capacity to think about the long-term effects of their decisions is at stake in this situation. This ability may also enable LLMs to perform other tasks because it is required to solve mathematical puzzles.
The purpose of large language models is to forecast the word that will come after a specific word in a string. However, in actual use, they are employed to forecast lengthy word sequences, like responses to queries from human users.
Currently, this is accomplished by expanding the original sequence by adding one word to the answer, followed by another, and so on. In technical terms, this kind of prediction is called “autoregressive.” But occasionally, LLMs might paint themselves into situations from which there is no way out.
Anticipated progress
Combining planning with deep neural networks—the kind of algorithms, or set of rules, that underpin the models—has been a key objective for LLM designers. The nervous system served as the original inspiration for deep neural networks. Through a procedure known as training, in which they are exposed to sizable data sets, they can get more proficient at what they do.
Executives from OpenAI and Meta have indicated that we may not have to wait much longer for LLMs with planning capabilities. But AI experts are not surprised by this; they have been anticipating this kind of development for a while.
Sam Altman, the CEO of OpenAI, was let go at the end of the previous year and then brought back. Rumors at the time suggested that the drama was related to the company’s creation of a sophisticated algorithm known as Q, but they have subsequently been refuted. Q‘s function is unclear, but when it was first announced, AI researchers took note of the name because it was similar to the names of already-existing planning techniques.
Yann LeCun, the leader of AI at Meta, responded to the rumors by writing on X (previously Twitter) that while replacing auto regression with planning in LLMs was difficult, nearly every elite lab was working on it. Additionally, he believed that Q* was probably OpenAI’s attempt to integrate planning into its LLMs.
LeCun was right when he suggested that the best labs should focus on planning, as evidenced by a recent patent filing from Google DeepMind.
It’s interesting to note that the inventors on the list were AlphaGo team members. The approach outlined in the application appears to be very similar to the one that directs AlphaGo toward its objectives. Additionally, it would work with the huge language models that now use neural network architectures.
This leads us to the remarks made about the capabilities of their enhancements by Meta and OpenAI officials. Vice-president of Meta’s AI research Joelle Pineau stated to the FT newspaper: “We are hard at work figuring out how to get these models to reason, plan, and have memory, in addition to just talking.”
If that succeeds, planning and reasoning could advance from straightforward, sequential word creation to the preparation of full talks or even negotiations. Then, AI might actually advance to a new level.
