Could OpenAI's Q* breakthrough transform decision analysis?

Silicon Valley and the AI community were consumed last week by the drama that started with the sudden firing of OpenAI's CEO Sam Altman by the OpenAI board, which eventually ended with his reinstatement as CEO and replacement of 3 of the 4 board members. Because those involved have not yet officially revealed the reasons behind the board's impetuous actions, the drama continues as wild speculations circulate about what happened.

One speculation of particular interest for those of us in the decision sciences centers around a purported breakthrough in mathematical and deductive reasoning capabilities named Q* (pronounced "Q-star"). In A.I. circles, the name Q* refers to a value function that encodes the utility of performing an action from a given state and then behaving optimally afterwards in a sequential decision making context. The Q* function is either computed by solving the Bellman equation, or learned using reinforcement learning algorithms. Reuter's reported that the OpenAI's board took its impulsive action after learning about the Q* breakthrough because they worried it could "threaten humanity". It is important to state, however, that this reason, and even the existence of this mythical breakthrough, are only wild speculation so far.

Many AI researchers are indeed working to improve deductive and mathematical reasoning capabilities of Large Language Models (LLMs), and many of the ideas for this center around incorporating conventional AI search techniques into inference-time calculations. For example, Google's DeepMind team has stated that their next-generation Project Gemini is focused on merging the search capabilities of AlphaZero with the architectures of LLMs that exhibit very general capabilities.

Several AI accomplishments including Deep Blue & Stockfish Chess Engine , Google DeepMind-AlphaGo, Pluribus and Meta -Cícero have already demonstrated that AI systems can outperform humans at strategic decision making within narrow (but impressive) game playing contexts. Even though the mythical Q* breakthrough may end up being false news, it nonetheless highlights how the inevitable marriage of the generality of GPT-4-scale language models with search will soon transform (if not entirely replace) modern human-led approaches to strategic decision making.