Researchers hope that this model will help bridge the divide between language models and human-like reasoning capabilities. AI Am When an AI is given an inner monologue, it appears to teach itself to become smarter. In a paper that has not yet undergone peer review, researchers from Stanford University and a group called "Notbad AI" have collaborated to develop an AI model that takes a moment to "think" before generating answers. It shows its thought process, presents its reasoning, and asks users to determine the most correct response. The team behind this reflective AI, known as Quiet Self-Taught Reasoner (Quiet-STaR for short), aimed to create a model that not only has the ability to self-learn reasoning, which they successfully achieved in 2022 with the original Self-Taught Reasoner algorithm, but also to do so silently before providing answers. This way, it operates akin to a human's inner monologue, ideally occurring prior to speaking. In an enthusiastic X-formerly-Twitter thread about the new model he contributed to, Eric Zelikam from Stanford University stated, "Excitingly, self-teaching reasoning on diverse web text automatically improves other reasoning!" To construct this contemplative AI, the research team built Quiet-STaR on Mistral 7B, an open-source large language model trained on seven billion parameters. It is regarded as outperforming the latest version of Meta's Llama model, according to the Hugging Face AI community. Quiet-STaR has been programmed to display its reasoning when providing outputs, and users are able to choose the most accurate response among the options.

As noted in the paper, this approach resulted in the model being correct 47. 2 percent of the time, which, while not particularly impressive, is an improvement from the 36. 3 percent accuracy achieved without the additional reasoning training. Although the model still performs poorly in mathematics, only answering 10. 9 percent of the questions correctly, the pre-trained Quiet-STaR only answered 5. 9 percent correctly, indicating that its mathematical proficiency doubled during training. While these results may not be groundbreaking, they are intriguing because chatbots such as OpenAI's ChatGPT and Google's Gemini have historically struggled with common-sense reasoning. The researchers propose in their paper that Quiet-STaR could serve as a stepping stone to narrow the gap between language models and the reasoning capabilities of humans. Perhaps OpenAI's mysterious Q* (pronounced "queue star") model, which sounds strikingly similar, holds similar potential. Only time will reveal the truth. For more information on advancements in AI, refer to the State Department Report that warns of an AI apocalypse and suggests limiting compute power for training.