Google’s has improved Bard using a technique called “implicit code execution.” This enhancement has enhanced Bard’s abilities in math, coding, and string manipulation. Bard can now identify and execute code in the background, leading to more accurate responses in these areas.
The blog post by Google explains that Bard, like other large language models, works by predicting the most likely words to follow a given prompt. While this makes them excellent at tasks like writing emails and essays, it also reveals their limitations in software development.
The new method used in this update combines code generation and execution, inspired by the concepts of System 1 and System 2 thinking from Daniel Kahneman’s book “Thinking, Fast and Slow.” System 1 thinking is fast and intuitive, while System 2 thinking is deliberate and effortful. Language models mainly rely on System 1 thinking, which allows them to generate text quickly but may lead to limitations. For example, in math problem-solving, relying solely on System 1 thinking can result in impulsive responses without proper calculation. Traditional computation, on the other hand, aligns with System 2 thinking, allowing for a sequential and meticulous approach to solving complex problems.
Google’s latest update combines the power of language models and traditional code to improve Bard’s accuracy. By executing code in the background, Bard can identify and handle prompts that require logical code, leading to more precise responses. This approach has shown a significant 30% improvement in accuracy when dealing with word and math problems.
However, Bard may still have limitations and may not always provide perfect responses.
The sources for this piece include an article in TechCrunch.