Revolutionary AI Breakthrough: Google's Bard Employs Coding for Accurate Computation
Google's innovative approach utilises AI models to write programs for precise computation, enhancing accuracy by 30%.
Large language models (LLMs) such as ChatGPT and Google Bard excel at answering specific questions but struggle with computational tasks.
To address this challenge, Google has devised a novel solution: employing AI to generate programs. Instead of merely displaying the language model's output, Bard, for instance, will write a program, execute it and present the program's output as the answer when faced with computational queries like math or string manipulation.
In a blog post, Google cites an example where ChatGPT fails to correctly reverse the word 'Lollipop,' producing the inaccurate result "pillopoL." In contrast, Bard accurately reverses the word as "popilloL" while also sharing the python code employed to solve the query. While programming enthusiasts may appreciate this peek under the hood, it may seem bewildering and irrelevant to regular users. It is akin to Gmail displaying lines of code instead of fetching emails—undoubtedly peculiar.
Bard, focus on the task at hand!
Google likens the process of an AI model writing a program to humans performing long division, as it represents a distinct mode of thinking. This approach, referred to as "writing code on the fly," will also be utilised for queries such as determining prime factors or calculating growth rates. According to Google, this method has boosted Bard's accuracy in computation-based problems by approximately 30% based on internal challenge datasets. However, Google cautions that Bard, like all of us, may still stumble due to misinterpreting the question or encountering code that requires refinement.
Experience Bard's real-time coding prowess firsthand here and put it to the test.
Google introduces an AI solution to enhance language models' ability to perform computations accurately.
Instead of displaying the language model's output, Bard writes a program, executes it and the python code used to solve the queries will also be shared.
Google compares the AI model writing a program to humans performing long division and that the "writing code on the fly" approach improves Bard's accuracy by around 30% in computation-based problems.