Jun 08, 2023 12:24:00

Google's interactive AI `` Bard '' introduces `` implicit code execution '' that allows code to be executed in the background, improving answer accuracy for complex tasks including string manipulation and logic / reasoning

`` Bard '' is an interactive AI

announced by Google in February 2023 and released to the public in March of the same year. Google introduced a technique called ' implicit code execution' that allows the Bard to run code in the background and respond accurately to prompts about mathematical tasks, coding questions, and string manipulation. It was announced on June 7, 2023 local time that it was introduced.

Bard updates: More accurate responses, export to Google Sheets
https://blog.google/technology/ai/bard-improved-reasoning-google-sheets-export/

A new technique called 'implicit code execution' helps Bard detect prompts that require computation and execute code in the background. Through 'implicit code execution', previously difficult-to-answer requests such as 'what are the prime factors of 15683615', 'calculate the growth rate of savings', 'turn the word 'Lollipop' upside down', He said he was able to give an accurate answer.

Below is an example of asking Bard the question 'Turn the word'Lollipop' upside down'. Type reverse the word 'Lollipop' for me to Bard.

Bard then replied, 'The reversed word of 'Lollipop' is 'popilloL'. (The reverse of Lollipop is popilloL).'

In addition, I am presenting the code I used to get this answer in Python format.

According to Google, a general large-scale language model has a structure like a prediction engine, and when receiving instructions from the user, it predicts the words that will come next and generates a response. Therefore, large-scale language models are very good at performing creative tasks such as generating sentences, but they are said to be unfamiliar with the fields of reasoning and mathematics.

Therefore, Google has introduced `` implicit code execution '' that enhances reasoning and mathematics ability by having Bard automatically generate and execute code. Google says that this approach was inspired by the `` dichotomy in human intelligence '' typified by psychologist Daniel Kahneman 's book `` Thinking, Fast and Slow ''.

According to Google, humans use both 'System 1' thinking, which is quick and intuitive, and simple thinking, and 'System 2' thinking, which is slow and effortful. The idea of System 1 is represented by improvisational performances by jazz musicians and blind touches. On the other hand, System 2 thinking is used when learning how to perform division with many digits and how to play musical instruments.

A typical large-scale language model uses only this system1 to create sentences quickly, while excluding complex thinking. If you try to solve a complicated math problem using System 1 alone, you will not be able to do complicated calculations, and you will only answer the first random number that comes to your mind. When doing actual mathematical calculations, it is necessary to solve the problem with the system 2 way of thinking, which is formulaic, lacks flexibility and creativity, but takes the right steps to produce an appropriate solution. is possible.

By implementing 'implicit code execution' in Bard, Google has evolved Bard into an interactive AI that has both System 1 and System 2 functions. Implicit Code Execution identifies and creates code for prompts that can be answered by creating logical code. You can run that code in the background and use the results to generate an accurate answer.

Google said, ``With the introduction of 'implicit code execution,' the accuracy of Bard's answers to words and mathematical problems that require calculation in Google's dataset has improved by about 30%. reporting. Bard also said that by introducing 'implicit code execution', Bard was able to significantly outperform other conversational AIs, such as ChatGPT and Bing's AI chat, which are equipped with popular large-scale language models. I'm here.

On the other hand, Google says that 'implicit code execution' is in the development stage, and Bard does not generate code to properly answer questions from users, or the generated code is wrong, or even generates code. He says that there are cases where it is not used in the answer. However, Google said, 'The ability to generate structured and logical responses by generating code is an important step in making Bard a more useful AI.'

Bard also announced that when it receives a request such as 'create a table for animal shelter volunteer registration', it will now be possible to export the results generated directly to Google Sheets. I'm here.