Google’s language model, Bard, is getting a major update today aimed at improving its logic and reasoning capabilities.
This was announced in a blog post by Jack Krawczyk, product manager at Bard, and Amarnag Subramanya, vice president of technology at Bard.
A Leap Forward in Reasoning and Mathematics
These updates aim to improve Bard’s ability to do math, answer coding questions, and handle string manipulation prompts.
To achieve this, developers integrate “implicit code execution”. This new method allows Bard to recognize computational prompts and run code in the background, allowing him to more accurately respond to complex tasks.
“This allows it to more accurately respond to math tasks, coding questions, and string manipulation prompts,” the Google team said in the announcement.
System 1 and System 2 thinking: A blend of intuition and logic
The approach used in the update is inspired by the well-studied dichotomy of human intelligence as discussed in Daniel Kahneman’s book Thinking, Fast and Slow.
The concept of “System 1” and “System 2” thinking is central to Bard’s enhanced abilities.
System 1 is fast, intuitive and effortless, comparable to a jazz musician improvising on the spot.
However, System 2 is slow, deliberate, and laborious, comparable to performing long divisions or learning an instrument.
Large Language Models (LLMs) like Bard typically operate under System 1 and generate text quickly but without deep thought.
Traditional calculations are closer to System 2 as they are formulaic and inflexible, but when done correctly they can produce impressive results.
“You can think of LLMs as functioning exclusively according to System 1 – they produce text quickly, but without much thought,” the blog post says. However, “with this latest update, we have combined the capabilities of LLMs (System 1) and traditional code (System 2) to improve the accuracy of Bard’s responses.”
One step closer to enhanced AI capabilities
The new updates represent a significant advance in the field of AI language models and enhance Bard’s ability to provide more accurate answers.
However, the team recognizes that there is still room for improvement:
“Even with these buffs, Bard won’t always get it right…this enhanced ability to react with structured, logic-driven abilities is an important step in making Bard even more helpful.”
While the improvements are notable, they present potential limitations and challenges.
It’s plausible that Bard doesn’t always generate the correct code, or doesn’t include the executed code in his answer.
It can also happen that Bard does not generate any code at all. Additionally, the effectiveness of “implicit code execution” might depend on the complexity of the task.
In total
As Bard integrates advanced reasoning capabilities, users can look forward to more accurate, helpful, and intuitive AI support.
However, every AI technology has limitations and disadvantages.
As with any tool, you should consider approaching it with a balanced perspective and understanding the capabilities and challenges.
Featured image: Amir Sajjad/Shutterstock
