Google’s AI-powered chatbot, Bard, is showing signs of improvement when it comes to tasks involving logic and reasoning, according to a recent blog post by the tech giant. By utilizing a technique called “implicit code execution,” Bard has made significant advancements, specifically in the areas of math and coding.
The blog post explains that large language models (LLMs) like Bard are primarily prediction engines. When given a prompt, they generate responses by anticipating the next words in a sentence. While this makes them excellent at writing emails and essays, their performance in software development tasks has been somewhat error-prone.
Although code-generating models like GitHub’s Copilot and Amazon’s CodeWhisperer exist, they are not designed for general-purpose use. Unlike Bard and similar models such as ChatGPT, which were trained on a wide range of text samples, including web content, e-books, and other resources, code-generating models were mostly trained on code samples.
To address the limitations of general LLMs in coding and mathematics, Google developed implicit code execution, enabling Bard to write and execute its own code. The latest version of Bard identifies prompts that would benefit from logical code, generates the code internally, tests it, and incorporates the result into its response, aiming for greater accuracy.
According to internal benchmarking, Google claims that the new Bard has shown a 30% improvement in responding to “computation-based” word and math problems compared to the previous release. However, these claims will need to be verified through independent testing.
When Bard was initially launched earlier this year, it faced criticism and unfavorable comparisons to competitors such as Bing Chat and ChatGPT. The rollout experienced a setback when a Google ad featuring a wrong answer from Bard briefly caused an 8% drop in the company’s stock.
To address these issues and regain ground, Google has introduced implicit code generation and other enhancements, including support for new languages, multimodal queries, and image generation.
However, it remains to be seen whether these improvements will be sufficient to compete with leading generative AI chatbots in the field.