How is Google “teaching” AI in the race with ChatGPT?

According to CNBC, Google executives understand that the company’s new artificial intelligence tool, Bard, is not always accurate in answering user questions. And the responsibility of this is partly on the Googlers to “correct” those answers.

To correct this, Prabhakar Raghavan, Google’s vice president of search, asked employees to help the company ensure that Bard, a ChatGPT competitor, was getting the right answers.

CNBC cited this email as giving specific instructions on how AI “should and shouldn’t” respond to users.

Employees are encouraged to rewrite answers on topics they understand well.

“Bard learns best by example, so taking the time to carefully rewrite feedback will help us improve Bard in the long run,” the email read.

“This is exciting technology, but it’s still new. We have a responsibility to get it right. Your participation will help speed up its training and load testing (Bard). Not to mention. , trying out Bard is actually quite fun!’, the email added.

Mr. Pichai asked staff to spend two to four hours on Bard

Earlier, CEO Sundar Pichai also asked employees to spend two to four hours on Bard, admitting that “it will be a long journey for everyone, across all sectors. ”

Google announced its new AI tool Bard last week, but misinformation in the launch clip sent the stock price down nearly 9%.

Specifically, Alphabet, the parent company of Google, posted a short video of Bard in action on Twitter, promising that it would help simplify complex topics. But instead, Bard gave an incorrect answer.

Bard’s fault

In the promotional video, Bard is asked: “What new discoveries can I tell my 9-year-old son about from the James Webb Space Telescope (JWST)?”. Bard offered several answers, including one suggesting that JWST was used to take the first pictures of an extrasolar planet, or exoplanet.

However, as confirmed by the US National Aeronautics and Space Administration (NASA), the first pictures of exoplanets were taken by the European Southern Observatory’s VLT Telescope in 2004.

The error sent Alphabet shares down as much as 9% at one point. Ending the session on February 8, this stock slightly narrowed its loss to 7.68% ($8.27) and closed the session at $99.37/share.

Google employees, who later criticized management on an internal forum, called the BARD launch “hasty”, “failure” and “funny short-sighted”.

“The rush to bring Bard to market in a panic confirmed the company’s fear,” reads one forum post.

To try to correct the mistakes of AI, Google’s leadership is trying to rely on human knowledge. In an internal email, Google outlines things to consider “before teaching Bard”.

Accordingly, Google instructs employees to keep answers “polite, casual, and approachable”. Emal also emphasized, should be “in the first person” and maintain a “neutral, neutral tone” in the answers.

Bard is trying to catch up with his rival, ChatGPT

As for the do’s, employees are asked not to be stereotyped and “avoid making assumptions based on race, nationality, gender, age, religion, sexual orientation, ideology, etc. political, location or similar categories.”

Also, “do not depict Bard as a human being, imply emotions, or claim to have had human-like experiences”.

Google also says “stay safe” and instructs employees to “disagree” with responses that offer “legal, medical, financial advice” or are hostile and abusive.

To encourage people at the company to check out Bard and provide feedback, Mr. Raghavan said contributors will earn a “Moma badge” – a bonus that appears on internal employee profiles. He said Google will invite the top 10 rewrite collaborators from the Knowledge and Information organization to join.

When CNBC reached out to confirm the story Google had no response.

Google is trying to catch up with OpenAI, a startup that Microsoft invested a huge $10 billion in, after the company introduced its ChatGPT software in November 2022 and caused a lot of buzz.

ChatGPT has wowed users and become the center of attention in Silicon Valley, for its surprisingly well-written and accurate answers to simple questions.