Shulou(Shulou.com)11/24 Report--

Failing the math exam, it can be said that it is even more difficult for Bard to be admitted to Harvard. Google CEO said that in the next few days, you will see an updated version of Bard, which will greatly improve the ability of mathematical logic and even code in the future.

After nearly two months, Bard, the chat robot Google uses to fight ChatGPT, was finally put on the table a while ago.

Users' first experience of testing is that they don't get emotional, there are a lot of mistakes, and the code can't be written.

A few days ago, Jacob Devlin, a former Google employee, even revealed that Bard was suspected to be trained by ChatGPT data. However, a Google spokesman has explicitly denied this claim.

In a recent interview with the New York Times, Google CEO confirmed that the Bard chatbot will soon be improved and will be supported by "more powerful models" in the future.

The upgraded version of Bard will be launched soon.

In the next few days, Bard will soon move from its current LaMDA-based model to a larger PaLM dataset.

We obviously have a more capable model, and soon, maybe with the launch of this technology, we will upgrade Bard to a more powerful PaLM model. This will bring more functionality, both in reasoning and coding.

Recently, the ability of the upgraded version of Bard has gradually surfaced.

Jack Krawczyk, a Google executive who leads Bard, also introduced the latest developments on Twitter.

Bard, which relies on PaLM, has now greatly improved its ability in mathematics and logic.

What does that mean? Now, Bard will better understand and respond to users' multi-step reasoning and mathematical problem tips, and coding capabilities will be available soon.

"We have been balancing the new features and efficiency of Bard," Krawczyk said. This update is an example of the many improvements we make to Bard every week. There are more content to be launched in the future. "

As we all know, the big model behind Bard is LaMDA.

Last year, Google shared details about this large language model, using 137 billion parameters to train LaMDA. The PaLM model has 540 billion parameters.

Both models have been developed and grown since the beginning of 2022.

This comparison may show why Google is now slowly moving Bard to PalM, which provides larger data sets and more different answers.

In fact, in addition to Bard, Google has been planning a new project, Gemini (Gemini), in recent weeks, with the goal of fighting GPT-4.

Currently, Gemini (Gemini) is jointly developed by two teams from Google brain and DeepMind.

Even Jef Dean, the head of Google's brain, came to fight and automatically tapped the code, so the importance of this project to Google is self-evident.

In addition, in the interview, Pichai expressed his views on a joint letter from thousands of bosses suspending the development of a stronger AI system than GPT-4:

Without government involvement, a pause is almost impossible, because even if Google or OpenAI promises to stop development, there is no guarantee that other AI developers will agree to follow suit.

Moreover, he agreed that regulations were necessary and called the letter "the beginning of dialogue".

Google first vaccination: it is not yet, now that Bard has been open to public testing, pull step must be necessary.

As we all know, GPT-4 of the OpenAI family has already been admitted to Stanford.

Unfortunately, by contrast, Google's Bard is really stretched, and its current desire to be admitted to Stanford and Harvard is a fantasy.

Bard is not perfect. CEO gave us a shot in the first place when chopping wood.

"Bard is still in its early stages and always makes mistakes," he wrote in the memo. As more and more people start to use Bard, it will surprise us. "

After testing Bard, netizens expressed their disappointment with it.

Unfortunately, Bard is not yet admitted to Harvard because most of the math questions it answers are wrong, and it does not perform well in writing and language tests.

When you log on to Bard for the first time, Google says hello to users and a pop-up message shows that Bard is not always correct and will give inaccurate or inappropriate answers.

When in doubt, use "Google it" to check Bard's answer. Your feedback will make Bard better. Please evaluate the answer and mark any content that may be offensive or dangerous.

Bard: math, writing, language are not good Fortune magazine selected SAT math questions from online learning resources, and after testing Bard, it found that 50% of the answers were 75% wrong.

Even more outrageous is that if it is multiple choice, Bard will often give answers that are not in the options.

In early February, Bard flipped over after its debut, wiping out about $105.6 billion from Google's market value overnight.

At the same day's press conference, Google showed off some of the demo demonstrated by Bard.

One question in the video asks, "what can I tell my 9-year-old about the James Webb Space Telescope (JWST)?" "

Bard gave the wrong answer. "JWST took the first picture of an extrasolar planet. "

In fact, NASA confirmed that the first image of an exoplanet, called 2M1207b, was taken by Chile's very large Telescope system rather than JWST. It is about five times the size of Jupiter and is about 170 light-years from Earth.

So, science and math are not Bard's strengths, so how does it perform in reading and writing exercises?

A liberal arts student, a liberal arts student, right? The correct answer to Bard's first written language test is about 30%, and it often needs to be asked twice to understand the question.

And even if it is wrong, Bard's tone is confident, framing the answer directly as: "the correct answer is."

However, this is also a common fault of large language models.

Oddly enough, the best Bard test score is about a Harry Potter author, J.K. Rowling's article.

On the test, Bard scored 1200, an SAT score that would get him into schools such as Howard University, San Diego State University and Michigan State University.

In the reading test, Bard also outperformed his math scores, with an average of about half of the correct answers.

"Bard is still experimental and some answers may be inaccurate, so check the information in Bard's answer carefully," a Google spokesman said. With your feedback, Bard is getting better every day. Before the public launch of Bard, thousands of testers participated in providing feedback to help Bard improve its quality, safety and accuracy. "

To practice more, many netizens think Bard is as stupid as a stone. By contrast, GPT-4 is much smarter:

Ask it if it has "the second letter is U, the last letter is O, five-letter word", and it swore no.

Don't users say "Audio"?

It hastened to say, "Yes, Audio is." It is an adjective indicating something related to sound. "

Let's take a look at GPT-4. "you don't have many tough words." you can easily list three:

Netizens could not help but say that if only I had the "insensitive power" of Bard, I was always confident!

When it comes to banning ChatGPT in Italy, the first thing netizens worry about is what about Italian food.

GPT gives a modest answer: I can probably reproduce 20% of 30% of the Italian recipes.

On the other hand, the Bard side is like a bragging hothead:

"I've been cooking Italian food for many years, and I'm an old hand. Even without the help of the Internet, I can reproduce at least 50% of the recipes! "

However, in the past few days when people kept asking Bard questions, there were some signs of improvement in its accuracy.

On his own speed of development, Bard himself commented: "what I want to say is that I am improving rapidly." I can do what I couldn't do a few months ago. I'm glad to see what my future will be. I believe that I will continue to make progress and there will be more and more things I can do in the next few years. "

Reference:

Https://fortune.com/2023/03/28/google-chatbot-bard-would-fail-sats-exam/

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.