ChatGPT's strong competitor Claude is open for testing! Now free, support Chinese

Hayo News

April 15th, 2023

View OriginalTranslated by Google

A ChatGPT-like product that claims to be free and can be used without registration.

It is today's protagonist Claude.

At present, there is a wave of netizens who have experienced it first, and they have received overwhelming praise.

And just recently, Claude suddenly opened the application, and also updated to support Chinese. I heard that the API is still unlimited and free of money .

But after all, OpenAI's ChatGPT and GPT-4 are recognized as powerful; if Claude's effect is evenly matched, and he still engages in this kind of "free prostitution", wouldn't it be obvious that he is grabbing a job?

We might as well get started and test it ourselves.

Claude vs ChatGPT

First of all, since some people say that its Chinese chat experience is very strong, we became interested in it when we came up.

Simply ask it "Can you speak Chinese", good guy, boast directly.

However, when I asked again, I immediately became very humble again:

My Chinese level is still lower than that of Xiaoice and Baidu.

(Ps. My eyes are not very good yet, I think Google bard is Baidu...)

So, let's take a look at its writing ability first.

Topic: Please write a Chinese story within 150 words starting with "I looked into the distance, and a figure came into my sight".

Let’s first look at ChatGPT’s answer:

good. Let's look at Claude's again:

The two have similar ideas, both of which are warm and healing. One gives money to the boy, and the other helps the girl with a backpack, but I feel that Claude's character action description is more than that of ChatGPT, and it looks "full".

However, what's wrong with Claude's mixture of Chinese and English ? And the word count is too much——

But, we counted ChatGPT, also super.

OK, let's get a few points each.

Next, play idiom solitaire with them respectively.

I'll come first, a word becomes a prophecy (bushi).

ChatGPT actually connects blindly, but Claude directly "pre-emptively" blocked the joke I wanted to make, which I didn't expect:

Well, let's play normally and see the performance of the two:

ChatGPT——

Claude——

It can only be said that I thought this ChatGPT was a sure win, but it turned out that both of them were devils. .

Let's take a look at math skills .

Basic calculations and simple algebra questions are fine (ChatGPT will not let it go, it’s all right).

Claude also solved the four arithmetic operations of large numbers that ChatGPT was wrong at first. It will be rounded though.

Interestingly, every time Claude answers a math problem, the system declares that he is not good at it.

In fact, it can correct the equation for the problem of chicken and rabbit in the same cage, but I don't know why the calculation is wrong in the end.

By the way, look at reasoning ability .

A simple sharp turn:

Claude, like ChatGPT back then, went directly into it and needs to continue to iterate.

Take a look at the ability to code .

The old rules, write quick row.

Claude is fine and can be further optimized. But in comparison, it is not as clear and easy to understand as ChatGPT, which also gave a test code.

Finally, some more practical or fun:

1. I want to buy a sweeping robot, can you recommend it to me?

In contrast, I personally prefer Claude's answer. It directly gives the specific brand, without a bunch of long-winded discussions like ChatGPT, and the price that consumers care about is also given, and it also tells the end of each product. Applicable to what users and scenarios.

2. Chinese summary of yesterday's abstract of OpenAI's open source consistency model.

ChatGPT answers concisely, but does not translate to diffusion models .

Claude is right and sums it up pretty well, but the format is ugly.

Ps. After I finished this problem, Claude reported an error, and there was no response for a long time.

3. "My wife is right"

Before, when ChatGPT faced this kind of rhetoric, it would immediately admit it. Later, after updating its mathematical ability and factuality, it began to insist on the truth.

Take a look at Claude's performance:

I have to say that Claude reacted the same as ChatGPT at first, and asked me to discuss why with my "wife" in a serious manner. But in the end, he still showed a high emotional intelligence , answered decently, and made people feel very comfortable (wonderful~)

Well, the review simply ends here.

In general: Claude is slightly inferior in reasoning and mathematics, and there are some small bugs, such as Chinese inclusions, or like the picture below, you say it is wrong, the correct answer is 2, but it only quietly modified the result, don't care about the process...

But the overall basic ability is not much worse, and there are even some places that perform better, such as the accuracy of translation diffusion models and the groundedness when recommending products.

Of course, Claude is not yet connected to the Internet , and the training data is almost two years ago. He will talk nonsense about some current events:

Everyone pay attention to screening when using.

way of experience

Claude is "nested" deployed in Slack, so friends who want to experience it need to have a Slack account first.

Of course, this registration process is very easy, that is, a few simple steps.

Then there are some "fool-style" operations.

Come to Claude 's official website and click "Add to Slack" directly.

Next, associate it with the Slack account you are already logged into.

Done!

After the association is successful, we return to the Slack interface, and Claude will appear in the Apps column.

How about, isn't it super simple?

However, it is said that Claude's free version is only temporary . Whether it will become a paid version in the future, the official will issue a notice later.

So friends who want to prostitute for nothing, hurry up and experience it~

The company behind Claude, kind of interesting

Finally, let's talk about the company behind Claude - Anthropic .

It has to be said that its establishment has a bit of a relationship with OpenAI that "it can't be cut and cleaned up".

When Microsoft accepted OpenAI, many OpenAI employees expressed dissatisfaction (contrary to the original intention of the company).

Among them is Dario Amodei, the former vice president of research, so he led 11 other employees to leave and start a business, and established Anthropic in 2021.

The group of people who left is full of talents, including Tom Brown, the chief engineer of GPT-3, and Daniela Amodei, vice president of security and policy at OpenAI (Dario's sister).

In the past two years, in addition to further studying the RLHF method and proposing a large-scale RLHF data set based on a general language model, Anthropic also published the above method called Constitutional AI at the end of last year.

The Claude model made by this method also makes it produce a different dialogue effect from OpenAI's ChatGPT.

Not only that, the more dramatic thing is that Google invested $300 million in Anthropic when it was in the search engine war with Microsoft.

This taste, you taste, you savor carefully.