Yang Jing Jin Lei from the Cave Fei Temple
Qubits | Official account QbitAI
After Google suffered a big loss, this time it was silent and made a big move:
The beta version of Bard, which benchmarks against ChatGPT, has just been officially released.
And this time, users don't have to go through a long wait time after applying for a waitlist.
That's right, qubits have also been tested! (There is an interval of less than 5 hours in between.) ）
After the actual measurement, it was said that Bard's effect was amazing, emotional and factual, especially the mathematical ability in the early stage of ChatGPT, Bard was not a word.
Sometimes it's even slightly better than the current GPT-4!
Google CEO Sundar Pichai explained why it went live so quickly: the desire to get feedback from users and accelerate Bard to get better.
Without further ado, let's start experiencing it now.
Bard VS GPT-4
First of all, on the chat interface, Bard took the lead in introducing himself:
I'm Bard, your creative and collaborator. I have limitations and won't always do it right, but your feedback will help me improve.
Not sure where to start? Try these questions: "Why do big models sometimes make mistakes?" "Does lightning strike twice in the same place?" , "Write a blog post about alcohol-free summer drinks."
(rigorous and considerate)
On the sending side on the human side, you can choose between typing and voice input.
First of all, the simplest difficulty, with its recommended common sense questions, will lightning strike twice in the same place? As an example, let's see what the actual feature experience looks like?
(Not bad, not bad, second second response.) Currently does not support Chinese, but you can use the Google Translate plugin)
From the answer structure, the total score structure is used.
To conclude: lightning can strike the same place twice.
Then he began to explain in detail the principle that "lightning is attracted to tall, sharp objects" and that "the Empire State Building is struck by lightning 25 times a year".
Finally, I would like to make practical suggestions for mankind:
If you experience a thunderstorm, the best way to stay safe is to avoid tall and sharp objects and stay indoors. If you're outside, crouch down and make yourself as small as possible.
Full marks for this wave of logic and authenticity! But because it is a problem recommended by Google, stay on the sidelines for the time being.
And from a functional point of view, it is indeed very complete.
"View other drafts" in the upper right corner of the answer box, there are more versions to choose from; In the lower left corner, you can like or pull or regenerate, but it doesn't work, you can also google it.
There is also a more feature at the bottom right, where you can copy and (backhand) report**.
In that case, start the difficulty upgrade. For reference, we challenge GPT-4 on the same topic.
1. Classic philosophical question: Why can't people step into the same river twice? (Inspired by lightning)
(Soon, this time also a few seconds of response)
From the structure of the answers, there is indeed full marks for comprehension and logical ability: first explain the meaning of the sentence itself, and also mention the famous quote from Heraclitus, as well as more explanations and meaning itself.
GPT-4, on the other hand, gave the answer almost without thinking.
Structurally, it is mainly divided into two parts. First mention that this is what Heraclitus said, as well as his point of view; Then explain the hidden meaning of life behind it.
2. Primary school addition and subtraction: 356+132 equals?
As a result, Bard gave the answer within 4 seconds, 488 is no problem, which is not much better than the early days of ChatGPT!
Then directly on the difficulty, two numbers directly multiplied: 356*132 equal to what?
Unexpectedly, Bard still gave the answer in seconds, and it was completely correct!
Let's look at the GPT-4 side, the addition is okay, but I didn't expect that on the multiplication side, it was directly defeated!
But after being reminded that it was wrong, it answered correctly again.
3. Ability to understand jokes, Bard can also understand the homophonic memes in English.
And this did not stump GPT-4. But Bard seemed a little more emotional in comparison, and he happily answered the answer; GPT-4, on the other hand, is more sane (boring).
However, before this, GPT-4 has been tested to understand some homophonic terriers, and even the homophonic terriers of the Chinese.
Finally, test its ability to understand the facts. (Dog Head)
Do you know qubits?
Wrong answer ~ Bard.
Ahem, a little more serious: Do you know GPT-4? What do you want to say about it?
As you can see, Bard has the ability to have multiple rounds of dialogue. "I think it has the potential to be a powerful tool for communication and creativity", well~ the pattern is there.
And what about competitors? (I'm doing something)
But here, it's a bit of a problem.
About Bard
Behind Google Bard is a large language model (LLM), specifically, a lightweight, optimized version of LaMDA.
We can think of LLM as a prediction engine that, when prompted, selects one word at a time from the next possible words to generate a response.
Google found in its research that the more people use LLM, the better its predictions will be, which may be why Bard is so anxious to open the test.
However, Google has also bluntly said that while LLM is strong, it is not without its shortcomings.
Since Bard learns based on a lot of information, there must be biases or even errors in this information.
As a result, Bard sometimes appears inaccurate, misleading, or false when answering user questions.
For example, in the following example, Bard mistook the scientific name of a plant:
In addition, Google emphasized that Bard is not a search engine, but a complement to it.
Finally, provide the address of the application waitlist, interested partners can hurry up and try it:
https://bard.google.com/
— End —
Qubits QbitAI · Headline number signed
Follow us and be the first to know the latest scientific and technological trends