When I woke up, GPT-4 brushed the screen, how to say this scene, a little chaotic. The melon-eating masses shivered while relishing.
Talking about the revolutionary new features, shivering that silicon-based people are one step closer to replacing carbon-based people.
So, what are the changes in GPT-4 compared to the previous generation model?
There have been a lot of screen brushing in the morning, so we will briefly summarize a few points ↓
❶ It can read the picture and speak
The prelude version of GPT, just NLP (natural language processing) models, can only process text information.
GPT-4 is a multimodal model that can not only accept text, but now also accept image input and output text results.
For example, we can throw a picture to it and ask it to tell where the meme of this diagram is.
Because there is no actual measured entrance yet, I can only look at the official example ↓
Do you understand? The user gave a combination diagram of the "VGA Apple phone charging cable" and asked the GPT-4 to point out the "laughing points" one by one, and it was obvious that the GPT-4 answered well, and it really understood.
If I were to do a test demo, wouldn't it be more intuitive for me to let GPT-4 play "Find the Difference".
In fact, this is nothing, what is more bullish is the official developer demo video ↓
You can sketch a very rudimentary website UI on paper, like the following.
Then, take a picture with your phone and upload it to GPT-4, asking it to generate detailed HTML and JS code based on this sketch.
The command in the demo is:
Then, GPT-4 is like a skilled front-end engineer, writing the code in an instant, and seeing this, I am a little square.
Previously, ChatGPT can help coders write code, help network workers knock CLI We have seen no surprise, after all, the code is good, the network configuration command is good, the reuse rate itself is relatively high, as long as the training is in place, it is not difficult to help you write some examples.
For example, this is something I was bored to let ChatGPT do, but there is actually no technical content ↓
In this case, if we need to let ChatGPT help us really complete a job, we need to write "Prompt" step by step in great detail, which is actually very considering the professional ability and skill of the questioner.
But now, if GPT-4 really has the ability to read and read pictures, it will be really terrible↓
Although GPT-4's ability to read maps has yet to be tested, this day does not seem far away.
❷ It is a small town in California
The official gave a set of GPT-4 brush scores, although it is too early to say that it is a super student, but the results of some subjects are already quite against the sky.
Looking at this result, the most impressive is the Biology Olympiad test and the GRE test, both of which beat 99% of the references.
The Uniform Bar Exam is also very good, ranking in the top 10%. GPT-3.5 to take the bar exam, can only rank in the bottom 10%.
Other questions are also unambiguous, such as SAT math, which can get 700 points (equivalent to 92%), while GPT-3.5 is 590 points (about 70% percentile).
Combined with the previous ability to read pictures, I think the more ultimate exam scenario is ↓
You take a picture of the "Huanggang Scroll" and it will directly write the answer to you
❸ It's harder than ever to fool
In the past, when using ChatGPT (based on GPT-3.5), I often got some ridiculous answers, even contrary to the facts.
Or you can "brainwash" it so that he gives the wrong answer.
For example, the following question, the first answer has obvious errors, you can further fool it.
Next, I changed the ChatGPT model to GPT-4 and asked the same question again, this time the answer looked much more reliable, and it seemed that I had done a lot of homework.
But if you continue to fool it, you can still fool it lame.
According to the official introduction, compared with GPT-3.5, GPT-4 has greatly improved the factual accuracy rate for 9 major categories of problems.
In previous versions, it was generally below 60%.
Of course, the current flaws are still relatively obvious, that is, our answers to ChatGPT/GPT still need to be carefully reviewed manually.
For example, the following one is very absurd. I believe you are a ghost!
❹ It's more expensive than before
In GPT-3.5, the price of the API was 0.2 cents/1000 words (token), and in version 4.0, the price skyrocketed, becoming: 6 cents/1000 words (token)
Maybe the cost of training a model is too high, and the cost of inference will not drop, but I believe that this price will gradually decrease later.
In addition, the input text length supported by GPT-4 has been greatly improved, up to 32,000 (token).
This is still useful, you want it to help you proofread an article, or help you check for bugs in a piece of code that may have had to be divided into many paragraphs to input and output before.
Now, to exaggerate, Ctrl+A/Ctrl+C/Ctrl+V, you can wait for the result.
❺ ChatGPT Plus users have benefits
This time, for ChatGPT Plus users, the $20-knife monthly subscription fee is not in vain. You can try the GPT-4 model first, and after logging in, you can flexibly choose between the three models.
Moreover, the official also gives the difference between the three models very intuitively:
The default paid version GPT-3.5, response speed 5 stars, reasoning logic 3 stars, simplicity 2 stars
The free version of GPT-3.5, response speed 2 stars, reasoning logic 3 stars, conciseness 1 star
GPT-4 trial version, response speed 2 stars, reasoning logic 5 stars, conciseness 4 stars
I tried the GPT-4 mode in ChatGPT, and it was indeed a bit stuck, and in order to prevent too many requests from the enthusiastic melon-eating masses, the official also made a limit: every four hours, up to 100 messages.
And because the image input mode is only preview, the GPT-4 mode of ChatGPT can only enter text.
Another problem with the GPT-4 model is that its training dataset is the same as GPT-3.5, as of September 2021.
In other words, GPT-4 has a smarter brain and clearer brain circuits, but its knowledge structure is still the same as before. I don't read much, it's still those old sesame rotten grains.
Okay, GPT-4 melon, let's eat it here first.
The overall feeling is that it is still too early to say that the singularity of artificial intelligence has arrived, but everyone needs a breakthrough too much, so they will "".
But do we really need that many big models?
There are too many large models, and the customer scenario is a bit inadequate