On March 14, OpenAI released GPT-4, dropping another "nuclear bomb" on the tech world.

According to OpenAI's demo, we know that GPT-4 has more power than GPT-3.5: summarizing articles, writing code, filing taxes, writing poems, etc. But if we go deeper into OpenAI's technical report, we may also find out about GPT-4.

But if we go deeper into the technical report published by OpenAI, we may also find more features about GPT-4 ......

As well as some details that OpenAI doesn't name and proclaim that might be a chill in the back.

The new Bing is loaded with GPT-4

Naturally, when GPT-4 was released, the new Bing was also loaded with the latest version.

According to a tweet by Microsoft Bing VP Jordi Ribas, the new Bing loaded with GPT-4 has upped the Q&A limit to 15 questions at a time, with a maximum of 150 questions a day.

Eight times longer text length

On GPT-4, the text length has been significantly increased.

Previously, we knew that GPT API calls were charged by "token", and one token usually corresponds to about 4 characters, while one Chinese character is roughly 2~2.5 tokens.

Before GPT-4, the token limit was around 4096, which was equivalent to 3072 English words. Once the length of the conversation exceeded this limit, the model would generate incoherent and meaningless content.

However, by GPT-4, the maximum number of tokens is 32,768, which is equivalent to about 24576 words, and the text length is expanded by a factor of eight.

That is, GPT-4 can now answer longer texts.

OpenAI says in the documentation that the context length limit for GPT-4 is now limited to 8192 tokens, and the version that allows 32,768 tokens is called GPT-4-32K, which currently restricts access for now. In the near future, this feature may be opened up.

Model parameters become secret

We know that the GPT-3.5 model has 200 billion parameters and GPT-3 has 175 billion parameters, but this was changed in GPT-4.

In the report, OpenAI stated that

Given the competitive landscape and the security implications of large models (such as GPT-4), this report does not contain further details about the architecture (including model size), hardware, training computation, dataset construction, training methods, or the like.

This means that OpenAI did not disclose any more about the size of the GPT-4 model, the number of parameters, or the hardware used.

OpenAI says this move was made in light of concerns about competitors, which may be a hint of its strategy towards a competitor, Google Bard.

In addition, OpenAI mentions "the security implications of large models," which, although not further explained, also alludes to the more serious issues that generative AI may face.

Selective expression of "excellence"

With the introduction of GPT-4, we have all seen how much better the model is than its predecessor.

GPT-4 scored in the top 10% of test takers on the mock bar exam; by comparison, GPT-3.5 scored in the bottom 10%. But this is actually a trick by OpenAI - it only shows you the best part of the GPT-4, while more secrets are hidden in the report.

The chart below shows the performance of GPT-4 and GPT-3.5 on some exams. As you can see, the GPT-4 does not perform that well on all exams, and the GPT-3.5 does not always suck.

Improved "prediction" accuracy

Since the launch of ChatGPT, we all know that the model is often "serious nonsense", giving a lot of arguments that seem to make sense but don't really exist.

Especially when it comes to predicting certain things, the fact that the model has data from the past leads to a cognitive bias called "hindsight", which makes the model quite confident in its predictions.

In the report, OpenAI says that the accuracy of the model should have declined as the size of the model increased, but GPT-4 reversed this trend, with the graph below showing that the accuracy of the prediction improved to 100.

On March 14, OpenAI released GPT-4, dropping another "nuclear bomb" on the tech world.

According to OpenAI's demo, we know that GPT-4 has more power than GPT-3.5: summarizing articles, writing code, filing taxes, writing poems, etc. But if we go deeper into OpenAI's technical report, we may also find out about GPT-4.

But if we go deeper into the technical report published by OpenAI, we may also find more features about GPT-4 ......

As well as some details that OpenAI doesn't name and proclaim that might be a chill in the back.

The new Bing is loaded with GPT-4

Naturally, when GPT-4 was released, the new Bing was also loaded with the latest version.

According to a tweet by Microsoft Bing VP Jordi Ribas, the new Bing loaded with GPT-4 has upped the Q&A limit to 15 questions at a time, with a maximum of 150 questions a day.

Eight times longer text length

On GPT-4, the text length has been significantly increased.

Previously, we knew that GPT API calls were charged by "token", and one token usually corresponds to about 4 characters, while one Chinese character is roughly 2~2.5 tokens.

Before GPT-4, the token limit was around 4096, which was equivalent to 3072 English words. Once the length of the conversation exceeded this limit, the model would generate incoherent and meaningless content.

However, by GPT-4, the maximum number of tokens is 32,768, which is equivalent to about 24576 words, and the text length is expanded by a factor of eight.

That is, GPT-4 can now answer longer texts.

OpenAI says in the documentation that the context length limit for GPT-4 is now limited to 8192 tokens, and the version that allows 32,768 tokens is called GPT-4-32K, which currently restricts access for now. In the near future, this feature may be opened up.

Model parameters become secret

We know that the GPT-3.5 model has 200 billion parameters and GPT-3 has 175 billion parameters, but this was changed in GPT-4.

In the report, OpenAI stated that

Given the competitive landscape and the security implications of large models (such as GPT-4), this report does not contain further details about the architecture (including model size), hardware, training computation, dataset construction, training methods, or the like.

This means that OpenAI did not disclose any more about the size of the GPT-4 model, the number of parameters, or the hardware used.

OpenAI says this move was made in light of concerns about competitors, which may be a hint of its strategy towards a competitor, Google Bard.

In addition, OpenAI mentions "the security implications of large models," which, although not further explained, also alludes to the more serious issues that generative AI may face.

Selective expression of "excellence"

With the introduction of GPT-4, we have all seen how much better the model is than its predecessor.

GPT-4 scored in the top 10% of test takers on the mock bar exam; by comparison, GPT-3.5 scored in the bottom 10%. But this is actually a trick by OpenAI - it only shows you the best part of the GPT-4, while more secrets are hidden in the report.

The chart below shows the performance of GPT-4 and GPT-3.5 on some exams. As you can see, the GPT-4 does not perform that well on all exams, and the GPT-3.5 does not always suck.

Improved "prediction" accuracy

Since the launch of ChatGPT, we all know that the model is often "serious nonsense", giving a lot of arguments that seem to make sense but don't really exist.

Especially when it comes to predicting certain things, the fact that the model has data from the past leads to a cognitive bias called "hindsight", which makes the model quite confident in its predictions.

In the report, OpenAI says that the accuracy of the model should have declined as the size of the model increased, but GPT-4 reversed this trend, with the graph below showing that the accuracy of the prediction improved to 100.