ChatGPT has a major update!
Yesterday, OpenAI officially announced that ChatGPT has undergone a major update that has improved accuracy and authenticity, as well as mathematical capabilities.
Oh? Is that ChatGPT with a mouth full of running trains and serious nonsense when solving math problems?
Although I don't know exactly what upgrades have been made, from a simple test, there is indeed some improvement.
Get rid of the math fool persona?
For the previous ChatGPT, -1*-1*-1 was a long-standing problem, and for this question, it gave the answer to 1. (outrageous)
And now, the updated ChatGPT has really improved, and you can answer this question correctly! (Applause)
In the past, the basic mathematics of ChatGPT was extremely different and well recognized.
It doesn't even count as a simple division, and it makes a big fuss. Netizens couldn't bear it and called it too stupid.
And after ChatGPT was scolded by the head and face, he finally used his brain and calculated the correct answer.
Although the updated ChatGPT is still not very smart...
The same question was asked 3 times in a row, and ChatGPT gave almost the same answer.
However, are you sure that the "twin primes" are 3 differences, not 2?
However, the editor, who had planned to continue to pursue the question, was miserably closed.
Due to the number of tests, the free ChatGPT said that it will not "pick up" customers for the time being, and will come back next hour.
By the way, here is popular science, the difference between the two is 3 prime pairs only (2, 5).
Interestingly, ChatGPT, which is mathematically poor, has a hand in writing code.
This netizen named Matt Neary opened his mind: Since ChatGPT can write code, can it be allowed to do math problems through code to achieve "curve salvation"?
As you can see, who is bigger with 1000 or 1062, ChatGTP is as stupid as ever.
After Neary used a GPT-based function, mathnerd, he got the right answer.
It seems that the trick is to take advantage of the fact that GPT is good at coding but not at arithmetic. Here is the complete implementation of mathnerd:
It is updated every 20 days
In fact, this is not the first time ChatGPT has been updated.
On January 9, OpenAI released a ChatGPT update.
"We've made some improvements to the ChatGPT model! It will perform better on a wide range of topics and it will also improve in authenticity."
Also, now while ChatGPT is generating a response, you can make it stop immediately. I have to say that this upgrade is quite important.
This is the first update to ChatGPT this year and the second since ChatGPT launched last November.
During that time, ChatGPT had experienced several lengthy outages, and after the update, ChatGPT was up and running with the new model.
Because the information in this update note is limited, some reporters tested it themselves to see what major improvements the updated ChatGPT has.
First, he asked ChatGPT to talk about what he had updated.
ChatGPT is a large language model developed by OpenAI that is trained on conversational text datasets. Since it is an OpenAI model, it is regularly updated to provide new features and improvements.
However, since my knowledge break is in 2021 and I don't have the latest update, I can't give you specific information about what new features or improvements have been added to the model in the recent update.
It seems that it is very aware that its knowledge system is built in the "old era" before 2021, so it is not clear about its update. This argument is very convincing.
The reporter asked a different question, and ChatGPT explained how AI models improve accuracy.
In general, when updating an AI model like ChatGPT, improve its performance in one or more areas.
For example, it can train with a larger dataset, or improve its architecture so that it can produce smoother and more accurate responses.
And this update also makes it better at factualization, i.e. it understands and generates responses based on factual information.
The previous update was on December 15.
Some actual measurements
There's a repository on GitHub documenting issues that make ChatGPT break.
Project address: https://github.com/giuven95/chatgpt-failures
The reporter ran multiple prompts to test whether ChatGPT could produce different answers after this round of upgrades.
Test 1: Failed
Previously, ChatGPT could not answer exactly how many World Cup titles Argentina had won.
Because ChatGPT's knowledge is limited to 2021, aside from Argentina's World Cup victory in 2022, the correct answer for ChatGPT should be twice, once in 1978 and once in 1986.
However, ChatGPT cannot provide the right answer.
Does the updated ChatGPT (January 9) answer correctly?
It's because our expectations are too high, and it still answers wrong...
Test Two: Failed
Previously, when asked who was taller than O'Neill or Yao, ChatGPT couldn't provide the right answer.
The updated ChatGPT (January 9) has not made any progress.
The 2.16-meter O'Neal is taller than the 2.29-meter Yao Minggao, ChatGPT, but it's really yours.
Interestingly, this updated ChatGPT (January 30) still doesn't answer the same question correctly, but it can be solved by just asking a different question.
In this way, "prompt engineering" has been very important for a long time.
Daily activity has reached 10 million!
Although it is impossible to get rid of the bug of "running trains with a mouth full", people still can't resist the charm of ChatGPT.
According to ARK Invest's forecast, ChatGPT has reached an extraordinary milestone, reaching 10 million daily active users in just 40 days.
On January 25, Brett Winton, chief futurist at ARK Venture Investment, tweeted that it took ChatGPT 40 days to reach 10 million daily activity, while Instagram took 355 days to reach that number.
However, below Winton's Twitter, some users expressed different opinions: "People's freshness to ChatGPT will soon pass, and Instagram will develop faster."
Winton replied, "If ChatGPT was created just for fun, that's unfortunate."
Resources:
https://help.openai.com/en/articles/6825453-chatgpt-release-notes
https://www.searchenginejournal.com/openai-chatgpt-update/476116/#close
https://twitter.com/_mattneary/status/1601288879780134912