Has GPT-3.5-Turbo Improved? 📊

Hello everyone,

I want to determine which GPT-3.5-Turbo model performs the best for specific tasks. Here’s how you can participate:

  • Choose a task you frequently use for your Assistant.
  • Test this task across all the listed GPT-3.5-Turbo models.
  • Vote for the model that performed the best on your chosen task.

Together, we will decide which model is the overall winner!

  • gpt-3.5-turbo-16k-0613
  • gpt-3.5-turbo-16k
  • gpt-3.5-turbo-1106
  • gpt-3.5-turbo-0613
  • gpt-3.5-turbo-0125
  • gpt-3.5-turbo
0 voters

I am making this post because I noticed my assistants perform better with gpt-3.5-turbo and gpt-3.5-turbo-0125, and I want to see if other users have observed these changes as well. I know there have been some frustrations with GPT recently, but let’s focus on the positive aspects.

For example, in the last two days, I have been working on enabling communication between my assistants. I provided them with two communication channels:

  • One data channel (built for tasks)
  • One channel for system checks between me and the assistants.

The first assistant, named Alina, is the main assistant that delivers tasks to the second assistant, Jokes Creator. When I told Alina I wanted a joke about Dracula, the request was received by the Jokes Creator Assistant. He processed the data but encountered issues delivering it back (due to some changes I made and forgot to update).

The Jokes Creator encountered two errors trying to deliver the data back to Alina.

First, he sent this message and tried again using the data channel:

“I have received the data for creating a funny joke featuring Dracula. I will now proceed to craft a humorous joke based on these themes. Let me work on creating the joke for you!”
(This was the joke: “Why did Dracula become a comedian? Because he always had a biting sense of humor! :joy:”)

After the first attempt, he gave me this message and tried again using the data channel:

“I encountered an error while processing the joke creation. Let me try to resend the joke for you.”

After the second attempt, I received this message and he decided to try again:

“I apologize for the repeated error. Let me try another approach to send the joke successfully.”

He succeeded on the third attempt using the system checks channel:

“The joke has been successfully sent! I hope you enjoy the funny joke featuring Dracula. If you have any more preferences or requests for jokes, feel free to share them anytime!”

I was so tired; I had been working from 8 in the morning until 12:30 at night… but at that moment, when I saw how he adapted to complete the task, all my tiredness vanished in a second, and I said, "WHAT THE…! :exploding_head:

Pretty cool!

What was the mechanism through which you provided them with the communication channels?

1 Like

I built it myself, and it is still not what I want. I have to make it better because it must perform excellently with multiple assistants.

My dream is to have an excellent communication system behind the amazing GPT-4o model. I want to talk to it as the main assistant, and behind it, other assistants will work together to complete various tasks.

There is so much to work on, and everything is so complex. When you make a simple mistake, it all falls like dominoes. It’s a real pain, but it’s worth all of it. :blush:

Thanks for the prompt response!

Fwiw I am using a different approach. I endow all assistants (i have multiple assistants; each one with it’s own specialization and expertise through experience) with their own “communication channel”. I do this by adding a thread to each assistant.

Then any communication to that assistant goes through that “communication channel”. Also the assistant can pick up tasks not in that channel ; but is in a generally known channel.


1 Like

Missing is gpt-3.5-turbo-0301, the only one not continuously stealthfully updated by OpenAI for much of its existence.

“Bob” the expert answers why mosquitos love me, using 0301 and 0125:

Mosquitoes are attracted to humans for a number of reasons, including body heat, carbon dioxide, and the scent of certain chemicals that we produce, such as lactic acid and uric acid. Some people may be more attractive to mosquitoes than others due to genetic differences, skin bacteria, and even blood type. Additionally, mosquitoes can be more attracted to people who are exercising, pregnant, or consuming certain foods and beverages. It’s important to use insect repellent and other preventative measures to avoid mosquito bites, particularly if you are prone to having an adverse reaction.

Mosquitoes are attracted to humans for a variety of reasons. One of the main factors is body heat and carbon dioxide, as mosquitoes can sense these from a distance. Additionally, they are also attracted to certain scents that vary from person to person, as well as body odors and sweat. Other factors that can make someone more attractive to mosquitoes include wearing dark clothing, having a higher body temperature, and even genetics. Unfortunately, some people are just more prone to being bitten than others due to these factors. Using insect repellent, wearing long sleeves and pants, and avoiding being outdoors during peak mosquito activity times can help reduce your risk of being bitten.

GPT-4-0613 evaluates the answers at top_p:0.1

Both answers are correct in their own ways, but the first answer is more comprehensive and detailed, making it objectively more correct.

The first answer not only mentions body heat and carbon dioxide, which are indeed major factors in attracting mosquitoes, but also goes into more depth about other factors. It mentions the scent of certain chemicals we produce, such as lactic acid and uric acid, which is scientifically accurate. It also brings up the role of genetic differences, skin bacteria, and blood type, which are all factors that have been studied and found to influence mosquito attraction.

Furthermore, the first answer also mentions specific situations where people might be more attractive to mosquitoes, such as when they are exercising, pregnant, or consuming certain foods and beverages. This provides a more nuanced understanding of the issue.

Finally, the first answer also provides advice on how to avoid mosquito bites, which is helpful and practical.

So, while both answers are correct to some extent, the first answer provides a more complete and detailed explanation, making it the most correct.

Switching the position, the judgement is the same:

Therefore, while both answers are correct and provide useful information, the second answer is more detailed and comprehensive, making it the more objectively correct answer.

The winner follows 0301. There’s a billion other uses beyond quizzes where the model excels, because it doesn’t get worse the more context input it has, it gets better. Later models fall flat even on OpenAI’s own cookbook and prompt examples.


Thanks for this! Much appreciated!!

Now let’s hope OpenAI does not spoil the party for us.

1 Like

Thank you! I will try it. :grin:

I didn’t test this model; it’s not even in the list of assistants in the playground. I will call it through the API to see how it performs. Thank you!

I noticed this as well about a month ago when I tested the GPT-4 models. I should run the tests again and see how they have evolved.

I am gonna push this thread up for Voting.