GPT-4o is the best ChatGPT model, but it has one weakness

OpenAI GPT-4o ChatGPT UpgradeIt was a surprising display of power announced earlier this week. The company has canceled a large-scale AI live demo ahead of Google I/O 2024. So Google will only talk about Gemini AI innovations in the app and services ecosystem.

After watching both events, I thought, GPT-4o wins the AI ​​assistant battle. OpenAI can provide significantly improved AI assistant capabilities. Chat GPT right now. Chatbots can support text, audio, photo, and video input.

Google can match GPT-4o’s multimodality with Project Astra. However, Google had prerecorded a demo of a GPT-4o-like conversation with an AI. And Gemini AI won’t have similar functionality until later this year.

In other words, if you’re interested in AI, you need to use GPT-4o on your device today. Because GPT-4o is the best ChatGPT model that OpenAI can offer to the public at this time. Plus, it’s free for everyone, even if you don’t pay for ChatGPT Plus.

However, GPT-4o has one drawback that I didn’t expect. OpenAI did not offer the most extensive knowledge database for the latest and greatest models. The knowledge cutoff for GPT-4o is his May 2023, which is slightly worse than his December 2023 cutoff for GPT-4.

When OpenAI made GPT-4o available on Monday, I briefly tested the new ChatGPT model and asked about its knowledge database. At the time, it gave me a mysterious date of September 2021. I was relieved to be able to view the latest data I already knew on the web.

Since OpenAI was still rolling out GPT-4o to various markets and devices, I thought this might be a mistake. I decided to wait and asked the same question on Friday morning. This time we asked all three models.

GPT-4o had a more favorable deadline of May 2023.

When we transitioned to GPT-4, the ChatGPT model provided the most up-to-date deadline of the three models: December 2023.

GPT-3.5 was the last option on the list, and the model told me that its training data had information up to January 2022. Unlike the other two ChatGPT models, GPT-3.5 cannot search the web.

This isn’t necessarily a big deal. After all, GPT-4o can search the web for everything you need. In fact, I haven’t been paying attention to the limits of knowledge in recent months. However, this is still an interesting detail as it provides further insight into how OpenAI trains the model.

GPT-4o may be “smart and fast,” as OpenAI puts it, but it does not inherit the knowledge database of previous models. Instead, the chatbot could have been trained from scratch.

As we announce new privacy tools to prevent the use of copyrighted content Chat GPT trainingOpenAI too explained How to train the model. Now we know that each model goes through a new training process.

Each new generation of foundational models is trained from scratch on a new dataset. We are constantly improving our architecture and significantly increasing the size and diversity of our dataset over previous models.

This explains why the knowledge cutoff of GPT-4o does not match GPT-4. We hope that OpenAI will improve our training cutoffs, as we have done with other models over the past few months.

I would also like to point out one interesting feature of the redesigned ChatGPT UI on the web. As you can see from the screenshot above, you can choose the model that answers your question. Simply select GPT-4o, GPT-4, or GPT-3.5 using the drop-down menu to see how each responds to the prompts.