OpenAI’s GPT-4o ChatGPT upgrade, announced earlier this week, was an amazing display of power. The company dropped a big AI live demo ahead of Google’s I/O 2024, where Google would talk only about Gemini AI innovations for its ecosystem of apps and services.
After watching both events, I thought GPT-4o won the AI assistant battle. OpenAI can offer significantly improved AI assistant abilities in ChatGPT right now. The chatbot can support text, voice, photo, and video inputs.
Google can match the multimodality of GPT-4o with Project Astra. But Google had a prerecorded demo of a GPT-4o-like conversation with AI. And Gemini AI won’t get similar features until later this year.
In other words, if you’re interested in AI, you should be using GPT-4o on your devices right now, as it’s the best ChatGPT model that OpenAI has to offer the public right now. Plus, it’s free for everyone, even if you don’t pay for ChatGPT Plus.
But there is one downside to GPT-4o that I didn’t expect. OpenAI didn’t give its latest and greatest model the most extensive knowledge database. GPT-4o’s knowledge cutoff is May 2023, which is slightly worse than the December 2023 cutoff for GPT-4.
When OpenAI made GPT-4o available on Monday, I briefly tested the new ChatGPT model and asked it about its knowledge database. At the time, it gave me a puzzling September 2021 date. It reassured me that it could browse the web for up-to-date data, which I already knew.
I thought it might be a mistake, as OpenAI was still rolling GPT-4o out to various markets and devices. I decided to wait and then asked the same question on Friday morning. This time around, I asked all three models.
GPT-4o offered a much more palatable cutoff date: May 2023.
I moved on to GPT-4, and the ChatGPT model offered the most up-to-date cutoff date of the three: December 2023.
GPT-3.5 was the last option on the list, and the model told me its training data had information up to January 2022. Unlike the other two ChatGPT models, GPT-3.5 can’t search the web for me.
This isn’t necessarily a big deal. After all, GPT-4o can search the web for anything I might need. In fact, I haven’t paid attention to knowledge cutoffs for months. But it’s still an interesting detail, as it gives us further insight into how OpenAI trains its models.
GPT-4o might be “smart and fast,” as OpenAI describes it, but it doesn’t inherit the knowledge database of its previous models. Instead, the chatbot might have been trained from the ground up.
When announcing a new privacy tool to prevent copyrighted content from being used for ChatGPT training, OpenAI also explained how it trains its models. That’s where we learned each model goes through a fresh training process:
Each new generation of foundation models is trained from scratch on a new dataset. We constantly improve our architecture and increase the scale and diversity of our datasets significantly beyond our previous models.
This would explain why GPT-4o’s knowledge cutoff doesn’t match GPT-4. I expect OpenAI to improve the training cutoff, just like it did with the other models in the past few months.
I’ll also point out one interesting feature in the redesigned ChatGPT UI on the web. As you can see in the screenshots above, you can choose which model to answer your question. Just use the dropdown menu and select between GPT-4o, GPT-4, and GPT-3.5 to see how each responds to your prompts.
