 | |
|
Earlier this week, OpenAI unveiled the impressive capabilities of its latest upgrade, GPT-4o ChatGPT. This announcement, marked by a compelling live demo, came just ahead of Google’s I/O 2024, where Google focused on its Gemini AI innovations for its ecosystem of apps and services.
After watching both events, I concluded that GPT-4o emerged as the superior AI assistant. OpenAI's new chatbot boasts significantly enhanced abilities, supporting text, voice, photo, and video inputs, making it a versatile tool.
While Google’s Project Astra matches GPT-4o’s multimodal capabilities, their presentation included only a pre recorded demo of a GPT-4o-like conversation. Additionally, Gemini AI won’t feature similar enhancements until later this year.
Therefore, if you're interested in AI, GPT-4o is the best ChatGPT model available right now. It's accessible to everyone for free, even without a ChatGPT Plus subscription.
However, there is one unexpected downside to GPT-4o: it doesn’t have the most up-to-date knowledge database. Its knowledge cutoff is May 2023, slightly earlier than the December 2023 cutoff for GPT-4.
When OpenAI released GPT-4o on Monday, I briefly tested it and inquired about its knowledge database. It surprisingly cited a cutoff date of September 2021, though it assured me it could browse the web for current information, which I already knew.
 |
| The first response GPT-4o gave me about its training data set. Image source: Chris Smith, BGR |
I suspected it might be an error, given that OpenAI was still in the process of rolling out GPT-4o to various markets and devices. So, I decided to wait and recheck on Friday morning. This time, I posed the same question to all three models.
GPT-4o provided a much more satisfactory cutoff date: May 2023.
 |
| A few days later, GPT-4o offered an improved cutoff date. Image source: Chris Smith, BGR |
Next, I checked GPT-4, and this ChatGPT model provided the most up-to-date cutoff date of the three: December 2023.
 |
| GPT-4 has the best data set of the three ChatGPT models. Image source: Chris Smith, BGR |
Finally, I turned to GPT-3.5, which informed me that its training data only included information up to January 2022. Unlike the other two ChatGPT models, GPT-3.5 cannot browse the web for updates.
 |
| GPT-3.5 is stuck in the more distant past. Image source: Chris Smith, BGR |
This isn’t a significant issue since GPT-4o can search the web for any information I might need. In fact, I haven’t focused on knowledge cutoffs for months. However, it's an interesting detail that sheds light on how OpenAI trains its models.
Although OpenAI describes GPT-4o as "smart and fast," it doesn’t inherit the knowledge database of its predecessors. Instead, it appears the chatbot was trained from scratch.
When OpenAI announced a new privacy tool to prevent the use of copyrighted content in ChatGPT training, they also detailed their training process. This revealed that each model undergoes a fresh training cycle.
Each new generation of foundation models is trained from scratch using a new dataset. We continually enhance our architecture and significantly expand the scale and diversity of our datasets, surpassing those of previous models.
This explains why GPT-4o’s knowledge cutoff differs from that of GPT-4. I anticipate that OpenAI will update the training cutoff, as it has done with other models in recent months.
 |
| ChatGPT can generate responses for the same prompt using any of the three models. Image source: Chris Smith, BGR |
One interesting feature in the redesigned ChatGPT. ChatGPT web interface is the ability to choose which model answers your question. As shown in the screenshots above, you can use the dropdown menu to select between GPT-4o, GPT-4, and GPT-3.5, allowing you to compare how each model responds to your prompts.