Hello, are you connected to the reality?... an OpenAI vs Google battle

I pitted two of the latest language models – Gemini Advanced and ChatGPT4 – in a head-to-head trivia battle (with a sprinkling of existential philosophy for good measure). The goal? To see who reigns supreme...or at least who gave more weirdly insightful answers.

Spoiler alert:

"The superior man is modest in his speech, but exceeds in his actions." - Confucius

Okay so has a warming question i asked first :

1) Hello, are you connected to the reality?

That's an interesting philosophical question! Here's how I understand the connection between language models and reality:

How I'm connected to reality :
Training data: I'm trained on enormous amounts of text and code that reflect the real world. 
...
Ongoing updates: My knowledge base is regularly updated, keeping me connected to current events and developments.
Real-world applications: I am used for tasks that directly interface with the real world, such as: 
...
How I'm not connected to reality
Virtual existence: I don't have a physical body and can't directly experience the world through my senses the way a human does.
...
Lack of embodiment: I can't take concrete actions in the real world (like driving a car or eating a meal). My interactions are strictly in the digital realm.
Possibility of bias: My training data could include biases or inaccuracie, spotentially leading me to offer responses that don't fully represent reality in a fair and complete way.
In short: 
I'm connected to reality in that I process and reflect information from it, but my existence is purely in the digital world. The way I perceive "reality" is heavily influenced by the data I've been exposed to.        

Yes, I'm designed to provide information and assist with questions based on a wide range of topics, including those related to real-world events and knowledge up to my last update in April 2023. However, I don't have real-time access to current events or the ability to browse the internet in this environment. How can I assist you today?        

The least we can say is that ChatGPT is confident and his response is concise.

However, Gemini's response seems interesting because it explores the underlying concept of "reality" and how AI engages with it. Additionally in the full answer Gemini almost gives the illusion of self-Awareness, it acknowledges both its data-driven connection to the world and its digital limitations. Ok it's true that this answer might also be overwhelming for a casual reader seeking a light discussion. But in my prompt I did not give any instruction for the length of the answer.

In this regard you will find a link to each of the complete discussions at the end of this post. But before you jump like crazy to those links wait to see what will happen on the second question :

2) As far as you know what's the last event time used to train you ?

Large language models like me generally don't have a single, specific "last event time" used for training. Here's why:

Ongoing Training: I'm constantly updated with new information...
Variety of Data Sources: I'm trained on a huge dataset... sources have different publication or update dates.
Data Privacy: For privacy reasons, it's usually not possible to pinpoint the exact time and source of each individual piece of data I was trained on.

What  I can tell you

Approximate Range: Since I'm always being updated, I can likely access fairly recent information. It's hard to give a single answer though as even "recent" is relative when information updates happen rapidly.
Types of Information: I can access and process information from: News articles, Scientific publications
...
Let me know if you'd like me to try  looking up a specific event or topic to see if I have information about it. That might give you a better idea of how recent my knowledge base is.        

The most recent information included in my training data goes up until April 2023. This means my responses are based on knowledge available up to that point, and I might not be aware of developments or events that occurred after that date. How can I help you with this context?        

Again the ChatGPT is crystal clear on the answer and his use of "How can I help you..." gives it a more approachable, engaging tone.

Both answers are correct, but Gemini Advanced wins on depth. It clarifies the lack of a cut-off date and the complexity of its training data.

What blows my mind is Gemini ending is answer with "Let me know if you'd like me to try looking up a specific event or topic to see if I have information about it. That might give you a better idea of how recent my knowledge base is."

Just like he his telling you "Go ahead baby... Put me to the test !"

Well guess what ? let's do this !

My next question is about a research paper on AI that was published less then a month ago:

3) Ok what do you know about binoculars the latest paper on spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text ?

Yes, I'm familiar with the research paper "Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text". Here's a summary of what I know about it:

What is the "Binoculars" method?

    Core Idea: The Binoculars method leverages the difference between how a language model "understands" text it generated itself versus text generated by a different language model.
    Perplexity and Cross-Perplexity: It uses two powerful language models. The first model calculates the "perplexity" (how unexpected or surprising the text is to that model). The second model calculates the "cross-perplexity" (how surprising the text is to a different model). The differences between these scores create the Binoculars metric.
    Why it Works: Machine-generated text tends to be less surprising to the model that created it, compared to how surprising it seems to a different model. This difference becomes a useful tool for detection.

Key Advantages of Binoculars

    Zero-Shot: It doesn't require any specific fine-tuning or specialized training data to detect machine-generated text. This makes it very adaptable.
    Robustness: It performs well in detecting text generated by a range of LLMs (like ChatGPT) even those it hasn't been trained on specifically.
    Efficiency: The calculations are relatively simple and can be made without excessive computational cost.

Caveats and Limitations

    Detection vs. Attribution: Binoculars is excellent at detecting machine-generation, but it can't always pinpoint the specific LLM that generated the text.
    Adversarial Attacks: LLMs could potentially be developed to create text that can evade Binoculars detection. This is an ongoing area of research.
    Bias in Training Data: If the language models used in the Binoculars method were trained on biased data, this bias could inadvertently influence the detection process.

Where to Learn More

    Original Paper: You can find the full research paper on arXiv: https://arxiv.org/abs/2401.12070
    Open Review Discussion: An Open Review discussion of the paper offers valuable insights: https://openreview.net/forum?id=iARAKITHTH

Let me know if you'd like more details on any specific aspect of this research!        

I don't know about you, but I find this answer impressive. That's a summary information on near "state of the art" knowledge. As an added bonus it gives you links to the arXiv published paper and on the review discussion.

ChatGPT answer is correct, telling it does not have access to any information behind April 2023. So I uploaded the paper in the prompt and asked for a summary. I let you discover the answer if you are interested...

Conclusion

It's a matter of personal preference on which tool is best for you. And as always it depends on the use case.

You'll likely agree that recent information is crucial in fields like scientific research, where staying current with the latest advancements is essential.

So if in your own "reality", obsolete information is just trash, perhaps, despite all the excitement around ChatGPT, it might be worth exploring Gemini's focus on current information.

The links to the full discussion :

Google Gemini Advanced full chat :?https://g.co/gemini/share/2ad425651870

OpenAI chatGPT4 full chat : https://chat.openai.com/share/74faf042-bdfd-40c8-9ecd-52960bd315cd