#Gemini vs DeepSeek: Google event post report

This is my post about attending a Google event: Gemini vs. DeepSeek. It was an interesting experience, and the reason I wanted to share this post is to give insights into top-tier tech events and encourage everyone to explore state-of-the-art technology.

#About the Deepseek

For anyone who have not to found deepseek yet, this is AI model from research-focused company called Deepseek.

#AI & Multi-Agent Models

We started by discussing how Large Language Models (LLMs) could become cheaper than hiring humans.

One example we explored was in legal disputes. Imagine a scenario where the outcome of a lawsuit is determined by the quality of an AI lawyer—the better the AI, the higher the chances of winning. This could lead to massive disruptions in finance, law, and other industries, potentially making legal authority unfair.

A thought-provoking scenario we discussed was whether AI lawyers could eventually replace human lawyers altogether, leading to a world where legal disputes are entirely handled by AI, without human involvement.

#LLM cheaper than Human

But let's face the reality, AI is becoming better and better and less cost effective these days.

There was a case where lawyer cites fake cases generated by ChatGPT in legal brief in early 2023. I think there would be a case where legal AI is adapted to real case, sooner or later.

#Multi-agent: What they talk?

Another business owner introduced the concept of multi-agent AI models, where different AI systems communicate before providing an answer, theoretically leading to better decision-making.

Peter then asked "What do these AI models actually talk about?". We debated whether the conversations between AI models significantly influence decision-making and outcomes.

#LLM talks so we don't have to talk?

Peter also come up with a futuristic idea where we have our AI model instance so that we do not have to talk in person, but our ideas are transmitted fluently.

My thought after this is that if AI have a dream like human see a dream when we sleep. AI has numbers of question answering every seconds. Hypothetically speaking if AI has a rest, do they take a break and see a dream to organize their thoughts as of indexing their memory?

---

#Business Aspects Talk

Here are couple of interesting conversation between Peter and participants.

#Reference to a business in Brazil.

One of the participants, a business owner from Brazil, shared how his company implemented an AI-powered customer support system, reducing 70% of customer interactions with human agents while improving service quality.

#Frustration wins over convincing AI bot

Peter, the host of the event, raised a concern. He shared a personal experience where he was unknowingly charged 20 per month for a service he no longer needed. When he attempted to cancel the subscription, he was stuck in an endless loop with an AI chatbot and never got through to a human.

Frustrated, he eventually gave up and continued paying. This sparked a discussion about whether AI-driven customer support truly improves efficiency or just frustrates users into compliance.

---

#Technical Concepts

#Distillation - big model to train small model.

#Open Source/Closed Source.

We also explored the difference between open-source and closed-source AI models. As we know, ChatGPT (OpenAI) is closed-source, while DeepSeek’s latest model is open-source. We compiled a list of characteristics for each type

Open Source AI ✅	Closed Source AI 🚫
Cheap	Expensive
Worldwide	Restricted
Collaborative	Centralized
Flexible	Controllable
Private	Encapsulated
Transparent	Opaque
Hackable	More Secure
Runs on the edge (local)	Cloud-based (provider-dependent)
User owns data	Dependent on provider
Community-driven	Highly profitable for companies

#Prisoner Dilemma

The Prisoner’s Dilemma is a classic game theory question where two prisoners must decide whether to cooperate or betray each other.

• If both cooperate, they each get 3 points.
• If one defects while the other cooperates, the defector gets 2 points, and the cooperator gets 0 points.
• If both defect, they each get 1 point.

We applied this concept to Gemini vs. DeepSeek.

• Gemini initially chose to collaborate—showing that it is more trustworthy and cooperative.
• DeepSeek immediately chose to defect, seeking a competitive advantage.
• However, after this initial round, both models continued to defect, leading to a final score of DeepSeek = N+2 and Gemini = N.

This experiment highlighted how Gemini appears to be more cooperative, while DeepSeek is more ruthless and strategic.