DeepMind says its next AI model Gemini will eclipse ChatGPT

Google is digging into its greatest hits to take down ChatGPT.

Demis Hassabis, CEO of Google DeepMind, just revealed that the company is developing a new flagship AI system dubbed Gemini. The goal? Eclipse OpenAI's GPT-4 by blending large language models with the breakthrough techniques that made AlphaGo legendary.

Supercharging LLMs with AlphaGo's brain

On the surface, Gemini looks like standard fare: a massive language model handling text, similar in scale to GPT-4. But underneath, DeepMind is attempting a fundamentally different architectural bet.

Hassabis says the system will combine the linguistic mastery of state-of-the-art transformers with the reinforcement learning and tree search methods pioneered in AlphaGo. Back in 2016, AlphaGo stunned grandmasters because it didn't just memorize past games; it learned to explore branching possibilities, evaluate future states, and plan out winning sequences.

If DeepMind pulls this off, Gemini won't just predict the next plausible token. It will actively plan, deliberate, and solve multi-step problems, moving AI from passive text generators toward proactive agents capable of handling complex internet and software workflows.

DeepMind is also leaning into its broader scientific bench. Having just demonstrated an algorithm that controls various robotic arms, Hassabis hinted that multimodal learning from physical world interactions could help bypass the traditional dead ends of text-only models.

A multi-million-dollar counterpunch

Building Gemini is neither fast nor cheap. Hassabis says development will take several more months and could easily cost tens or hundreds of millions of dollars. For context, OpenAI CEO Sam Altman noted in April that GPT-4 cost north of $100 million to train.

This heavyweight project follows Google's sweeping April restructuring, which merged Hassabis' DeepMind with the internal Google Brain lab to form Google DeepMind. As Hassabis boldly put it, 80 to 90 percent of recent AI breakthroughs originated from one of those two camps.

Addressing safety fears, Hassabis dismissed calls for a blanket development pause as impractical to enforce. Instead, he argued for urgent research into frontier evaluation tests, adding that he wants academic researchers to get early access to these boundary-pushing models before deployment.

Why it matters

For developers and everyday users, Gemini represents the first serious attempt to solve the fundamental ceiling of current LLMs: planning and reasoning. If Google successfully bolts AlphaGo-style search trees onto a chatbot, we could see hallucinations plummet and autonomous coding or browsing agents finally become reliable enough for production work. That said, it is still months away, and access will likely trickle out through Google's corporate footprint first.

For the broader AI ecosystem, this officially ends OpenAI's uncontested reign at the frontier. While OpenAI showed how far sheer scale and human feedback could carry next-token prediction, DeepMind is leaning hard into its deep reinforcement learning moats. It forces the entire industry to shift focus from just scraping more text to rethinking agentic problem-solving.

Strategically, Google looks smarter for ending the internal rivalry between Brain and DeepMind to marshal all resources behind one flagship. The big question mark is viability. AlphaGo mastered a board game with crisp rules and a clear win-loss condition. Applying Monte Carlo tree search and RL to the messy, ambiguous world of human language could trigger astronomical inference costs that might be too heavy even for Google's servers to bear cheaply.