Summary of Key Points
This article focuses on the resignation of Demis Hassabis, the founder of DeepMind, and covers the entire journey of DeepMind from its inception to its acquisition by Google, the competition with rivals such as OpenAI in terms of technological approaches, and Hassabis' eventual shift towards AI in pharmaceuticals. It reveals the early financing challenges in the AI industry, the divergence of technical paths (reinforcement learning vs large language models), the strategic games among giants, and Hassabis' personal transformation from a "dreamer of AGI" to an "entrepreneur in pharmaceuticals," showcasing the tumultuous developments in the AI sector and the rationale behind his personal choices.
I. The Birth of DeepMind: From a Game Studio to the Financing Journey for AGI Dreams
Hassabis did not start out in AI. His first business venture was a game studio called "Alphabeta," which earned him several million pounds and allowed him to achieve financial freedom. In 2010, when AI was still in its infancy (the ImageNet dataset had not yet been created, and no one was optimistic about the commercialization of AI), he founded DeepMind with the goal of achieving "Artificial General Intelligence (AGI)"—making machines possess general intelligence similar to that of humans.
Early financing was extremely difficult:
- He met Peter Thiel at the "Singularity Summit" (which was like a gathering of AI enthusiasts) and exchanged half of his company for $2.3 million (like buying a ticket to enter the AGI arena, somewhat like playing the lottery);
- Through Thiel's introduction, he received an investment of $5 million from Elon Musk (who valued AI security and believed Hassabis was a "responsible person");
- During the Series C financing round, Zhou Kaixuan from Hong Kong (representing Li Ka-shing) invested $13.6 million.
However, Hassabis grew tired of the constant need for funding and ultimately chose to be acquired by Google: Google offered $650 million and promised extreme autonomy (allowing him to keep his London-based team and establish an ethics committee). Google saw potential in him because Larry Page recognized his abilities through AI expert David Silver and even advised him, "If your goal is AGI, why not use Google's resources?"
II. The Competition of Technical Paths: DeepMind's Reinforcement Learning vs OpenAI's Large Language Models
DeepMind and OpenAI represented two core approaches in the early days of AI:
- DeepMind's Reinforcement Learning: This approach allows AI to learn through trial and error (for example, AlphaGo learned by playing chess and adjusting its strategy after losses). Hassabis believed that language was not practical enough to directly reflect the real world, and reinforcement learning was the correct path to AGI. Early successes such as AlphaGo (defeating human chess players) and AlphaFold1 (protein folding) were examples of reinforcement learning's power.
- OpenAI's Large Language Models: Led by Ian Goodfellow (whom DeepMind tried to recruit), this approach argued that language is a compressed version of the world—if a model can predict the next word, it indicates an understanding of the underlying principles. OpenAI initially used Transformer technology (invented by Google Brain) to develop its GPT series, which gradually demonstrated versatility from GPT-1 to GPT-3.
These two paths eventually converged: ChatGPT combined pre-training with fine-tuning and Reinforcement Learning via Human Feedback (RLHF), combining the versatility of large language models with the optimization capabilities of reinforcement learning, sparking an AI revolution. DeepMind only began to catch up with large language models in 2020 but did not release its products to the public (such as the Sparrow chatbot) early on, thus missing out on the opportunity.
III. Google's Integration and the AI Race: From a Surprising Attack to Gemini's Counterattack
The release of ChatGPT in 2022 caught both Google and DeepMind off guard:
- Google hastily launched the Bard chatbot, which was ridiculed for factual errors, resulting in a loss of billions of dollars in market value;
- To address the crisis, Google merged Google Brain with DeepMind to form "Google DeepMind (GDM)" in 2023, with Hassabis taking on the role of CEO (essentially, DeepMind absorbed Google Brain).
GDM launched the Gemini series of models, but their initial response was lukewarm:
- In February 2024, Gemini 1.5 Pro (with a million-level context) was released, yet it lost momentum to OpenAI's Sora video model;
- It wasn't until the end of 2024 that Gemini gained industry recognition, only to be soon surpassed by new developments from Anthropic and OpenAI. The planned Gemini 3.5 Pro was delayed.
By this time, Hassabis had grown tired of the competition and officially announced his transition to Isomorphic Labs, an AI pharmaceutical company he founded in 2021, aiming to use AI to transform drug development.
IV. Hassabis' Transition: From an AI Laboratory to a New Frontier in Pharmaceuticals
Hassabis' shift was not sudden:
- Early collaborations with the UK NHS on medical AI projects (which failed due to privacy concerns) sparked his interest in healthcare;
- He believed that AI in pharmaceuticals was an extension of AGI—using AI to solve human diseases was more meaningful than mere technological competition.
Isomorphic Labs today is similar to DeepMind back then: no one fully understands its potential, but Hassabis believes this path can create a new paradigm and even "eradicate all diseases." His resignation is not a departure from AI; rather, it marks a shift to a field that is more closely aligned with human well-being.
V. The Butterfly Effect of the AI Industry: How DeepMind Inspired Rivals' Growth
The existence of DeepMind indirectly led to the emergence of several other AI giants:
- OpenAI: Disappointed by DeepMind's acquisition by Google, Elon Musk initiated OpenAI (with a non-profit focus to counter commercial companies controlling AGI);
- FAIR: After being rejected by Hassabis, Mark Zuckerberg hired Yang Liqun to establish Facebook AI Research, setting up a rival to DeepMind;
- Microsoft's Investment in OpenAI: Microsoft's CTO admitted they had been left behind by Google/DeepMind and bet on OpenAI, later using ChatGPT to overtake Google in search rankings.
DeepMind acted like a "catfish" in the AI industry, forcing rivals to improve and pushing the entire sector from a research laboratory to commercialization.
Conclusion
Hassabis' resignation marks the end of an era for DeepMind, but the story of AI continues. From a game studio to an AGI lab, and then to AI in pharmaceuticals, every step he took was into uncharted territory. Whether Isomorphic Labs is successful or not, his journey teaches us that the value of AI lies not only in technological breakthroughs but also in using it to solve real-world problems for humanity.