AI Talk: A Brief History of the AI Revolution

This blog is a book review: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence, by Sebastian Mallaby. The book is a biography of Nobel laureate Demis Hassabis, founder of DeepMind, an artificial intelligence company.

ChatGPT generating an image from my website on how I possibly look reading the book on my iPhone!

Why does my title say something else? The book covers the evolution—or shall I say, revolution—of AI that happened over the past decade. It covers it from the perspective of one of the famous actors in this saga—Demis Hassabis. It also touches on the various cast of characters in AI who have shaped, and are being shaped by, this revolution as we speak.

When I finished reading the book last week, my only reaction was—Wow! A book that covers a parallel universe that I didn’t know anything about as an AI researcher—the same events that I am very familiar with, but from a totally different perspective.

A word about the author—Sebastian Mallaby. He is a journalist who has written extensively about finance and technology. He has written half a dozen books, and The Infinity Machine, published this spring, is his latest. For this book, he interviewed Demis Hassabis and a cohort of characters in the AI race over a period of three years.

The early years of Hassabis expose a chess genius, excelling in chess tournaments from a very tender age. He interns at a London gaming company, developing complex multiplayer games at 17! From a very young age, he appears driven, working insane hours. He refuses half a million pounds to work on the next game and instead pursues a Cambridge education at 18. He is obsessed with understanding intelligence. He gets a degree in computer science and then a PhD in neuroscience—the rationale being, you have to understand how the brain works to develop intelligence in computers!

DeepMind Origins

From an early age, Hassabis’s focus on intelligence morphed into a quest for AGI. He founded DeepMind with a couple of kindred spirits in 2010. Peter Thiel provided the seed funding of $20 million. And how did they manage to locate Thiel? In Silicon Valley? By attending the Singularity Summit—a gathering of technology dreamers buying into Kurzweil’s vision of AGI.

The goal from day one was to develop superintelligence! And to give it away to society! It is a miracle such a proposition got funded! Their first foray into codifying intelligence—unsurprisingly—was in the gaming arena. Given Hassabis’s early exposure to complex computer games, the DeepMind team decided to develop computer software to play Atari games. Not any Atari game, but ALL games. They perfected a reinforcement-learning algorithm that learns by playing and eventually beat every one of these games. They called this Deep Q-Network (DQN). This was in December 2013. In parallel, the whole world was buzzing with the advancement of neural networks—the ImageNet competition had a new entrant from Hinton’s team in Toronto that blew away the competition. AlexNet—the solution from Alex Krizhevsky, Ilya Sutskever, and Geoff Hinton—started the turning point in neural network adoption.

The architecture used by Hassabis had nothing to do with neural networks. He was a strong believer in reinforcement learning. And for a while, these two streams of technology were evolving on their own. In 2014, Hassabis managed to sell DeepMind to Google for a handsome price of 400 million pounds. But what is more surprising is the stance to keep the research team in London and to compel Google to establish safety and ethical guardrails in this quest for AGI. A team focused on AGI without worrying about any commercial products or solutions.

Go, Go saga!

The next challenge that Hassabis and team went after was the ancient Chinese game of ‘Go’. The game has an insane number of possible moves—2.1 × 10^170. Now, instead of relying just on reinforcement learning as a strategy, they also incorporated deep learning neural networks into the mix. They incorporated supervised learning of a neural network based on a corpus of 150,000 games played by experts. By combining board-position assessment using a neural network with more traditional tree-pruning strategies and reinforcement learning, their version of a Go annihilator, dubbed AlphaGo, was shaping up. And in March 2016, Google challenged one of the world’s best Go players, Korean champion Lee Sedol, in South Korea. And it trounced him. IBM’s Deep Blue built its chess solution using traditional AI—DeepMind’s Go system was a triumph of the new-age AI. The DeepMind team didn’t rest there. Using reinforcement learning, they set up an experiment to let a novice model, called AlphaZero, learn by playing against itself. This new version, after continuously playing against itself, became the world champion in Go.

Foray into Health

Suleyman, Hassabis’s second in command and the voice of ethics, convinced the National Health Service in the UK to allow them to use AI to help operations. Laudable goal. But as a healthcare research scientist for over three decades, I know nothing is easy in healthcare! It is not technology—it is process, it is collaboration, it is convincing a whole lot of people involved—from patients, to nurses, to providers, to administration, to bean counters. And unsurprisingly, the technology solutions they created, including predicting acute kidney injury (AKI), worked great—but the project ultimately failed.

The Transformer enters the scene

In the fall of 2017, the landmark Transformer paper from Google shook up the whole world. Hassabis was focused on ethics guardrails, convinced their approach to Go was the path to superintelligence. He didn’t initially see the impact of the Transformer architecture. OpenAI, funded by Musk and others, was an answer to Google’s lead in AI. OpenAI saw the potential of the Transformer architecture, came up with a simplified version, and released its first model—Generative Pre-trained Transformer (GPT). The language model was born—though not large at that time. Then came GPT-2, using 1.5 billion parameters. It was already showing impressive results. A few researchers in DeepMind noticed the advance and were trying to convince Hassabis and others to take it seriously. But they didn’t—not at that time.

Intrigue and merry-go-round

The next few years saw a merry-go-round of key researchers shuttling between Google and OpenAI and other startups. Hassabis wanted to focus on AGI research. The application side of DeepMind (namely the health arena) was brought over to join the Google mothership in Silicon Valley.

AlphaFold

Buoyed by AlphaGo’s success, the DeepMind research team pivoted to investigate the next grand challenge—how to predict the 3D structure of a protein from a sequence of amino acids. Knowing the structure can accelerate drug discovery and help us understand causes of diseases like Alzheimer’s and Parkinson’s. This problem in biology has been around for over fifty years. DeepMind now incorporated a bit of Transformer architecture into the overall solution to solve the challenge. I was surprised to see Bidirectional Encoder Representations from Transformers (BERT) playing a part. We experimented with BERT when it first came out in 2018. Also referred to as a Masked Language Model, it is great at predicting missing pieces in sequences. We used it to disambiguate the meaning of words in context—like ‘cold’ meaning something totally different depending on the context. For Google, it was amino acids. They participated in the Critical Assessment of Techniques for Protein Structure Prediction (CASP)—an ongoing computational challenge that started in 1994. Google came first in the challenge in 2018—CASP 13—with its system to predict folds. In CASP-14, two years later, they blew the competition away with 90+% accuracy in prediction. That feat got Demis Hassabis and John Jumper (co-creator of the AlphaFold software) Nobel Prizes in 2024. A few years later, Google released over 66 million protein structures into the public domain. Anyone can now download this content for free. A true gift to mankind.

DeepMind enters the LLM race

In May 2020, OpenAI released the 175-billion-parameter GPT-3. I remember playing with it back then and marveling at all it could do. I also remember reading the paper: “Language Models are Few-Shot Learners.” GPT-3 was the catalyst for DeepMind to finally sit up and notice that they were sorely behind. On to the races, to build the next great model. ‘Gopher’ was born at DeepMind. Released the next year, it had 280 billion parameters. Scaling was the mantra chanted by all leading labs—in terms of the amount of data used for training, the amount of compute used, and the amount of time spent training the models. Then models started becoming multimodal. OpenAI released DALL-E, an image-generation model. DeepMind released Chinchilla, a smaller model that performed as well as the larger model but was trained on much more data for longer. DeepMind also started work on a conversational agent called Sparrow.

ChatGPT changes the race

November 2022 saw the release of ChatGPT. The use of the chatbot skyrocketed. Researchers and laymen alike were gobsmacked. Of course, the last three years have seen a myriad of model releases from OpenAI, Anthropic, and Google. But what was evident from reading Mallaby’s recounting of events was a great deal of focus on safety in the DeepMind camp. Anthropic founder Dario Amodei left OpenAI to focus on safe AI. Jailbreaking techniques were constantly spotlighting problems in chatbots. Hallucinations are evident from these models. Google releases Bard, and takes a financial hit as the model hallucinated. Multimodal models are now commonplace. The race changed when China entered the fray. DeepSeek-R1, released as open source with all the weights in January 2025, shocked the world—and caused Nvidia stock to drop like a stone. It recovered shortly, though. Meta had released its model as open source, but its capability lagged the frontier models. DeepSeek attempted to change these dynamics.

Ongoing race

The quest for AGI continues. Yann LeCun doesn’t believe LLMs are a path to AGI. He has another startup going after different architectures to reach that goal (more on this in the next blog). Hassabis has given up on believing that reinforcement learning (RL) alone can reach AGI. He now believes a combination of RL, neuroscience research, and LLMs are the core elements.

Just last month, in a major shuffle of leadership at Google, Hassabis stepped down from leading AI at Google to focus directly on AGI research. And safety.

Concluding thoughts

Sebastian Mallaby’s book The Infinity Machine is a fascinating read. I thoroughly enjoyed reliving the events of the past decade and a half. And I am excited to see what is going to happen in the next decade. Hassabis just turned 50 and he certainly has a long runway left to go after his AGI goal.

Acknowledgements

No AI was used in actually writing this blog post. I did, of course, use Google for research, and most of the content is paraphrasing events presented in Mallaby’s book. ChatGPT did a quick editorial review for grammar and spellings. And it generated a remarkable photo of me reading the book!