Google DeepMind

• 1. Introduction

  ›

  Google DeepMind began not with a grand corporate vision but with a simple experiment: teach an AI to play Breakout. Demis Hassabis, Shane Legg, and Mustafa Suleyman founded the company in November 2010 in the United Kingdom, and their early method was to drop an AI into old arcade games from the 1970s and 1980s, one game at a time, with no knowledge of the rules. The AI would fumble, adapt, and eventually master each game. Hassabis described the cognitive processes involved as very like those of a human who had never seen the game before. That humble origin poses the central question this documentary traces: how does a lab that started with Pong end up sharing a Nobel Prize and reshaping molecular biology, weather forecasting, and the structure of the internet's video traffic?

• 2. From Startup to Google

  ›

  Horizons Ventures and Founders Fund were among the major venture capital firms that backed the young company. Entrepreneurs including Peter Thiel, Scott Banister, and Elon Musk also invested. Jaan Tallinn joined as an early investor and adviser. Facebook reportedly entered negotiations to buy the company in 2013 but pulled back. On the 26th of January 2014, Google confirmed an acquisition at a price reported to range between 400 million and 650 million dollars. That same year, DeepMind received the Company of the Year award from Cambridge Computer Laboratory. The acquisition brought resources but also a years-long tension: DeepMind executives pushed steadily for greater autonomy from Google. That struggle ended in April 2023 when DeepMind merged with Google's Brain division to form Google DeepMind, a consolidation driven in part by the pressure of OpenAI's ChatGPT.

• 3. The Go Breakthrough

  ›

  In October 2015, AlphaGo beat Fan Hui, the European Go champion, five games to zero. Fan Hui held a 2-dan professional ranking out of a possible 9, and no artificial intelligence had ever before defeated a professional Go player. Go posed a particular challenge because its number of possible positions made brute-force calculation prohibitively difficult, unlike chess. The algorithm fed on more than 30 million historical tournament moves and then played against itself to improve its winning rate. It used two deep neural networks: a policy network to evaluate probable moves and a value network to assess board positions. In March 2016, AlphaGo beat Lee Sedol, a 9-dan professional, four games to one in a five-game match. At the 2017 Future of Go Summit, it then defeated Ke Jie, who had been the world's highest-ranked player for two years, winning a three-game match. Later that year, AlphaGo Zero, trained with no human game data at all and running on just four Google TPUs instead of AlphaGo's 48, defeated the original AlphaGo one hundred games to zero after only three days of self-play.

• 4. Solving Protein Folding

  ›

  In December 2018, AlphaFold entered the 13th Critical Assessment of Techniques for Protein Structure Prediction, known as CASP, and correctly predicted the most accurate structure for 25 out of 43 proteins. Hassabis called it "our first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem" in remarks to The Guardian. By the 14th CASP in 2020, AlphaFold's accuracy was regarded as comparable to laboratory techniques. Andriy Kryshtafovych, one of the scientific adjudicators, described the result as "truly remarkable" and said the folding problem had been "largely solved". In July 2021, AlphaFold2 was released as open-source software. A week later DeepMind announced predictions of nearly all human proteins plus the complete proteomes of 20 other widely studied organisms. By July 2022, predictions covering more than 200 million proteins representing virtually all known proteins were released on the AlphaFold database. AlphaFold3, released in May 2024, added the ability to predict interactions between proteins and DNA, RNA, and other molecules; on a benchmark test for DNA interactions it reached 65% accuracy against a prior state of the art of 28%. In October 2024, Demis Hassabis and John Jumper received half of the Nobel Prize in Chemistry for their work on protein structure prediction.

• 5. Algorithms in the Real World

  ›

  In 2014, a Google datacenter engineer began using supervised machine learning to predict the power usage effectiveness of Google's datacenters. Two years later, inspired by AlphaGo, he asked DeepMind to apply reinforcement learning to the problem. An early system read sensor data and sent email recommendations to human engineers every 15 minutes; despite producing strategies that seemed unintuitive to long-time operators, it still achieved a 15% saving in power usage effectiveness. A later, more autonomous version checked its actions against safety constraints and, when verified safe, acted without human approval. That system reached a 30% saving, with one notable cooling strategy exploiting winter conditions to produce colder-than-normal water. Google subsequently worked with Trane Technologies to deploy similar systems on HVAC equipment at facilities outside Google. Meanwhile, AlphaDev, announced in June 2023, discovered a sorting algorithm that was 70% faster for short sequences and 1.7% faster for sequences longer than 250,000 elements; it was accepted into the C++ Standard Library, the first addition to those algorithms in more than a decade and the first to involve an AI-discovered method. Google estimated these two algorithms are executed trillions of times every day.

• 6. Language, Gemini, and Generative Tools

  ›

  WaveNet, a text-to-speech system introduced by DeepMind in 2016, was initially too computationally expensive for consumer products. By late 2017 it was ready for deployment in Google Assistant, and in 2018 Google launched Cloud Text-to-Speech built on WaveNet. In 2022 DeepMind released AlphaCode, a coding system tested against Codeforces challenges and trained on GitHub data; it earned a rank equivalent to 54% of the median score on that competitive programming platform. Gemini, a multimodal large language model released on the 6th of December 2023, succeeded Google's LaMDA and PaLM 2 and was built to compete with OpenAI's GPT-4. It launched in three sizes: Nano, Pro, and Ultra. Gemini 2.0 Flash followed on the 12th of December 2024 with expanded multimodal output including images and audio. On the 25th of March 2025, Google released Gemini 2.5, a reasoning model that pauses to think before responding, with Google announcing that all future models would carry similar reasoning capability. On the 18th of November 2025, Gemini 3 Pro arrived, fully multimodal and integrated with Google Search on the same day. Alongside Gemini, DeepMind produces Veo for video generation, Imagen for images, and Lyria for music; Veo 3, released in May 2025, generates synchronized dialogue, sound effects, and ambient audio to accompany its video output.

• 7. Healthcare and Controversy

  ›

  In July 2016, DeepMind announced a collaboration with Moorfields Eye Hospital to analyze anonymized eye scans for early signs of diseases that lead to blindness. A month later, a separate partnership with University College London Hospital aimed at algorithms to distinguish healthy from cancerous tissue in head and neck areas. Staff at the Royal Free Hospital reported in December 2017 that the Streams app, which sent alerts about acute kidney injury to mobile phones, had made a phenomenal difference in patient management. Yet the same hospital trust sat at the center of a significant data-protection dispute. In April 2016, New Scientist obtained a data-sharing agreement showing DeepMind Health had access to records of approximately 1.6 million patients treated annually at three London hospitals, including sensitive details such as HIV status, depression history, and abortion records. A complaint filed with the Information Commissioner's Office followed. In July 2017, the ICO ruled that the Royal Free hospital had failed to comply with the Data Protection Act when handing over patient data, and a leaked letter from National Data Guardian Dame Fiona Caldicott stated that in her considered opinion the data-sharing had taken place on an inappropriate legal basis. On the 13th of November 2018, DeepMind announced its health division and the Streams app would be absorbed into Google Health, a move privacy advocates said contradicted earlier assurances that patient data would not be connected to Google accounts.

• 8. Mathematics and What Comes Next

  ›

  AlphaGeometry solved 25 out of 30 geometry problems from the International Mathematical Olympiad, a performance comparable to a gold medalist, by pairing a symbolic proof engine with a large language model trained on synthetic geometric data. At the 2024 International Mathematical Olympiad, AlphaProof combined a fine-tuned Gemini model with the AlphaZero reinforcement learning algorithm; working alongside an adapted version of AlphaGeometry, the combined system reached the level of a silver medalist. In May 2025, AlphaEvolve arrived as an evolutionary coding agent that discovered improvements across open mathematical problems; tested on 50 such problems, it matched state-of-the-art algorithms 75% of the time and found improved solutions in 20% of cases, including a new result for the Kissing number in 11 dimensions. It also developed a datacenter scheduling heuristic that recovers an average of 0.7% of Google's worldwide compute. In April 2025, DeepMind introduced DolphinGemma, a research model designed to learn the structure of dolphin vocalizations and generate novel dolphin-like sound sequences; researchers hope to use it as a foundation for eventually decoding dolphin communication.