What is Google’s DeepMind Known For? A Look Into Their Major Breakthroughs What is Google’s DeepMind Known For? A Look Into Their Major Breakthroughs
Google’s DeepMind is elbow-deep in some incredible pieces of research and AI innovation. So let’s take a quick dive and see... What is Google’s DeepMind Known For? A Look Into Their Major Breakthroughs

Google’s DeepMind is elbow-deep in some incredible pieces of research and AI innovation. So let’s take a quick dive and see what the team is up to and hopefully get an idea of what the future might hold for the prestigious research team.

In-Person and Virtual Conference

April 23rd to 25th, 2024

Join us for a deep dive into the latest data science and AI trends, tools, and techniques, from LLMs to data analytics and from machine learning to responsible AI.



Using general-purpose machine learning techniques, which include neural networks, self-play via reinforcement learning, multi-agent learning, and imitation learning, AlphaStar is the first AI to reach the top league of a widely popular esport without any game restrictions. According to their blog post, AlphaStar was able to challenge two of the world’s top players in StarCraft II.

The game is considered one of the most enduring and popular real-time strategy video games of all time. Now the program can play the full game at a Grandmaster level under professionally approved conditions.


This is a system by DeepMind that uses reinforcement learning to discover enhanced computer science algorithms. It even has been able to surpass those honed by scientists and engineers over decades in a bid to improve code and make it more powerful and sustainable in the long term. In short, the goal of AlphaDev is to help generate better algorithms using AI that will transform how we program computers and impact all aspects of our increasingly digital society, while keeping new sorting algorithms open-sourced, in the main C++ library.

Universal Speech Model

Universal Speech Model is a family of state-of-the-art speech models from DeepMind. They are 2B parameters and trained on 12 million hours of speech. This also includes 28 billion sentences of text, spanning 300+ languages. Many people don’t realize they even use this program as USM is currently being used in YouTube for closed captions, which allow for automatic speech recognition in many popular languages such as English and Madrian. 


As one of the first AI models to generate natural-sounding speech, WaveNet is a general-purpose piece of technology. To date, it has allowed researchers, teams, and others to unlock a range of new applications utilizing AI to help with many different issues. From improving video calls on even the weakest connections to helping people regain their original voice after losing the ability to speak. WaveNet seeks to bridge the gap in vocalized communication through AI.


The purpose of AlphaCode by Google’s DeepMind is to write computer programs at a competitive level. The AI uses transformer-based language models to generate code at an unprecedented scale, and then smartly filters to a small set of promising programs. So far, AlphaCode has come pretty far as it has achieved an estimated rank within the top 54% of participants in programming competitions. The rank is calculated by how the program is able to solve new problems that require a combination of critical thinking, logic, algorithms, coding, and natural language understanding.



Google’s DeepMind is claiming that AlphaTensor is the first artificial intelligence system that can be used for the discovery of novel, efficient, and provably correct algorithms for fundamental tasks such as matrix multiplication. This is a part of DeepMind’s mission to help advance science by using AI systems to unlock solutions to complex mathematical problems. The program builds upon AlphaZero an agent with excellent performance with games such as Go and Shogi. 

PaLM 2

PaLM 2 is Google’s next-generation LLM that was pre-trained on parallel multilingual text and on a much larger corpus of different languages than its predecessor, PaLM. From reasoning, logic, coding, math, question answering, and more, this model seeks to accomplish tasks quickly and effectively no matter their complexity. 


Using the PaLM LLM, PaLM-SayCan by Google’s DeepMind is an algorithm that specializes in robotics. It combines the power of large language models with help bots in a bid to showcase the potential of robotics and advanced AI technology. SayCan works by splitting prompts into sub-tasks for the helper robots to execute complex tasks one by one, bridging the gap between hardware limitations and advanced reasoning necessary in robotics. 


AlpahGo is the program that turned the game of Go on its head. It was able to do this by defeating the Go world champion a full decade before experts thought it was possible. This AI mode is built upon a system that combines deep neural networks with advanced search algorithms. The first neural network is known as the policy network, which makes predictions of movement, while the other neural network, known as the value network, predicts the winner of the game. Together, they’ve predicted movements by players in multiple games in an effective manner. 


Unlike AlphaGo which learned to play go analyzing game data, AlphaZero was taught how to play using reinforcement learning, with only the rules of games as a guide until it learned how to win said games. Real-world application of this model includes faster sorting of data, matrix multiplication algorithms, and ways of reducing internet traffic to help deliver content in a way that improves general access to information. 

In-Person Data Engineering Conference

April 23rd to 24th, 2024 – Boston, MA

At our second annual Data Engineering Summit, Ai+ and ODSC are partnering to bring together the leading experts in data engineering and thousands of practitioners to explore different strategies for making data actionable.



Similar to OpenAI’s Sora, this model is able to synthesize realistic video from textual prompt sequences. Phenaki also looks to address the issue of high computational costs associated with video generation by leveraging a transformer model that translates text embeddings to video tokens and an encoder-decoder model that compresses videos to discrete tokens with a tokenizer. 


Yes, another model with alpha in the name! Well, this one specializes in the field of biology, looking to enhance the discovery of protein and protein prediction. This was accomplished by teaching the program sequences and structures of around 100,000 known proteins and now it is able to predict the shape of one down to atomic accuracy rapidly. 


With the explosion of AI-generated content, the need to be able to identify what is AI-generated and what is not has become an important talking point. Well, with DeepMind’s SynthID, the first steps at watermarking AI-generated content have come online. The goal of this program isn’t just for identifying but to also watermark AI-generated images and auditions. The watermark is imperceptible to humans but is detectable to AI. The main goal of this program is to help build trust between AI systems and users. 

Imagen 2

Of course, Google’s Deepmind wasn’t going to not have its answer to OpenAI’s Dalle-2 or MidJourney. That’s where Imagen 2 comes in. By using advanced text-to-imager diffusion technology, this model is able to deliver high-quality and photorealistic images that are very close to aligned with user prompts. 


This is a family of lightweight open models that are built from the same foundations of the current Gemini Models Google has pushed out recently. Currently, the Gemma Models are seeing impressive benchmarks when compared with MIstral or Llama-2 while often being smaller in size. Google’s family of Gemma models provides a flexible framework that is compatible with PyTorch, TensorFlow, and JAX, as a means to provide researchers wiggle room on project requirements.

Tell me more!

It’s clear that Google’s DeepMind team is pushing the envelope in multiple subfields and areas outside just data science with its catalog of models and programs. What has been done so far is amazing, one can only imagine what the team has up their sleeves in the coming years. 

Now if you want to stay in the known with the latest AI developments, one of the best ways to do so is to meet, learn from, and network with the leading experts driving cutting-edge research forward. 

And the best place to do that is at ODSC East this April 23rd to 25th!



ODSC gathers the attendees, presenters, and companies that are shaping the present and future of data science and AI. ODSC hosts one of the largest gatherings of professional data scientists with major conferences in USA, Europe, and Asia.