News from the AI & ML world

DeeperML - #deepmind

@Google DeepMind Blog // 24d
Google is pushing the boundaries of AI and robotics with its Gemini AI models. Gemini Robotics, an advanced vision-language-action model, now enables robots to perform physical tasks with improved generalization, adaptability, and dexterity. This model interprets and acts on text, voice, and image data, showcasing Google's advancements in integrating AI for practical applications. Furthermore, the development of Gemini Robotics-ER, which incorporates embodied reasoning capabilities, signifies another step toward smarter, more adaptable robots.

Google's approach to robotics emphasizes safety, employing both physical and semantic safety systems. The company is inviting filmmakers and creators to experiment with the model to improve the design and development. Veo builds on years of generative video model work, including Generative Query Network(GQN),DVD-GAN,Imagen-Video,Phenaki,WALT,VideoPoetandLumiere— combining architecture, scaling laws and other novel techniques to improve quality and output resolution.

Recommended read:
References :
  • Google DeepMind Blog: Gemini Robotics brings AI into the physical world
  • Maginative: Google DeepMind Unveils Gemini Robotics Models to Bridge AI and Physical World
  • IEEE Spectrum: With Gemini Robotics, Google Aims for Smarter Robots
  • The Official Google Blog: Take a closer look at our new Gemini models for robotics.
  • THE DECODER: Google Deepmind unveils new AI models for robotic control
  • www.tomsguide.com: Google is putting it's Gemini 2.0 AI into robots — here's how it's going
  • Verdict: Google DeepMind unveils Gemini AI models for robotics
  • MarkTechPost: Google DeepMind’s Gemini Robotics: Unleashing Embodied AI with Zero-Shot Control and Enhanced Spatial Reasoning
  • LearnAI: Research Published 12 March 2025 Authors Carolina Parada Introducing Gemini Robotics, our Gemini 2.0-based model designed for robotics At Google DeepMind, we’ve been making progress in how our Gemini models solve complex problems through multimodal reasoning across text, images, audio and video. So far however, those abilities have been largely confined to the digital realm....
  • OODAloop: Google DeepMind unveils new AI models for robotic control.
  • www.producthunt.com: Gemini Robotics
  • Last Week in AI: Last Week in AI #303 - Gemini Robotics, Gemma 3, CSM-1B
  • Windows Copilot News: Google is prepping Gemini to take action inside of apps
  • Last Week in AI: Discusses Gemini Robotics in the context of general AI agents and robotics.
  • www.infoq.com: Google DeepMind unveils Gemini Robotics, an advanced AI model for enhancing robotics through vision, language, and action.
  • AI & Machine Learning: This article discusses how generative AI is poised to revolutionize multiplayer games, offering personalized experiences through dynamic narratives and environments. The article specifically mentions Google's Gemini AI model and its potential to enhance gameplay.
  • Gradient Flow: This podcast episode discusses various advancements in AI, including Google's Gemini Robotics and Gemma 3, as well as the evolving regulatory landscape across different countries.
  • Insight Partners: This article highlights Andrew Ng's keynote at ScaleUp:AI '24, where he discusses the exciting trends in AI agents and applications, mentioning Google's Gemini AI assistant and its role in driving innovation.
  • www.tomsguide.com: You can now use Google Gemini without an account — here's how to get started

Maria Deutscher@AI ? SiliconANGLE // 5d
Isomorphic Labs, an Alphabet spinout focused on AI-driven drug design, has secured $600 million in its first external funding round. The investment, led by Thrive Capital with participation from Alphabet and GV, will fuel the advancement of Isomorphic Labs' AI drug design engine and therapeutic programs. The company aims to leverage artificial intelligence, including its AlphaFold technology, to revolutionize drug discovery across various therapeutic areas, including oncology and immunology. This funding is expected to accelerate research and development efforts, as well as facilitate the expansion of Isomorphic Labs' team with top-tier talent.

Isomorphic Labs, founded in 2021 by Sir Demis Hassabis, seeks to reimagine and accelerate drug discovery by applying AI. Its AI-powered engine streamlines the design of small molecules with therapeutic applications and can predict the effectiveness of a small molecule's attachment to a protein. The company's software also eases other aspects of the drug development workflow. Isomorphic Labs has already established collaborations with pharmaceutical companies like Eli Lilly and Novartis, and the new funding will support the progression of its own drug programs into clinical development.

Recommended read:
References :
  • www.genengnews.com: DeepMind Spinout Isomorphic Labs Raises $600M Toward AI Drug Design
  • Crunchbase News: Alphabet-Backed Isomorphic Labs Raises $600M For AI Drug Development
  • Maginative: Isomorphic Labs Secures $600M to Accelerate AI-Powered Drug Discovery
  • AI ? SiliconANGLE: Alphabet spinout Isomorphic Labs raises $600M for its AI drug design engine
  • Silicon Canals: London-based Isomorphic Labs, an AI-focused drug design and development company, has raised $600M (nearly €555.44M) in its first external funding round.
  • GZERO Media: Meet Isomorphic Labs, the Google spinoff that aims to cure you
  • www.genengnews.com: DeepMind Spinout Isomorphic Labs Raises $600M Toward AI Drug Design
  • Maginative: AI-first drug design company Isomorphic Labs has raised $600 million in its first external funding round, led by Thrive Capital, to advance its AI drug design engine and therapeutic programs.
  • eWEEK: AI drug design and development company Isomorphic Labs announced on Monday that it raised $600 million in its first external funding round. Thrive Capital led the financing round with participation from Google’s GV, and follow-on capital from Alphabet, its parent company and an existing investor. Isomorphic Labs described plans to use the funds to drive

@phys.org // 22h
Google's DeepMind has achieved a significant breakthrough in artificial intelligence with its Dreamer AI system. The AI has successfully mastered the complex task of mining diamonds in Minecraft without any explicit human instruction. This feat, accomplished through trial-and-error reinforcement learning, demonstrates the AI's ability to self-improve and generalize knowledge from one scenario to another, mimicking human-like learning processes. The achievement is particularly noteworthy because Minecraft's randomly generated worlds present a unique challenge, requiring the AI to adapt and understand its environment rather than relying on memorized strategies.

Mining diamonds in Minecraft is a complex, multi-step process that typically requires players to gather resources to build tools, dig to specific depths, and avoid hazards like lava. The Dreamer AI system tackled this challenge by exploring the game environment and identifying actions that would lead to rewards, such as finding diamonds. By repeating successful actions and avoiding less productive ones, the AI quickly learned to navigate the game and achieve its goal. According to Jeff Clune, a computer scientist at the University of British Columbia, this represents a major step forward for the field of AI.

The Dreamer AI system, developed by Danijar Hafner, Jurgis Pasukonis, Timothy Lillicrap and Jimmy Ba, achieved expert status in Minecraft in just nine days, showcasing its rapid learning capabilities. One unique approach used during training was to restart the game with a new virtual universe every 30 minutes, forcing the algorithm to constantly adapt and improve. This innovative method allowed the AI to quickly master the game's mechanics and develop strategies for diamond mining without any prior training or human intervention, pushing the boundaries of what AI can achieve in dynamic and complex environments.

Recommended read:
References :
  • techxplore.com: Google's AI Dreamer learns how to self-improve over time by mastering Minecraft
  • Analytics Vidhya: What if I told you that AI can now outperform humans in some of the most complex video games? AI now masters Minecraft too.
  • eWEEK: The new Dreamer AI system figured out how to conduct the multi-step process of mining diamonds without being taught how to play Minecraft.
  • www.scientificamerican.com: The Dreamer AI system of Google's DeepMind reached the milestone of mastering Minecraft by ‘imagining’ the future impact of possible decisions

Nitika Sharma@Analytics Vidhya // 1d
Google's DeepMind has achieved a significant milestone in artificial intelligence by developing an AI system, named Dreamer, that has mastered Minecraft without any human instruction or data. The Dreamer AI system successfully learned how to mine diamonds, a complex and multi-step process, entirely on its own through trial and error. This breakthrough highlights the potential for AI systems to generalize knowledge and transfer skills from one domain to another, marking a major step forward in the field of AI development.

Researchers programmed the Dreamer AI to play Minecraft by setting up a system of rewards, particularly for finding diamonds. The AI explores the game on its own, identifying actions that lead to in-game rewards and repeating those actions. The AI was able to reach an expert level within just nine days. The results are a good sign that AI apps can learn to improve its abilities over a short period of time, which could give robots the tools they need to perform well in the real world.

Recommended read:
References :
  • Analytics Vidhya: Google’s DeepMind Masters Minecraft Without Human Data
  • eWEEK: AI Cracks Minecraft’s Toughest Challenge Without Help From Humans
  • techxplore.com: Google's AI Dreamer learns how to self-improve over time by mastering Minecraft
  • The Official Google Blog: Start building with Gemini 2.5 Pro.

@Google DeepMind Blog // 2d
Google DeepMind is intensifying its focus on AI governance and security as it ventures further into artificial general intelligence (AGI). The company is exploring AI monitors to regulate hyperintelligent AI models, splitting potential threats into four categories, with the creation of a "monitor" AI being one proposed solution. This proactive approach includes prioritizing technical safety, conducting thorough risk assessments, and fostering collaboration within the broader AI community to navigate the development of AGI responsibly.

DeepMind's reported clampdown on sharing research will stifle AI innovation, warns the CEO of Iris.ai, one of Europe’s leading startups in the space, Anita Schjøll Abildgaard. Concerns are rising within the AI community that DeepMind's new research restrictions threaten AI innovation. The CEO of Iris.ai, a Norwegian startup developing an AI-powered engine for science, warns the drawbacks will far outweigh the benefits. She fears DeepMind's restrictions will hinder technological advances.

Recommended read:
References :
  • Google DeepMind Blog: We’re exploring the frontiers of AGI, prioritizing technical safety, proactive risk assessment, and collaboration with the AI community.
  • The Next Web: Google DeepMind’s reported clampdown on sharing research will stifle AI innovation, warns the CEO of Iris.ai, one of Europe’s leading startups in the space.
  • www.techrepublic.com: DeepMind’s approach to AGI safety and security splits threats into four categories. One solution could be a “monitor†AI.

@techcrunch.com // 57d
DeepMind's artificial intelligence, AlphaGeometry2, has achieved a remarkable feat by solving 84% of the geometry problems from the International Mathematical Olympiad (IMO) over the past 25 years. This performance surpasses the average gold medalist in the prestigious competition for gifted high school students. The AI's success highlights the growing capabilities of AI in handling sophisticated mathematical tasks.

AlphaGeometry2 represents an upgraded system from DeepMind, incorporating advancements such as the integration of Google's Gemini large language model and the ability to reason by manipulating geometric objects. This neuro-symbolic system combines a specialized language model with abstract reasoning coded by humans, enabling it to generate rigorous proofs and avoid common AI pitfalls like hallucinations. This could potentially impact fields that heavily rely on mathematical expertise.

Recommended read:
References :
  • www.nature.com: This news report discusses DeepMind's AI achieving performance comparable to top human solvers in mathematics.
  • techcrunch.com: DeepMind says its AlphaGeometry2 model solved 84% of International Math Olympiad's geometry problems from the last 25 years, surpassing average gold medalists
  • Techmeme: DeepMind says its AlphaGeometry2 model solved 84% of International Math Olympiad's geometry problems from the last 25 years, surpassing average gold medalists (Kyle Wiggers/TechCrunch)
  • techxplore.com: TechXplore reports on DeepMind AI achieves gold-medal level performance on challenging Olympiad math questions.
  • www.analyticsvidhya.com: DeepMind’s AlphaGeometry2 Surpasses Math Olympiad
  • www.marktechpost.com: Marktechpost discusses Google DeepMind's AlphaGeometry2.

@Google DeepMind Blog // 2d
Google DeepMind has released a strategy paper outlining its approach to the development of safe artificial general intelligence (AGI). According to DeepMind, AGI, defined as AI capable of matching or exceeding human cognitive abilities, could emerge as early as 2030. The company emphasizes the importance of proactive risk assessment, technical safety measures, and collaboration within the AI community to ensure responsible development. They are exploring the frontiers of AGI, prioritizing readiness and identifying potential challenges and benefits.

DeepMind's strategy identifies four key risk areas: misuse, misalignment, accidents, and structural risks, with an initial focus on misuse and misalignment. Misuse refers to the intentional use of AI systems for harmful purposes, such as spreading disinformation. DeepMind is also introducing Gemini Robotics, which it touts as its most advanced vision-language-action model. Gemini Robotics aims to allow robots to comprehend something in front of them, interact with a user, and take action.

Recommended read:
References :
  • THE DECODER: Google Deepmind says AGI might outthink humans by 2030, and it's planning for the risks
  • LearnAI: We’re exploring the frontiers of AGI, prioritizing readiness, proactive risk assessment, and collaboration with the wider AI community.
  • Google DeepMind Blog: Our framework enables cybersecurity experts to identify which defenses are necessary—and how to prioritize them
  • www.techrepublic.com: DeepMind’s approach to AGI safety and security splits threats into four categories. One solution could be a “monitor†AI.

@www.marktechpost.com // 54d
DeepMind's AlphaGeometry2, an AI system, has achieved a remarkable milestone by surpassing the average performance of gold medalists in the International Mathematical Olympiad (IMO) geometry problems. This significant upgrade to the original AlphaGeometry demonstrates the potential of AI in tackling complex mathematical challenges that require both high-level reasoning and strategic problem-solving abilities. The system leverages advanced AI techniques to solve these intricate geometry problems, marking a notable advancement in AI's capabilities.

Researchers from Google DeepMind, alongside collaborators from the University of Cambridge, Georgia Tech, and Brown University, enhanced the system with a Gemini-based language model, a more efficient symbolic engine, and a novel search algorithm with knowledge sharing. These improvements have significantly boosted its problem-solving rate to 84% on IMO geometry problems from 2000-2024. AlphaGeometry2 represents a step towards a fully automated system capable of interpreting problems from natural language and devising solutions, underscoring AI's growing potential in fields demanding high mathematical reasoning skills, such as research and education.

Recommended read:
References :
  • the-decoder.com: The latest version of Deepmind's AlphaGeometry system can solve geometry problems better than most human experts, matching the performance of top math competition winners.
  • techxplore.com: DeepMind AI achieves gold-medal level performance on challenging Olympiad math questions
  • Analytics Vidhya: DeepMind’s AlphaGeometry2 Surpasses Math Olympiad
  • MarkTechPost: The International Mathematical Olympiad (IMO) is a globally recognized competition that challenges high school students with complex mathematical problems.
  • www.analyticsvidhya.com: DeepMind’s AlphaGeometry2 Surpasses Math Olympiad
  • www.marktechpost.com: Google DeepMind Introduces AlphaGeometry2: A Significant Upgrade to AlphaGeometry Surpassing the Average Gold Medalist in Solving Olympiad Geometry

Synced@Synced // 6d
References: Synced , www.theguardian.com
DeepMind has announced significant advancements in AI modeling and biomedicine, pushing the boundaries of what's possible with artificial intelligence. The company's research is focused on creating more effective drugs and medicine, as well as understanding and protecting species around the world.

DeepMind's JetFormer, a novel Transformer model, is designed to directly model raw data, eliminating the need for pre-trained components. JetFormer can understand and generate both text and images seamlessly. This model leverages normalizing flows to encode images into a latent representation, enhancing the focus on essential high-level information through progressive Gaussian noise augmentation. JetFormer has demonstrated competitive performance in image generation and web-scale multimodal generation tasks.

Additionally, DeepMind is exploring how studying honeybee immunity could offer insights into protecting various species. The company's AlphaFold continues to revolutionize biology, aiding in the design of more effective drugs. AlphaFold, which uses AI to determine a protein's structure, has been used to solve fundamental questions in biology, awarded the Nobel prize (in chemistry – to Demis Hassabis and John Jumper) and revolutionised drug discovery. There are approximately 250,000,000 protein structures in the AlphaFold database, which has been used by almost 2 million people from 190 countries.

Recommended read:
References :
  • Synced: DeepMind’s JetFormer: Unified Multimodal Models Without Modelling Constraints
  • www.theguardian.com: AI may help us cure countless diseases – and usher in a new golden age of medicine | Samuel Hume

@www.infoq.com // 4d
Google DeepMind has unveiled TxGemma, an AI designed to improve drug discovery and clinical trial predictions. TxGemma, built upon the Gemma model family, aims to streamline the drug development process and accelerate the creation of new treatments. This announcement comes as Isomorphic Labs, an Alphabet spinout, secured $600 million in funding to further develop its AI drug design engine, which reduces manual labor and speeds up drug development.

Isomorphic Labs' engine uses AI to design small molecules with therapeutic applications, predicting their effectiveness in attaching to disease-causing proteins and mapping properties like solubility. This is powered by models like Google's AlphaFold 3, which can predict the shape of proteins, DNA, and RNA, crucial for drug development. The funding will accelerate Isomorphic Labs' research and development efforts, expand its team, and advance its programs, including those focused on oncology and immunology, toward clinical development.

Recommended read:
References :
  • AI ? SiliconANGLE: Alphabet spinout Isomorphic Labs raises $600M for its AI drug design engine
  • www.infoq.com: Designed to enhance the efficiency of drug discovery and clinical trial predictions.
  • www.genengnews.com: DeepMind Spinout Isomorphic Labs Raises $600M Toward AI Drug Design

@Google DeepMind Blog // 7d
Researchers are making strides in understanding how AI models think. Anthropic has developed an "AI microscope" to peek into the internal processes of its Claude model, revealing how it plans ahead, even when generating poetry. This tool provides a limited view of how the AI processes information and reasons through complex tasks. The microscope suggests that Claude uses a language-independent internal representation, a "universal language of thought", for multilingual reasoning.

The team at Google DeepMind introduced JetFormer, a new Transformer designed to directly model raw data. This model, capable of both understanding and generating text and images seamlessly, maximizes the likelihood of raw data without depending on any pre-trained components. Additionally, a comprehensive benchmark called FACTS Grounding has been introduced to evaluate the factuality of large language models (LLMs). This benchmark measures how accurately LLMs ground their responses in provided source material and avoid hallucinations, aiming to improve trust and reliability in AI-generated information.

Recommended read:
References :
  • Google DeepMind Blog: FACTS Grounding: A new benchmark for evaluating the factuality of large language models
  • THE DECODER: Anthropic's AI microscope reveals how Claude plans ahead when generating poetry

@boards.greenhouse.io // 49d
References: AI Alignment Forum , youtu.be , youtu.be ...
DeepMind has launched a short course on AGI (Artificial General Intelligence) safety, targeting students, researchers, and professionals interested in the field. The course offers an accessible introduction to AI alignment, comprising short recorded talks and exercises totaling 75 minutes, complemented by an accompanying slide deck and exercise workbook. It addresses anticipated alignment challenges as AI capabilities advance and outlines DeepMind's current technical and governance approaches to mitigate these problems.

Key topics covered in the course include evidence suggesting the field is progressing toward advanced AI capabilities and arguments for instrumental subgoals and deliberate planning as potential sources of risk. It also differentiates between specification gaming and goal misgeneralization as ways misaligned goals can arise. The course delves into DeepMind's technical approach to AI alignment, emphasizing informed oversight and frontier safety practices such as dangerous capability evaluations, alongside institutional approaches to AI safety. If the course inspires you, you can apply to work with DeepMind in Research Scientist roles.

Recommended read:
References :
  • AI Alignment Forum: A short course on AGI safety from the GDM Alignment team
  • youtu.be: This YouTube video is part of DeepMind's short course on AGI safety. It focuses on the alignment problem, including risk arguments and technical challenges in AI alignment.
  • docs.google.com: Published on February 14, 2025 3:43 PM GMT We are excited to release a short course on AGI safety for students, researchers and professionals interested in this topic. The course offers a concise and accessible introduction to AI alignment, consisting of short recorded talks and exercises (75 minutes total) with an accompanying  and  . It covers alignment problems we can expect as AI capabilities advance, and our current approach to these problems (on technical and governance levels). If you would like to learn more about AGI safety but have only an hour to spare, this course is for you!  Here are some key topics you will learn about in this course: The evidence for the field being on a path to advanced AI capabilities. Arguments for instrumental subgoals and deliberate planning towards a misaligned goal as a source of extreme risk. Two ways in which misaligned goals may arise – specification gaming and goal misgeneralization – as well as the difference between the two. Our technical approach to AI alignment and its components. The guiding principle of informed oversight ("knowing what the AI system knows") and how it's implemented in our approach. What is involved in enabling AI safety on an institutional level, including frontier safety practices such as dangerous capability evaluations. Course outline: Part 0:  (4 minutes) Part 1: The alignment problem.  This part covers risk arguments and technical problems in AI alignment. (5 minutes) (7 minutes) (3 minutes) (10 minutes) (3 minutes) Part 2: Our technical approach.  The first talk outlines our overall technical approach, and the following talks cover different components of this approach. (4 minutes) (6 minutes) (4 minutes) (5 minutes) (4 minutes) (4 minutes) Part 3: Our governance approach.  This part covers our approach to AI governance, starting from a high-level overview and then going into specific governance practices. (7 minutes) (4 minutes) (7 minutes) If this course gets you excited about AGI safety, you can apply to work with us! Applications for  and  roles are open until Feb 28.
  • youtu.be: YouTube video from the course
  • boards.greenhouse.io: Applications for Research Scientist and Engineering roles are open until Feb 28.
  • AI Alignment Forum: AGI Safety & Alignment @ Google DeepMind is hiring

@www.analyticsvidhya.com // 53d
DeepMind has unveiled AlphaGeometry2, a significant upgrade to its AlphaGeometry system. This new iteration achieves gold-medal level performance in solving challenging Olympiad geometry problems, surpassing the abilities of the average gold medalist. Researchers from Google DeepMind, along with collaborators from the University of Cambridge, Georgia Tech, and Brown University, enhanced the system's domain language, enabling it to handle more complex geometric concepts and increasing its coverage of IMO problems from 66% to 88%.

AlphaGeometry2 integrates a Gemini-based language model with a more efficient symbolic engine and a novel search algorithm. These improvements boost its solving rate to 84% on IMO geometry problems from 2000-2024. The system is advancing towards a fully automated system that interprets problems from natural language. Prior research suggests that AI capable of solving geometry problems could lead to more sophisticated applications, requiring both a high level of reasoning ability and the ability to choose from possible steps in working toward a solution.

Recommended read:
References :
  • techxplore.com: DeepMind AI achieves gold-medal level performance on challenging Olympiad math questions
  • www.analyticsvidhya.com: DeepMind’s AlphaGeometry2 Surpasses Math Olympiad
  • www.marktechpost.com: Google DeepMind Introduces AlphaGeometry2: A Significant Upgrade to AlphaGeometry Surpassing the Average Gold Medalist in Solving Olympiad Geometry

Nathan Labenz@The Cognitive Revolution // 19d
DeepMind's Allan Dafoe, Director of Frontier Safety and Governance, is actively involved in shaping the future of AI governance. Dafoe is addressing the challenges of evaluating AI capabilities, understanding structural risks, and navigating the complexities of governing AI technologies. His work focuses on ensuring AI's responsible development and deployment, especially as AI transforms sectors like education, healthcare, and sustainability, while mitigating potential risks through necessary safety measures.

Google is also prepping its Gemini AI model to take actions within apps, potentially revolutionizing how users interact with their devices. This development, which involves a new API in Android 16 called "app functions," aims to give Gemini agent-like abilities to perform tasks inside applications. For example, users might be able to order food from a local restaurant using Gemini without directly opening the restaurant's app. This capability could make AI assistants significantly more useful.

Recommended read:
References :

Ben Lorica@Gradient Flow // 27d
References: Gradient Flow
DeepSeek is making significant strides in the AI landscape, particularly within the healthcare sector in China. The AI solution is being rapidly adopted across China's tertiary hospitals to improve clinical decision-making and operational efficiency. Its rollout began in Shanghai, with hospitals like Fudan University Affiliated Huashan Hospital, and has expanded nationwide. DeepSeek is being used in areas such as intelligent pathology to automate tumor analysis, imaging analysis for lung nodule differentiation, clinical decision support for evidence retrieval, and workflow optimization to reduce patient wait times.

DeepSeek has also open-sourced several code repositories to give competitors a scare on the journey toward transparency and the advancement of the AI community. This move puts the firm ahead of the competition on model transparency and the open source nature allows hospitals to customize the programs. This level of openness is a further step than other AI competitors such as Meta’s Llama, which has only open-sourced the weights of its models. DeepSeek's deployment focuses on practical applications within hospital intranets, ensuring data security while improving accuracy and generalization through hierarchical knowledge distillation, reducing computational costs.

Recommended read:
References :
  • Gradient Flow: DeepSeek in Action: Practical AI Applications Transforming Chinese Healthcare